(10/14) NEWEST COMFYUI 0.40 version sucks. It breaks a lot of workflows. Don't update if you can avoid it. Let them fix their shit.
I just made a Discord finally! Please join to connect with other AI Filmmakers or if you have questions about my workflows/process: https://discord.gg/PGAXv4xq7
[ free patreon for workflows/tutorials/support me: patreon.com/FoxFurEssence ]
[ new patreon for NSFW content I make: patreon.com/FoxFurSpice ]
This advanced filmmaker's workflow cleanly and quickly generates 3 low resolution sample preview gens, allowing you to select your favorite to then take to upscale finalization. The workflow has every feature an AI filmmaker could need, and easily produce the highest quality and resolution Minimax is capable of. It comes equipped with all quality of life features, the top known speed ups, and many options, from seamless video continuation to custom audio, to frame interpolation to go from 24 FPS to 48-60. Everything is here!
v2.5's Dialogue / Voice Cloner tool Quick Guide:
v2.0 Usage Guide (a bit outdated but introduces the core ideas):
Shortfilm entirely made using this workflow:
Some nodes in this workflow are too new to be in the Comfy Registry to be installed automatically with Comfy-Manager! Therefore, they need to be installed manually.
.: Latent Upscaler
- Node: https://github.com/LBH-123-AI/Comfyui_Minimax_h3_latent_Upscaler
- Latent Upscale Model: https://huggingface.co/LBH-123-AI/Minimax_h3_latent_Upscaler/blob/main/minimax_h3_latent_upscaler_3d_conv_v1/minimax_h3_latent_upscaler_3d_conv_v1_fp16.safetensors
(put into /latent_upscale_models/ folder and, since 9/17, this needs to be reselected in the WF's selection for latent upscaler choice due to the author renaming the file)
.: Pixaroma Node Suite
https://github.com/pixaroma/ComfyUI-Pixaroma
.: Prompt IDE
https://github.com/ethanfel/ComfyUI-H3-Prompt-IDE
.: Resources Loader (MiniMaxH3-PromptBuilder)
https://github.com/Adudeguyman/ComfyUI-Fantastic-MiniMaxH3-PromptBuilder
.: RefMods
https://github.com/Luisacaotica/ComfyUI-MiniMaxH3Mod
.: Turbo LoRA
NEW [Kijai's DMAD 4-step - for both FL2VA/REF2VA - 8 steps 1.0 strength recommended] https://huggingface.co/Kijai/MiniMax-H3-experimental/blob/main/loras/minimax_h3_DMAD_4step_full_lora_avg_rank_39_bf16.safetensors
Older:
This particular Turbo LoRA is my favorite. I recommend 6-8 steps if at 1.0 strength, or 8-10 steps if using at 0.6-0.8 strength. If you're upscaling, 6 first pass + 3 upscale pass works well.
If you use ref2va or Singularity, this is great too:
.: Kijai's int8 Video VAE (much faster than default!)
Everything else should be easily available from the ComfyUI-manager Install Missing Nodes screen!
.: TaeH3 (for Kijai's Model Preview Override):
https://huggingface.co/Kijai/MiniMax-H3-TAE/blob/main/vae_approx/taeh3.safetensors
(note, using Kijai's Model Previewer will slow gens down by 6-16% thereabouts, I have it disabled by default now, but you can turn it on to test for yourself).
If you support what I'm doing, please consider following or subscribing to my Patreon! It's free (if you want it to be) and all my content and updates go there!
Description
Fixed 4th Image: It wasn't connected to anything so your 4th reference image wouldn't apply. Also corrected an audio connection bug that could lead to errors depending on your configuration.
Fixed RIFE bug. Redownload if you have v1.2.0
Adds SEAMLESS VIDEO CONTINUATION, a very flexible tool for extending pre-existing videos. No new custom nodes.
Returns to Kijai's trusted Sol Attention (since SLA is basically the same thing now) -- so, one less custom node to download yay!
FAQ
Comments (206)
Does this one work without the whole oversharpening thing that other workflows suffer from w/ continuous generation? I've yet to find one that solves it - something I noticed with others like context loop and multishot is that for every shot the colors get more saturated and image gets sharper and sharper
If this one doesn't do it i'll be jumping for joy
Well, considering this one doesn't use any previously established methods, you'll have to just give it a shot and see for yourself. For 90% of outputs I can't see any difference between the original video and the new gen. I don't get any oversharpening issues -- especially if not using a turbo LoRA (which you really shouldn't for this kind of thing). 12 steps 1st pass, 4 steps final upscale. High quality input video + high quality output video yields really nice results in my testing.
@foxydits Thanks for responding so quick, I'm gonna give it a go right now - I've been fighting for days to get a seamless video and have tried 4-5 diff flows and tried vibe coding one out but I'm also quite dumb so hopefully this does it, I'll report back
@datlurkaa Keep in mind, a truly seamless video requires video editing on top to make sure lighting and framing remain constant. There is no 100% perfect solution in comfyUI alone. The hardest ones will always be still shots where it's very easy to tell that there's a subtle shift in the tone/lighting/quality.
@foxydits ah yea that's my issue - I dont mind editing, I've just found others have sharpening essentially destroy the image, movements in hue/saturation etc are simple enough to fix in davinci, its more or less just making sure it gets through the whole video without going wild on sharpening, generating now!
Hey - Something odd, I tried your wf before the continuation update, gave me that fake Rife error and proceeded to work perfectly, however with this new update the rife error is actually aborting the process, even when using the vanilla workflow here with zero changes not sure if this is due to a comfyui update or something but it refuses to generate for me. I loaded up your previous WF and its working great, very strange
@datlurkaa are you doing vid continuation or normal genning? i will need to look into it but need details for what you were attempting first.
@foxydits Just normal genning - Going to try on a fresh portable install and report back - it happened after updating some custom nodes so I'm wondering if one of them broke it
@foxydits okay on a brand new portable install w/ all required nodes it's stopping it in its tracks, not continuing its generation:
# ComfyUI Error Report ## Error Details - Node ID: 165 - Node Type: RIFEInterpolation - Exception Type: ValueError - Exception Message: ValueError: No images provided
Update - V10 Workflow works - V12 Stops, gonna keep playing with it but something changed that it apparently hates
@datlurkaa yeah, i'm seeing now the RIFE node no longer allows the bypass if images aren't supplied. i guess for now just disable RIFE in options, it's completely optional after all.
@foxydits Think I found the v12 RIFE issue. In v10 the switch feeding RIFE is bypassed during the initial run, so the “error” stays harmless. In v12 the Final Video Selection switch is active, so RIFE actually gets called with no images and hard-stops on No images provided.
I changed that switch from mode 0 to mode 4 to match v10 and its triggering
@datlurkaa No worries, I fixed it am about to upload the new version. It needed a failsafe, something that sends it an image if all the other outputs to it are turned off. So I send it a single black image that doesn't save if everything else isn't on.
@foxydits Awesome :) I'll update mine once you push it through, starting my long-form to see if it holds, I absolutely love this setup so I'm hoping it sticks, multiple low passes to choose from is brilliant
@datlurkaa thanks, and yeah, it's updated now.
Wanted to jump back in here and thank you for your help, though it didn't really solve the degradation thing happening. I'm trying to do a long form (a few minutes) of a person talking at 10 second chunks, with no cuts (no turbo lora)
I used the trim (22) and while the seam is fantastic, it does suffer from the image getting sharper and sharper for every generation, something I couldn't fix in post editing software, was able to solve for saturation but the burn-in from the image getting sharper made my characters skin and details go really wonky once I hit about 40 seconds.
Not sure how to solve this, or if its even solvable in the first place.
@datlurkaa you solve that by using a reference for your character and scene. load your char reference sheet or high quality images of the char, and optionally a high quality reference for the scene location, and make sure you prompt to include that the character(s) appearing in the shot are the ones in the ref. that way the model always has a high quality reference to go off of and you will never lose fidelity in your scenes.
@foxydits This is exactly what I did - I use my character (HQ) & a forest scene, tag em properly using Minimax's syntax, once I hit the 4th clip everything is noticeably more sharper with some saturation/luma changes, the character never really adjusts distance from camera so it's definitely odd, I'm at a loss
@datlurkaa I am assuming you're not just loading char pics into your picture references and hoping that's enough, right? Like you have to prompt correctly and assign them to their picture references. <Subject 1> is the girl in <Picture 1>. -- later in detailed_description: <Subject 1> is the running through the forest -- basically assigning the action from the extended video is associated with the picture references. That's all done through prompting, so if you're not doing that kind of thing, it's important.
@foxydits Oh 100% I'm doing that
I'm using a character with a solid grey background - using the Subject, picture assignment, I also have a picture for the background as well and use the correct prompting language.
I've tried different trim/overlap lengths as well, the seams are perfect but for every single video it just gets sharper and sharper, (It seems subtle but breaks down by clip 4-5) even when maintaining the same exact conversational distance, not sure if you can try replicating this but I'm curious if this is a me thing or a limitation thing. I really want to get this working as I'm loving this WF but can't seem to quite crack it, I've tried other WF's with the same issue though.
If you try 4-5 clips at 4 seconds each same conversational distance with no cuts of a character just talking you may be getting the same issue as me, or it might confirm I'm just doing something fundamentally wrong but
I use the same prompt each time, just changing the words, everything else remains constant
Do the low pass, then upscale, load up the upscaled audio, trim & set overlay
render,upscale, repeat - All from the vanilla WF with no messing around
Tried 0.3 w/ 1.0 scaling and also 0.9 with 1.0 scaling, still hitting the same wall
I tried fixing it in davinci but the sharpening is rough, much easier to deal with saturation
also; ty for taking time to help me out, you're the goat
@datlurkaa np, I'm very curious as to how this can be solved for you. So, the only thing that is leading to quality loss is those initial overlap frames from the previous gen. I assume what's happening is it's setting the 'tone' of what minimax thinks your scene should look like. I am assuming this is all supposed to be one continuous shot right? Like there is no room for a cut where you reset your scene from scratch and then just begin to continue off of that until it starts to degrade again? Because obviously you don't want to be using my workflow to be compiling a short film.. the vid extension feature is just for when you need to continue a shot without it looking like there was a cut. If there's a cut anywhere, just start a new gen from scratch and get that high quality back. You know? Everything gets edited into the final piece in davinci/video editor at the end anyway.
@foxydits I'm attempting to make 2-3 minutes of a no cut talking scene - basically a talking head scene, the issue is there is actually a lot of damage happening on each generation, I've actually been spending time vibe coding out color locking & anti-drift nodes and have it almost eliminated, its heavily modifying the workflow and isn't the most efficient right now for convenience but it seems to have improved it a lot
Original:
Mean edge accumulation +15.49%
Mean high-frequency accumulation +22.29%
Final seam edge accumulation +26.74%
Final seam HF accumulation +39.04%
Boundary RGB error 0.498 code
Modified
Mean edge accumulation +1.97% / 87.3% better
Mean high-frequency accumulation −4.42% / 80.2% lower magnitude
Final seam edge accumulation +4.06% / 84.8% better
Final seam HF accumulation −2.19% / 94.4% better
Boundary RGB error 0.0000089 code / 99.9982% better
It's not perfect yet but improved quite a lot
How do you prompt for the continuation video feature? Can you give an example?
You can use natural language or throw in bits and pieces of reference syntax.
Here's the prompt for the baseball sample:the cute asian woman throws a baseball, and when the baseball hits the blocks in the distance, the entire side of the building explodes and begins to collapse, smoke and dust pluming out towards the girl in the foreground as she turns around and says: <d>[Japanese] Yatta, ne?!</d>. Use <Picture 1> fully_preserved as a reference for her face and outfit. she is not carrying a baseball in her hand at the end of the video.
@foxydits I'm still so confused on how to use this feature. I had a 10s video I generated from the previous v10 workflow. I want to do another 10s that continue from this video. I loaded it in the "Seamless Video Continuation". Do I also load the ref pics from the previous finished video? And do I reference the previous video as <Video 1> or don't reference it at all in the prompt? I want to do a 10s continuation with various actions and sound effects. And I'm not sure how natural language without a ref2va format would work...
It'd be great if you can make a tutorial for this feature or explain it in more details on how to set up the ref and prompts.
@teiji25762 the tutorial video was supposed to come out with this workflow but i ran out of time. i put a brief explanation in the description here on this civit page. you don't need to reference any videos in the prompt. this is completely automatic in that way, just prompt what you want to happen.
God bless you
can you make a video about the video continuation? it's a bit confusing how to use it.
yes, that was planned for today but time has run away from me. it will come out soon. for now i have an explanation in the civit page description.
Having issues with the latest version. Saying I'm missing the SolAttnPatch. Any help. Thx
Install Sol Attention node. The link is in the bright red note that says "CUSTOM NODE LINKS", it's also in the civit page description.
@foxydits I from the link in the red comments https://github.com/kijai/ComfyUI-SolAttn_triton and installed in the custom nodes folder. Still throws the same error.
@freewellgear if you don't have triton installed, you can't use it. just disable it in the options, it's not required for the workflow to run, it's just a speedup optimization. then you can focus on getting triton installed on your own later.
Is there any optimizations for upscaler? rtx 5080 and upscaling proces take too long even with scale 1.25 from 0.4 MP.
Should definitely not take long if you're 0.4 MP -> 1.25x upscale. That's not really much of an upscale at all. Do you have enough system RAM?
@foxydits 64GB RAM
@lukaskolembus245 Yeah something else is very wrong then. 0.4 MP to 0.6 MP (a 1.25x upscale) should take like... 2 minutes.
Hi, I’m getting an error when running this workflow.
The workflow stops at the RIFEInterpolation node (Node 165) with the following error:
ValueError: No images provided
The log also shows several upstream nodes with missing required inputs, including model, conditioning, vae, sampler, and sigmas.
I’ve pasted the full error log below. Could you please let me know what is causing this and how I can fix it?
Thank you.
ノードの実行中にエラーが発生しました。
# ComfyUI Error Report
## Error Details
- Node ID: 165
- Node Type: RIFEInterpolation
- Exception Type: ValueError
- Exception Message: ValueError: No images provided
## Stack Trace
```
File "E:\Ai\ComfyUI_windows_portable\ComfyUI\execution.py", line 545, in execute
output_data, output_ui, has_subgraph, has_pending_tasks = await get_output_data(prompt_id, unique_id, obj, input_data_all, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb, v3_data=v3_data)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "E:\Ai\ComfyUI_windows_portable\ComfyUI\execution.py", line 344, in get_output_data
return_values = await asyncmap_node_over_list(prompt_id, unique_id, obj, input_data_all, obj.FUNCTION, allow_interrupt=True, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb, v3_data=v3_data)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "E:\Ai\ComfyUI_windows_portable\ComfyUI\execution.py", line 318, in asyncmap_node_over_list
await process_inputs(input_dict, i)
File "E:\Ai\ComfyUI_windows_portable\ComfyUI\execution.py", line 306, in process_inputs
result = f(**inputs)
File "E:\Ai\ComfyUI_windows_portable\ComfyUI\custom_nodes\ComfyUI-VFI\nodes.py", line 105, in interpolate
raise ValueError("No images provided")
```
## System Information
- ComfyUI Version: 0.34.0
- Arguments: ComfyUI\main.py --windows-standalone-build --enable-manager
- OS: win32
- Python Version: 3.13.14 (tags/v3.13.14:fd17997, Jun 10 2026, 13:03:48) [MSC v.1944 64 bit (AMD64)]
- Embedded Python: true
- PyTorch Version: 2.13.0+cu130
## Devices
- Name: cuda:0 NVIDIA GeForce RTX 4070 : cudaMallocAsync
- Type: cuda
- VRAM Total: 12878086144
- VRAM Free: 11647582208
- Torch VRAM Total: 0
- Torch VRAM Free: 0
## Logs
```
2026-08-29T10:51:35.817223 - [92m[save_image_extended][0m version: 2.64[0m2026-08-29T10:51:35.817326 -
2026-08-29T10:51:35.819811 -
======================================================================2026-08-29T10:51:35.819924 -
2026-08-29T10:51:35.820009 - [SAM3] ComfyUI SAM3 Node - Loading (TBG)2026-08-29T10:51:35.820088 -
2026-08-29T10:51:35.820157 - ======================================================================2026-08-29T10:51:35.820220 -
2026-08-29T10:51:35.820934 - [SAM3] SAM3 Python package detected.2026-08-29T10:51:35.821023 -
2026-08-29T10:51:35.862679 - [SAM3] Registered 6 nodes:2026-08-29T10:51:35.862786 -
2026-08-29T10:51:35.862862 - [SAM3] - TBG SAM3 Model Loader (TBGLoadSAM3Model)2026-08-29T10:51:35.862942 -
2026-08-29T10:51:35.863027 - [SAM3] - TBG SAM3 Model Loader and Downloader (TBGSAM3ModelLoaderAdvanced)2026-08-29T10:51:35.863094 -
2026-08-29T10:51:35.863160 - [SAM3] - TBG SAM3 Segmentation (TBGSam3Segmentation)2026-08-29T10:51:35.863217 -
2026-08-29T10:51:35.863273 - [SAM3] - TBG SAM3 Batch Selector (TBGSam3SegmentationBatch)2026-08-29T10:51:35.863326 -
2026-08-29T10:51:35.863382 - [SAM3] - TBG SAM3 Selector (TBGSAM3PromptCollector)2026-08-29T10:51:35.863433 -
2026-08-29T10:51:35.863487 - [SAM3] - TBG SAM3 Depth Map (TBGSAM3DepthMap)2026-08-29T10:51:35.863537 -
2026-08-29T10:51:35.863589 - ======================================================================2026-08-29T10:51:35.863642 -
2026-08-29T10:51:35.892298 - [32m[INFO][0m
Import times for custom nodes:
2026-08-29T10:51:35.892482 - [32m[INFO][0m 0.0 seconds: E:\Ai\ComfyUI_windows_portable\ComfyUI\custom_nodes\websocket_image_save.py
2026-08-29T10:51:35.892601 - [32m[INFO][0m 0.0 seconds: E:\Ai\modeldata\custom_nodes\WildcardSizeProcessor_ChatGPT
2026-08-29T10:51:35.892703 - [32m[INFO][0m 0.0 seconds: E:\Ai\modeldata\custom_nodes\Prompt_join_ChatGPT
2026-08-29T10:51:35.892801 - [32m[INFO][0m 0.0 seconds: E:\Ai\modeldata\custom_nodes\tag_filter_node_ChatGPT
2026-08-29T10:51:35.892933 - [32m[INFO][0m 0.0 seconds: E:\Ai\modeldata\custom_nodes\undo_text_multiline_ChatGPT
2026-08-29T10:51:35.893051 - [32m[INFO][0m 0.0 seconds: E:\Ai\modeldata\custom_nodes\Resolution_Preset_Switch_ChatGPT
2026-08-29T10:51:35.893169 - [32m[INFO][0m 0.0 seconds: E:\Ai\modeldata\custom_nodes\dynamic_text_multiline_ChatGPT
2026-08-29T10:51:35.893281 - [32m[INFO][0m 0.0 seconds: E:\Ai\ComfyUI_windows_portable\ComfyUI\custom_nodes\comfyui-obvpm
2026-08-29T10:51:35.893384 - [32m[INFO][0m 0.0 seconds: E:\Ai\ComfyUI_windows_portable\ComfyUI\custom_nodes\ComfyUI_MinimaxH3HybridLoader
2026-08-29T10:51:35.893501 - [32m[INFO][0m 0.0 seconds: E:\Ai\ComfyUI_windows_portable\ComfyUI\custom_nodes\comfyui-minimax-h3-turbo
2026-08-29T10:51:35.893612 - [32m[INFO][0m 0.0 seconds: E:\Ai\ComfyUI_windows_portable\ComfyUI\custom_nodes\ComfyUI-Spectrum-sdxl
2026-08-29T10:51:35.893724 - [32m[INFO][0m 0.0 seconds: E:\Ai\ComfyUI_windows_portable\ComfyUI\custom_nodes\ComfyUI-MiniMaxH3Mod
2026-08-29T10:51:35.893830 - [32m[INFO][0m 0.0 seconds: E:\Ai\ComfyUI_windows_portable\ComfyUI\custom_nodes\ComfyUI-Anima-LLLite
2026-08-29T10:51:35.893938 - [32m[INFO][0m 0.0 seconds: E:\Ai\ComfyUI_windows_portable\ComfyUI\custom_nodes\Comfyui_Minimax_h3_latent_Upscaler
2026-08-29T10:51:35.894044 - [32m[INFO][0m 0.0 seconds: E:\Ai\modeldata\custom_nodes\comfyui-dynamicprompts
2026-08-29T10:51:35.894155 - [32m[INFO][0m 0.0 seconds: E:\Ai\ComfyUI_windows_portable\ComfyUI\custom_nodes\wavespeed
2026-08-29T10:51:35.894263 - [32m[INFO][0m 0.0 seconds: E:\Ai\ComfyUI_windows_portable\ComfyUI\custom_nodes\ComfyUI-VFI
2026-08-29T10:51:35.894369 - [32m[INFO][0m 0.0 seconds: E:\Ai\ComfyUI_windows_portable\ComfyUI\custom_nodes\comfy-ex-tagcomplete
2026-08-29T10:51:35.894475 - [32m[INFO][0m 0.0 seconds: E:\Ai\ComfyUI_windows_portable\ComfyUI\custom_nodes\WhatDreamsCost-ComfyUI
2026-08-29T10:51:35.894581 - [32m[INFO][0m 0.0 seconds: E:\Ai\ComfyUI_windows_portable\ComfyUI\custom_nodes\save-image-extended-comfyui
2026-08-29T10:51:35.894686 - [32m[INFO][0m 0.0 seconds: E:\Ai\ComfyUI_windows_portable\ComfyUI\custom_nodes\comfyui_essentials
2026-08-29T10:51:35.894791 - [32m[INFO][0m 0.0 seconds: E:\Ai\ComfyUI_windows_portable\ComfyUI\custom_nodes\comfyui-spectrum-sdxl-proper
2026-08-29T10:51:35.894896 - [32m[INFO][0m 0.0 seconds: E:\Ai\ComfyUI_windows_portable\ComfyUI\custom_nodes\comfyui-custom-scripts
2026-08-29T10:51:35.894998 - [32m[INFO][0m 0.0 seconds: E:\Ai\ComfyUI_windows_portable\ComfyUI\custom_nodes\comfyui-extended
2026-08-29T10:51:35.895104 - [32m[INFO][0m 0.0 seconds: E:\Ai\ComfyUI_windows_portable\ComfyUI\custom_nodes\comfyui-frame-interpolation
2026-08-29T10:51:35.895207 - [32m[INFO][0m 0.0 seconds: E:\Ai\ComfyUI_windows_portable\ComfyUI\custom_nodes\ComfyUI-SolAttn_triton
2026-08-29T10:51:35.895322 - [32m[INFO][0m 0.0 seconds: E:\Ai\ComfyUI_windows_portable\ComfyUI\custom_nodes\rgthree-comfy
2026-08-29T10:51:35.895426 - [32m[INFO][0m 0.0 seconds: E:\Ai\ComfyUI_windows_portable\ComfyUI\custom_nodes\cg-use-everywhere
2026-08-29T10:51:35.895556 - [32m[INFO][0m 0.0 seconds: E:\Ai\ComfyUI_windows_portable\ComfyUI\custom_nodes\ComfyUI-MiniMaxH3
2026-08-29T10:51:35.895699 - [32m[INFO][0m 0.0 seconds: E:\Ai\ComfyUI_windows_portable\ComfyUI\custom_nodes\comfyui-spectrum-minimax-h3
2026-08-29T10:51:35.895836 - [32m[INFO][0m 0.0 seconds: E:\Ai\ComfyUI_windows_portable\ComfyUI\custom_nodes\tbg-sam3
2026-08-29T10:51:35.895965 - [32m[INFO][0m 0.1 seconds: E:\Ai\ComfyUI_windows_portable\ComfyUI\custom_nodes\comfyui-kjnodes
2026-08-29T10:51:35.896078 - [32m[INFO][0m 0.1 seconds: E:\Ai\ComfyUI_windows_portable\ComfyUI\custom_nodes\comfyui_nvidia_rtx_nodes
2026-08-29T10:51:35.896186 - [32m[INFO][0m 0.1 seconds: E:\Ai\ComfyUI_windows_portable\ComfyUI\custom_nodes\comfyui-videohelpersuite
2026-08-29T10:51:35.896292 - [32m[INFO][0m 0.1 seconds: E:\Ai\ComfyUI_windows_portable\ComfyUI\custom_nodes\comfyui-easy-use
2026-08-29T10:51:35.896397 - [32m[INFO][0m 0.2 seconds: E:\Ai\ComfyUI_windows_portable\ComfyUI\custom_nodes\comfyui-impact-pack
2026-08-29T10:51:35.896506 - [32m[INFO][0m 0.2 seconds: E:\Ai\ComfyUI_windows_portable\ComfyUI\custom_nodes\ComfyUI-GIMM-VFI
2026-08-29T10:51:35.896613 - [32m[INFO][0m 0.3 seconds: E:\Ai\ComfyUI_windows_portable\ComfyUI\custom_nodes\comfyui_fizznodes
2026-08-29T10:51:35.896720 - [32m[INFO][0m 0.3 seconds: E:\Ai\ComfyUI_windows_portable\ComfyUI\custom_nodes\ComfyUI-DaSiWa-Nodes
2026-08-29T10:51:35.896824 - [32m[INFO][0m 0.4 seconds: E:\Ai\ComfyUI_windows_portable\ComfyUI\custom_nodes\comfyui-easy-sam3
2026-08-29T10:51:35.896927 - [32m[INFO][0m 1.0 seconds: E:\Ai\ComfyUI_windows_portable\ComfyUI\custom_nodes\comfyui-dynamicprompts
2026-08-29T10:51:35.897085 - [32m[INFO][0m
2026-08-29T10:51:35.901016 - [32m[INFO][0m Context impl SQLiteImpl.
2026-08-29T10:51:35.901179 - [32m[INFO][0m Will assume non-transactional DDL.
2026-08-29T10:51:36.018010 - [32m[INFO][0m Using RAM pressure cache.
2026-08-29T10:51:36.019092 - [32m[INFO][0m Starting server
2026-08-29T10:51:36.019970 - [32m[INFO][0m To see the GUI go to: http://127.0.0.1:8188
2026-08-29T10:51:40.201496 - [32m[INFO][0m [ComfyUI-Manager] The ComfyRegistry cache update is still in progress, so an outdated cache is being used.
2026-08-29T10:51:40.347586 - FETCH DATA from: E:\Ai\ComfyUI_windows_portable\python_embeded\Lib\site-packages\comfyui_manager\custom-node-list.json2026-08-29T10:51:40.347696 - 2026-08-29T10:51:40.370390 - [DONE]2026-08-29T10:51:40.370517 -
2026-08-29T10:54:08.106055 - [32m[INFO][0m FETCH ComfyRegistry Data [DONE]
2026-08-29T10:54:08.449967 - [32m[INFO][0m [ComfyUI-Manager] default cache updated: https://api.comfy.org/nodes
2026-08-29T10:54:08.516205 - FETCH DATA from: E:\Ai\ComfyUI_windows_portable\ComfyUI\user\__manager\cache\1514988643_custom-node-list.json2026-08-29T10:54:08.516345 - 2026-08-29T10:54:08.553192 - [DONE]2026-08-29T10:54:08.553307 -
2026-08-29T10:54:08.619853 - [32m[INFO][0m [ComfyUI-Manager] All startup tasks have been completed.
2026-08-29T11:06:56.225813 - [32m[INFO][0m got prompt
2026-08-29T11:06:56.234771 - [1m[31m[ERROR][0m Failed to validate prompt for output 144:
2026-08-29T11:06:56.235012 - [1m[31m[ERROR][0m * BasicGuider 143:140:
2026-08-29T11:06:56.235129 - [1m[31m[ERROR][0m - Required input is missing: model
2026-08-29T11:06:56.235260 - [1m[31m[ERROR][0m - Required input is missing: conditioning
2026-08-29T11:06:56.235373 - [1m[31m[ERROR][0m * MiniMaxH3ReferenceToVideo 5:
2026-08-29T11:06:56.235472 - [1m[31m[ERROR][0m - Required input is missing: vae
2026-08-29T11:06:56.235568 - [1m[31m[ERROR][0m * SamplerCustomAdvanced 143:139:
2026-08-29T11:06:56.235660 - [1m[31m[ERROR][0m - Required input is missing: sampler
2026-08-29T11:06:56.235752 - [1m[31m[ERROR][0m - Required input is missing: sigmas
2026-08-29T11:06:56.235846 - [1m[31m[ERROR][0m * VAEDecode 180:
2026-08-29T11:06:56.235939 - [1m[31m[ERROR][0m - Required input is missing: vae
2026-08-29T11:06:56.236032 - [1m[31m[ERROR][0m Output will be ignored
2026-08-29T11:06:56.236516 - [1m[31m[ERROR][0m Failed to validate prompt for output 18:
2026-08-29T11:06:56.236643 - [1m[31m[ERROR][0m * BasicGuider 125:14:
2026-08-29T11:06:56.236742 - [1m[31m[ERROR][0m - Required input is missing: model
2026-08-29T11:06:56.236836 - [1m[31m[ERROR][0m - Required input is missing: conditioning
2026-08-29T11:06:56.236930 - [1m[31m[ERROR][0m * SamplerCustomAdvanced 125:12:
2026-08-29T11:06:56.237022 - [1m[31m[ERROR][0m - Required input is missing: sampler
2026-08-29T11:06:56.237114 - [1m[31m[ERROR][0m - Required input is missing: sigmas
2026-08-29T11:06:56.237207 - [1m[31m[ERROR][0m * VAEDecode 176:
2026-08-29T11:06:56.237307 - [1m[31m[ERROR][0m - Required input is missing: vae
2026-08-29T11:06:56.237409 - [1m[31m[ERROR][0m Output will be ignored
2026-08-29T11:06:56.238035 - [1m[31m[ERROR][0m Failed to validate prompt for output 134:
2026-08-29T11:06:56.238154 - [1m[31m[ERROR][0m * BasicGuider 133:130:
2026-08-29T11:06:56.238252 - [1m[31m[ERROR][0m - Required input is missing: model
2026-08-29T11:06:56.238346 - [1m[31m[ERROR][0m - Required input is missing: conditioning
2026-08-29T11:06:56.238439 - [1m[31m[ERROR][0m * SamplerCustomAdvanced 133:129:
2026-08-29T11:06:56.238532 - [1m[31m[ERROR][0m - Required input is missing: sampler
2026-08-29T11:06:56.238623 - [1m[31m[ERROR][0m - Required input is missing: sigmas
2026-08-29T11:06:56.238713 - [1m[31m[ERROR][0m Output will be ignored
2026-08-29T11:06:56.267624 - [1m[31m[ERROR][0m !!! Exception during processing !!! No images provided
2026-08-29T11:06:56.299432 - [1m[31m[ERROR][0m Traceback (most recent call last):
File "E:\Ai\ComfyUI_windows_portable\ComfyUI\execution.py", line 545, in execute
output_data, output_ui, has_subgraph, has_pending_tasks = await get_output_data(prompt_id, unique_id, obj, input_data_all, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb, v3_data=v3_data)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "E:\Ai\ComfyUI_windows_portable\ComfyUI\execution.py", line 344, in get_output_data
return_values = await asyncmap_node_over_list(prompt_id, unique_id, obj, input_data_all, obj.FUNCTION, allow_interrupt=True, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb, v3_data=v3_data)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "E:\Ai\ComfyUI_windows_portable\ComfyUI\execution.py", line 318, in asyncmap_node_over_list
await process_inputs(input_dict, i)
File "E:\Ai\ComfyUI_windows_portable\ComfyUI\execution.py", line 306, in process_inputs
result = f(**inputs)
File "E:\Ai\ComfyUI_windows_portable\ComfyUI\custom_nodes\ComfyUI-VFI\nodes.py", line 105, in interpolate
raise ValueError("No images provided")
ValueError: No images provided
2026-08-29T11:06:56.301530 - [32m[INFO][0m [32mPrompt executed in 0.05 seconds[0m
2026-08-29T11:10:29.892073 - [32m[INFO][0m got prompt
2026-08-29T11:10:29.900952 - [1m[31m[ERROR][0m Failed to validate prompt for output 144:
2026-08-29T11:10:29.901140 - [1m[31m[ERROR][0m * BasicGuider 143:140:
2026-08-29T11:10:29.901253 - [1m[31m[ERROR][0m - Required input is missing: model
2026-08-29T11:10:29.901377 - [1m[31m[ERROR][0m - Required input is missing: conditioning
2026-08-29T11:10:29.901499 - [1m[31m[ERROR][0m * MiniMaxH3ReferenceToVideo 5:
2026-08-29T11:10:29.901606 - [1m[31m[ERROR][0m - Required input is missing: vae
2026-08-29T11:10:29.901735 - [1m[31m[ERROR][0m * SamplerCustomAdvanced 143:139:
2026-08-29T11:10:29.901844 - [1m[31m[ERROR][0m - Required input is missing: sampler
2026-08-29T11:10:29.901959 - [1m[31m[ERROR][0m - Required input is missing: sigmas
2026-08-29T11:10:29.902071 - [1m[31m[ERROR][0m * VAEDecode 180:
2026-08-29T11:10:29.902198 - [1m[31m[ERROR][0m - Required input is missing: vae
2026-08-29T11:10:29.902310 - [1m[31m[ERROR][0m Output will be ignored
2026-08-29T11:10:29.902755 - [1m[31m[ERROR][0m Failed to validate prompt for output 18:
2026-08-29T11:10:29.902883 - [1m[31m[ERROR][0m * BasicGuider 125:14:
2026-08-29T11:10:29.903000 - [1m[31m[ERROR][0m - Required input is missing: model
2026-08-29T11:10:29.903104 - [1m[31m[ERROR][0m - Required input is missing: conditioning
2026-08-29T11:10:29.903199 - [1m[31m[ERROR][0m * SamplerCustomAdvanced 125:12:
2026-08-29T11:10:29.903311 - [1m[31m[ERROR][0m - Required input is missing: sampler
2026-08-29T11:10:29.903435 - [1m[31m[ERROR][0m - Required input is missing: sigmas
2026-08-29T11:10:29.903546 - [1m[31m[ERROR][0m * VAEDecode 176:
2026-08-29T11:10:29.903651 - [1m[31m[ERROR][0m - Required input is missing: vae
2026-08-29T11:10:29.903765 - [1m[31m[ERROR][0m Output will be ignored
2026-08-29T11:10:29.904419 - [1m[31m[ERROR][0m Failed to validate prompt for output 134:
2026-08-29T11:10:29.904546 - [1m[31m[ERROR][0m * BasicGuider 133:130:
2026-08-29T11:10:29.904654 - [1m[31m[ERROR][0m - Required input is missing: model
2026-08-29T11:10:29.904767 - [1m[31m[ERROR][0m - Required input is missing: conditioning
2026-08-29T11:10:29.904867 - [1m[31m[ERROR][0m * SamplerCustomAdvanced 133:129:
2026-08-29T11:10:29.904983 - [1m[31m[ERROR][0m - Required input is missing: sampler
2026-08-29T11:10:29.905093 - [1m[31m[ERROR][0m - Required input is missing: sigmas
2026-08-29T11:10:29.905959 - [1m[31m[ERROR][0m Output will be ignored
2026-08-29T11:10:29.924103 - [1m[31m[ERROR][0m !!! Exception during processing !!! No images provided
2026-08-29T11:10:29.925342 - [1m[31m[ERROR][0m Traceback (most recent call last):
File "E:\Ai\ComfyUI_windows_portable\ComfyUI\execution.py", line 545, in execute
output_data, output_ui, has_subgraph, has_pending_tasks = await get_output_data(prompt_id, unique_id, obj, input_data_all, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb, v3_data=v3_data)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "E:\Ai\ComfyUI_windows_portable\ComfyUI\execution.py", line 344, in get_output_data
return_values = await asyncmap_node_over_list(prompt_id, unique_id, obj, input_data_all, obj.FUNCTION, allow_interrupt=True, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb, v3_data=v3_data)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "E:\Ai\ComfyUI_windows_portable\ComfyUI\execution.py", line 318, in asyncmap_node_over_list
await process_inputs(input_dict, i)
File "E:\Ai\ComfyUI_windows_portable\ComfyUI\execution.py", line 306, in process_inputs
result = f(**inputs)
File "E:\Ai\ComfyUI_windows_portable\ComfyUI\custom_nodes\ComfyUI-VFI\nodes.py", line 105, in interpolate
raise ValueError("No images provided")
ValueError: No images provided
2026-08-29T11:10:29.927004 - [32m[INFO][0m [32mPrompt executed in 0.01 seconds[0m
2026-08-29T11:14:22.464110 - [32m[INFO][0m got prompt
2026-08-29T11:14:22.472036 - [1m[31m[ERROR][0m Failed to validate prompt for output 144:
2026-08-29T11:14:22.472206 - [1m[31m[ERROR][0m * BasicGuider 143:140:
2026-08-29T11:14:22.472343 - [1m[31m[ERROR][0m - Required input is missing: model
2026-08-29T11:14:22.472454 - [1m[31m[ERROR][0m - Required input is missing: conditioning
2026-08-29T11:14:22.472553 - [1m[31m[ERROR][0m * MiniMaxH3ReferenceToVideo 5:
2026-08-29T11:14:22.472678 - [1m[31m[ERROR][0m - Required input is missing: vae
2026-08-29T11:14:22.472820 - [1m[31m[ERROR][0m * SamplerCustomAdvanced 143:139:
2026-08-29T11:14:22.473733 - [1m[31m[ERROR][0m - Required input is missing: sampler
2026-08-29T11:14:22.473846 - [1m[31m[ERROR][0m - Required input is missing: sigmas
2026-08-29T11:14:22.473976 - [1m[31m[ERROR][0m * VAEDecode 180:
2026-08-29T11:14:22.474089 - [1m[31m[ERROR][0m - Required input is missing: vae
2026-08-29T11:14:22.474218 - [1m[31m[ERROR][0m Output will be ignored
2026-08-29T11:14:22.474896 - [1m[31m[ERROR][0m Failed to validate prompt for output 18:
2026-08-29T11:14:22.475034 - [1m[31m[ERROR][0m * BasicGuider 125:14:
2026-08-29T11:14:22.475171 - [1m[31m[ERROR][0m - Required input is missing: model
2026-08-29T11:14:22.475302 - [1m[31m[ERROR][0m - Required input is missing: conditioning
2026-08-29T11:14:22.475407 - [1m[31m[ERROR][0m * SamplerCustomAdvanced 125:12:
2026-08-29T11:14:22.475508 - [1m[31m[ERROR][0m - Required input is missing: sampler
2026-08-29T11:14:22.475651 - [1m[31m[ERROR][0m - Required input is missing: sigmas
2026-08-29T11:14:22.475789 - [1m[31m[ERROR][0m * VAEDecode 176:
2026-08-29T11:14:22.475920 - [1m[31m[ERROR][0m - Required input is missing: vae
2026-08-29T11:14:22.476029 - [1m[31m[ERROR][0m Output will be ignored
2026-08-29T11:14:22.477476 - [1m[31m[ERROR][0m Failed to validate prompt for output 134:
2026-08-29T11:14:22.477692 - [1m[31m[ERROR][0m * BasicGuider 133:130:
2026-08-29T11:14:22.477810 - [1m[31m[ERROR][0m - Required input is missing: model
2026-08-29T11:14:22.477937 - [1m[31m[ERROR][0m - Required input is missing: conditioning
2026-08-29T11:14:22.478064 - [1m[31m[ERROR][0m * SamplerCustomAdvanced 133:129:
2026-08-29T11:14:22.478198 - [1m[31m[ERROR][0m - Required input is missing: sampler
2026-08-29T11:14:22.478318 - [1m[31m[ERROR][0m - Required input is missing: sigmas
2026-08-29T11:14:22.478420 - [1m[31m[ERROR][0m Output will be ignored
2026-08-29T11:14:22.633710 - [32m[INFO][0m VAE load device: cuda:0, offload device: cpu, dtype: torch.float32
2026-08-29T11:14:22.923107 - [32m[INFO][0m Found quantization metadata version 1
2026-08-29T11:14:22.924266 - [32m[INFO][0m Using MixedPrecisionOps for text encoder
2026-08-29T11:14:23.623959 - [32m[INFO][0m CLIP/text encoder model load device: cuda:0, offload device: cpu, current: cpu, dtype: torch.float16
2026-08-29T11:14:23.641568 - [1m[31m[ERROR][0m !!! Exception during processing !!! MiniMaxH3ReferenceToVideo.execute() missing 1 required positional argument: 'vae'
2026-08-29T11:14:23.671921 - [1m[31m[ERROR][0m Traceback (most recent call last):
File "E:\Ai\ComfyUI_windows_portable\ComfyUI\execution.py", line 545, in execute
output_data, output_ui, has_subgraph, has_pending_tasks = await get_output_data(prompt_id, unique_id, obj, input_data_all, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb, v3_data=v3_data)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "E:\Ai\ComfyUI_windows_portable\ComfyUI\execution.py", line 344, in get_output_data
return_values = await asyncmap_node_over_list(prompt_id, unique_id, obj, input_data_all, obj.FUNCTION, allow_interrupt=True, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb, v3_data=v3_data)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "E:\Ai\ComfyUI_windows_portable\ComfyUI\execution.py", line 318, in asyncmap_node_over_list
await process_inputs(input_dict, i)
File "E:\Ai\ComfyUI_windows_portable\ComfyUI\execution.py", line 306, in process_inputs
result = f(**inputs)
File "E:\Ai\ComfyUI_windows_portable\ComfyUI\comfy_api\internal\__init__.py", line 149, in wrapped_func
return method(locked_class, **inputs)
File "E:\Ai\ComfyUI_windows_portable\ComfyUI\comfy_api\latest\_io.py", line 1990, in EXECUTE_NORMALIZED
to_return = cls.execute(*args, **kwargs)
TypeError: MiniMaxH3ReferenceToVideo.execute() missing 1 required positional argument: 'vae'
2026-08-29T11:14:23.674046 - [32m[INFO][0m [32mPrompt executed in 1.19 seconds[0m
2026-08-29T11:16:23.750619 - [32m[INFO][0m got prompt
2026-08-29T11:16:23.758873 - [1m[31m[ERROR][0m Failed to validate prompt for output 144:
2026-08-29T11:16:23.759042 - [1m[31m[ERROR][0m * BasicGuider 143:140:
2026-08-29T11:16:23.759170 - [1m[31m[ERROR][0m - Required input is missing: model
2026-08-29T11:16:23.759302 - [1m[31m[ERROR][0m - Required input is missing: conditioning
2026-08-29T11:16:23.759436 - [1m[31m[ERROR][0m * MiniMaxH3ReferenceToVideo 5:
2026-08-29T11:16:23.759544 - [1m[31m[ERROR][0m - Required input is missing: vae
2026-08-29T11:16:23.759833 - [1m[31m[ERROR][0m * SamplerCustomAdvanced 143:139:
2026-08-29T11:16:23.760032 - [1m[31m[ERROR][0m - Required input is missing: sampler
2026-08-29T11:16:23.760154 - [1m[31m[ERROR][0m - Required input is missing: sigmas
2026-08-29T11:16:23.760261 - [1m[31m[ERROR][0m * VAEDecode 180:
2026-08-29T11:16:23.761123 - [1m[31m[ERROR][0m - Required input is missing: vae
2026-08-29T11:16:23.761388 - [1m[31m[ERROR][0m Output will be ignored
2026-08-29T11:16:23.762007 - [1m[31m[ERROR][0m Failed to validate prompt for output 18:
2026-08-29T11:16:23.762159 - [1m[31m[ERROR][0m * BasicGuider 125:14:
2026-08-29T11:16:23.762301 - [1m[31m[ERROR][0m - Required input is missing: model
2026-08-29T11:16:23.762430 - [1m[31m[ERROR][0m - Required input is missing: conditioning
2026-08-29T11:16:23.762557 - [1m[31m[ERROR][0m * SamplerCustomAdvanced 125:12:
2026-08-29T11:16:23.762676 - [1m[31m[ERROR][0m - Required input is missing: sampler
2026-08-29T11:16:23.762810 - [1m[31m[ERROR][0m - Required input is missing: sigmas
2026-08-29T11:16:23.762932 - [1m[31m[ERROR][0m * VAEDecode 176:
2026-08-29T11:16:23.763049 - [1m[31m[ERROR][0m - Required input is missing: vae
2026-08-29T11:16:23.763151 - [1m[31m[ERROR][0m Output will be ignored
2026-08-29T11:16:23.763872 - [1m[31m[ERROR][0m Failed to validate prompt for output 134:
2026-08-29T11:16:23.764004 - [1m[31m[ERROR][0m * BasicGuider 133:130:
2026-08-29T11:16:23.764119 - [1m[31m[ERROR][0m - Required input is missing: model
2026-08-29T11:16:23.764218 - [1m[31m[ERROR][0m - Required input is missing: conditioning
2026-08-29T11:16:23.764338 - [1m[31m[ERROR][0m * SamplerCustomAdvanced 133:129:
2026-08-29T11:16:23.764863 - [1m[31m[ERROR][0m - Required input is missing: sampler
2026-08-29T11:16:23.765112 - [1m[31m[ERROR][0m - Required input is missing: sigmas
2026-08-29T11:16:23.765269 - [1m[31m[ERROR][0m Output will be ignored
2026-08-29T11:16:24.240819 - [32m[INFO][0m SELECTED: input1
2026-08-29T11:16:24.247260 - [1m[31m[ERROR][0m !!! Exception during processing !!! MiniMaxH3ReferenceToVideo.execute() missing 1 required positional argument: 'vae'
2026-08-29T11:16:24.248392 - [1m[31m[ERROR][0m Traceback (most recent call last):
File "E:\Ai\ComfyUI_windows_portable\ComfyUI\execution.py", line 545, in execute
output_data, output_ui, has_subgraph, has_pending_tasks = await get_output_data(prompt_id, unique_id, obj, input_data_all, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb, v3_data=v3_data)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "E:\Ai\ComfyUI_windows_portable\ComfyUI\execution.py", line 344, in get_output_data
return_values = await asyncmap_node_over_list(prompt_id, unique_id, obj, input_data_all, obj.FUNCTION, allow_interrupt=True, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb, v3_data=v3_data)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "E:\Ai\ComfyUI_windows_portable\ComfyUI\execution.py", line 318, in asyncmap_node_over_list
await process_inputs(input_dict, i)
File "E:\Ai\ComfyUI_windows_portable\ComfyUI\execution.py", line 306, in process_inputs
result = f(**inputs)
File "E:\Ai\ComfyUI_windows_portable\ComfyUI\comfy_api\internal\__init__.py", line 149, in wrapped_func
return method(locked_class, **inputs)
File "E:\Ai\ComfyUI_windows_portable\ComfyUI\comfy_api\latest\_io.py", line 1990, in EXECUTE_NORMALIZED
to_return = cls.execute(*args, **kwargs)
TypeError: MiniMaxH3ReferenceToVideo.execute() missing 1 required positional argument: 'vae'
2026-08-29T11:16:24.250124 - [32m[INFO][0m [32mPrompt executed in 0.47 seconds[0m
2026-08-29T11:18:49.173570 - [32m[INFO][0m got prompt
2026-08-29T11:18:49.181753 - [1m[31m[ERROR][0m Failed to validate prompt for output 144:
2026-08-29T11:18:49.181931 - [1m[31m[ERROR][0m * BasicGuider 143:140:
2026-08-29T11:18:49.182852 - [1m[31m[ERROR][0m - Required input is missing: model
2026-08-29T11:18:49.183071 - [1m[31m[ERROR][0m - Required input is missing: conditioning
2026-08-29T11:18:49.183213 - [1m[31m[ERROR][0m * SamplerCustomAdvanced 143:139:
2026-08-29T11:18:49.183326 - [1m[31m[ERROR][0m - Required input is missing: sampler
2026-08-29T11:18:49.183431 - [1m[31m[ERROR][0m - Required input is missing: sigmas
2026-08-29T11:18:49.183529 - [1m[31m[ERROR][0m * VAEDecode 180:
2026-08-29T11:18:49.183656 - [1m[31m[ERROR][0m - Required input is missing: vae
2026-08-29T11:18:49.183788 - [1m[31m[ERROR][0m Output will be ignored
2026-08-29T11:18:49.184394 - [1m[31m[ERROR][0m Failed to validate prompt for output 18:
2026-08-29T11:18:49.184541 - [1m[31m[ERROR][0m * BasicGuider 125:14:
2026-08-29T11:18:49.184674 - [1m[31m[ERROR][0m - Required input is missing: model
2026-08-29T11:18:49.184783 - [1m[31m[ERROR][0m - Required input is missing: conditioning
2026-08-29T11:18:49.184905 - [1m[31m[ERROR][0m * SamplerCustomAdvanced 125:12:
2026-08-29T11:18:49.185028 - [1m[31m[ERROR][0m - Required input is missing: sampler
2026-08-29T11:18:49.185125 - [1m[31m[ERROR][0m - Required input is missing: sigmas
2026-08-29T11:18:49.185230 - [1m[31m[ERROR][0m * VAEDecode 176:
2026-08-29T11:18:49.185665 - [1m[31m[ERROR][0m - Required input is missing: vae
2026-08-29T11:18:49.185922 - [1m[31m[ERROR][0m Output will be ignored
2026-08-29T11:18:49.186593 - [1m[31m[ERROR][0m Failed to validate prompt for output 134:
2026-08-29T11:18:49.186814 - [1m[31m[ERROR][0m * BasicGuider 133:130:
2026-08-29T11:18:49.186940 - [1m[31m[ERROR][0m - Required input is missing: model
2026-08-29T11:18:49.187056 - [1m[31m[ERROR][0m - Required input is missing: conditioning
2026-08-29T11:18:49.187315 - [1m[31m[ERROR][0m * SamplerCustomAdvanced 133:129:
2026-08-29T11:18:49.187440 - [1m[31m[ERROR][0m - Required input is missing: sampler
2026-08-29T11:18:49.187544 - [1m[31m[ERROR][0m - Required input is missing: sigmas
2026-08-29T11:18:49.187655 - [1m[31m[ERROR][0m Output will be ignored
2026-08-29T11:18:49.554155 - [32m[INFO][0m SELECTED: input1
2026-08-29T11:18:49.581920 - [32m[INFO][0m Found quantization metadata version 1
2026-08-29T11:18:49.834273 - [32m[INFO][0m VAE load device: cuda:0, offload device: cpu, dtype: torch.float16
2026-08-29T11:18:49.883951 - [32m[INFO][0m Requested to load MiniMaxH3VideoVAE
2026-08-29T11:18:50.030492 - [32m[INFO][0m Model MiniMaxH3VideoVAE prepared for dynamic VRAM loading. 2677MB Staged. 0 patches attached. Force pre-loaded 128 weights: 692 KB.
2026-08-29T11:18:51.318474 - [32m[INFO][0m Requested to load MiniMaxH3TEModel_
2026-08-29T11:18:51.472509 - [32m[INFO][0m Model MiniMaxH3TEModel_ prepared for dynamic VRAM loading. 25882MB Staged. 0 patches attached. Force pre-loaded 310 weights: 1272 KB.
2026-08-29T11:19:03.787112 - [1m[31m[ERROR][0m !!! Exception during processing !!! BasicGuider.execute() missing 1 required positional argument: 'model'
2026-08-29T11:19:03.788883 - [1m[31m[ERROR][0m Traceback (most recent call last):
File "E:\Ai\ComfyUI_windows_portable\ComfyUI\execution.py", line 545, in execute
output_data, output_ui, has_subgraph, has_pending_tasks = await get_output_data(prompt_id, unique_id, obj, input_data_all, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb, v3_data=v3_data)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "E:\Ai\ComfyUI_windows_portable\ComfyUI\execution.py", line 344, in get_output_data
return_values = await asyncmap_node_over_list(prompt_id, unique_id, obj, input_data_all, obj.FUNCTION, allow_interrupt=True, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb, v3_data=v3_data)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "E:\Ai\ComfyUI_windows_portable\ComfyUI\execution.py", line 318, in asyncmap_node_over_list
await process_inputs(input_dict, i)
File "E:\Ai\ComfyUI_windows_portable\ComfyUI\execution.py", line 306, in process_inputs
result = f(**inputs)
File "E:\Ai\ComfyUI_windows_portable\ComfyUI\comfy_api\internal\__init__.py", line 149, in wrapped_func
return method(locked_class, **inputs)
File "E:\Ai\ComfyUI_windows_portable\ComfyUI\comfy_api\latest\_io.py", line 1990, in EXECUTE_NORMALIZED
to_return = cls.execute(*args, **kwargs)
TypeError: BasicGuider.execute() missing 1 required positional argument: 'model'
2026-08-29T11:19:03.790861 - [32m[INFO][0m [32mPrompt executed in 14.60 seconds[0m
2026-08-29T11:20:32.387632 - [32m[INFO][0m got prompt
2026-08-29T11:20:32.396335 - [1m[31m[ERROR][0m Failed to validate prompt for output 144:
2026-08-29T11:20:32.396520 - [1m[31m[ERROR][0m * BasicGuider 143:140:
2026-08-29T11:20:32.396670 - [1m[31m[ERROR][0m - Required input is missing: model
2026-08-29T11:20:32.396892 - [1m[31m[ERROR][0m - Required input is missing: conditioning
2026-08-29T11:20:32.397042 - [1m[31m[ERROR][0m * MiniMaxH3ReferenceToVideo 5:
2026-08-29T11:20:32.397172 - [1m[31m[ERROR][0m - Required input is missing: vae
2026-08-29T11:20:32.397308 - [1m[31m[ERROR][0m * SamplerCustomAdvanced 143:139:
2026-08-29T11:20:32.397444 - [1m[31m[ERROR][0m - Required input is missing: sampler
2026-08-29T11:20:32.397582 - [1m[31m[ERROR][0m - Required input is missing: sigmas
2026-08-29T11:20:32.397696 - [1m[31m[ERROR][0m * VAEDecode 180:
2026-08-29T11:20:32.397803 - [1m[31m[ERROR][0m - Required input is missing: vae
2026-08-29T11:20:32.397907 - [1m[31m[ERROR][0m Output will be ignored
2026-08-29T11:20:32.398351 - [1m[31m[ERROR][0m Failed to validate prompt for output 18:
2026-08-29T11:20:32.398470 - [1m[31m[ERROR][0m * BasicGuider 125:14:
2026-08-29T11:20:32.399311 - [1m[31m[ERROR][0m - Required input is missing: model
2026-08-29T11:20:32.399543 - [1m[31m[ERROR][0m - Required input is missing: conditioning
2026-08-29T11:20:32.399662 - [1m[31m[ERROR][0m * SamplerCustomAdvanced 125:12:
2026-08-29T11:20:32.399790 - [1m[31m[ERROR][0m - Required input is missing: sampler
2026-08-29T11:20:32.399905 - [1m[31m[ERROR][0m - Required input is missing: sigmas
2026-08-29T11:20:32.400010 - [1m[31m[ERROR][0m * VAEDecode 176:
2026-08-29T11:20:32.400122 - [1m[31m[ERROR][0m - Required input is missing: vae
2026-08-29T11:20:32.400216 - [1m[31m[ERROR][0m Output will be ignored
2026-08-29T11:20:32.401090 - [1m[31m[ERROR][0m Failed to validate prompt for output 134:
2026-08-29T11:20:32.401223 - [1m[31m[ERROR][0m * BasicGuider 133:130:
2026-08-29T11:20:32.401350 - [1m[31m[ERROR][0m - Required input is missing: model
2026-08-29T11:20:32.401494 - [1m[31m[ERROR][0m - Required input is missing: conditioning
2026-08-29T11:20:32.401644 - [1m[31m[ERROR][0m * SamplerCustomAdvanced 133:129:
2026-08-29T11:20:32.401781 - [1m[31m[ERROR][0m - Required input is missing: sampler
2026-08-29T11:20:32.401894 - [1m[31m[ERROR][0m - Required input is missing: sigmas
2026-08-29T11:20:32.402012 - [1m[31m[ERROR][0m Output will be ignored
2026-08-29T11:20:32.422823 - [1m[31m[ERROR][0m !!! Exception during processing !!! No images provided
2026-08-29T11:20:32.424147 - [1m[31m[ERROR][0m Traceback (most recent call last):
File "E:\Ai\ComfyUI_windows_portable\ComfyUI\execution.py", line 545, in execute
output_data, output_ui, has_subgraph, has_pending_tasks = await get_output_data(prompt_id, unique_id, obj, input_data_all, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb, v3_data=v3_data)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "E:\Ai\ComfyUI_windows_portable\ComfyUI\execution.py", line 344, in get_output_data
return_values = await asyncmap_node_over_list(prompt_id, unique_id, obj, input_data_all, obj.FUNCTION, allow_interrupt=True, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb, v3_data=v3_data)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "E:\Ai\ComfyUI_windows_portable\ComfyUI\execution.py", line 318, in asyncmap_node_over_list
await process_inputs(input_dict, i)
File "E:\Ai\ComfyUI_windows_portable\ComfyUI\execution.py", line 306, in process_inputs
result = f(**inputs)
File "E:\Ai\ComfyUI_windows_portable\ComfyUI\custom_nodes\ComfyUI-VFI\nodes.py", line 105, in interpolate
raise ValueError("No images provided")
ValueError: No images provided
2026-08-29T11:20:32.425954 - [32m[INFO][0m [32mPrompt executed in 0.02 seconds[0m
```
## Attached Workflow
Please make sure that workflow does not contain any sensitive information such as API keys or passwords.
```
Workflow too large. Please manually upload the workflow from local file system.
```
## Additional Context
(Please add any additional context or steps to reproduce the error here)
hi, for now just disable RIFE in the options toggles on the left side. I need to figure out why it stopped allowing image bypasses. RIFE is completely unrequired for the workflow so turning it off is completely fine. it's toggleable.
I had the same error. Worked fine in the previous version. Claude fixed it for me by changing the behaviour of the Any switch node.
RIFE crashes because a bypassed finalization chain makes the upstream Any Switch output None rather than a dropped link, so RIFEInterpolation raises ValueError("No images provided") and kills the queue instead of the branch being skipped.
@maartmo216 this is completely fixed in v1.2.1 released last night. it just needed a failsafe so it always had access to images even when only first stage is running (and rife wouldn't be utilized).
Thank you so much for sharing! I encountered the following error when trying to open your workflow, even though my ComfyUI is already updated to the latest version. Any idea what might be causing this? “Loading aborted due to error reloading workflow data
TypeError: t.findInputSlot is not a function
This may be due to the following script:
/extensions/ComfyUI-Impact-Pack/impact-pack.js”
That would simply be that your impact node pack isn't up to date, or is missing a python dependency preventing its functioning. Impact pack is a popular node suite that is used for some of the latent processing in the workflow.
For me your workflow has been the best in terms of technical ability and ease of use. Great work!
I truly don't understand why my system hates this workflow so much. And really, anything with sparse/solattention. It takes my 16 gb vram card 12 minutes just to generate a single low res 0.2 mp clip (10 seconds) with all the optimizations installed and the latent upscale part doubles that time. Seems like everyone else is seeing massive improvements with this WF and my system just chokes on it every time.
maybe you have some launch settings or graphics card settings (like nvidia app) that prevent file swapping? or it's going into an extremely slow storage? in worst case try a fresh install of comfyui and see if it's the comfyui itself the problem or your hardware/windows setup
love the workflow, thank you very much.
Ver 1.21 had Minimax H3 Latent Upscaler (3D) setting on cuda and it was giving an error. the last workflow had fp16 and i switched it to said fp16 and it works now
@teiji25762 safe to say you found it? I'm referencing your post in your other review that says you fixed it.
@foxydits It was a different issue from this. But I found it and changed it as this user said and no more errors. Thanks!
im getting this error when tring to use the video continutation. how do i fix this?
"This workflow was created with a newer version of ComfyUI (0.33.4). Some nodes may not work correctly.
Core nodes from version 0.33.0:MiniMaxH3AddGuide"
Hi, the most recent comfyUI update should address this. It seems as if you're using the briefly existing version of comfy where they removed/altered the Minimax H3 Add Guide node that's supposed to exist alongside all the other core Comfy nodes.
@foxydits but i am already using the latest version of comfy ui 0.33.4 and am still getting this error. i dont know how to fix it.
Are you able to use this for I2VA or is it strictly for ref2va? I got good results using ref2va but just wondering if anything other than that can be used for this workflow.
You have it backwards, this workflow exclusively use fl2va. But you will find that the model is plugged into the ref2va h3 node, which makes it function virtually identically to ref2va, but is using the fl2va model. You'll have to try it if you don't believe me; they're actually 98% the same model and work exactly the same way, but fl2va just produces better visuals/audio and works with LoRAs. if you want classic i2v/fflf, there are toggles in the ref area titled: "<Picture 1> is FIRST FRAME", "<Picture 2> is LAST FRAME", which makes the workflow function like the normal fl2va implementation.
Awesome workflow! Works great even without Sol-Attn (haven't figured out how to install that with trident yet, docs don't exist yet I guess). The video extension flow is really cool, amazing work!
Noob question: on the video extension flow, is it possible to get the new clip saved out along with the extended video?
Oh you mean like, it saves just the gen + also saves the stitched OG+gen? I thought about adding a "don't stitch OG video to new gen" option but things were getting really complicated under the hood as it was so I decided to leave it for now. I do want to add that feature though, it'll just have to be in v1.3
@foxydits Sounds good, I'll look forward to it. This really is an awesome workflow, my little 8GB 3070 is doing way more than you'd think it could. Thank you!
"Hey! Just wanted to share a quick heads-up on the performance between Filmmaking 2.0 and SEED HUNTER 1.2.1
Honestly, the Filmmaking 2.0 workflow runs way faster for me! Here’s what I got on my setup (RTX 4070 Ti, 32GB RAM):
Test: 10s Ref2VA video (0.5 MP, 12 steps, ref2v_lightx2v_turbo_4step_v0.1 Turbo LoRA, 3 ref images + 1 ref audio)
Filmmaking 2.0: ~430s or less
SEED HUNTER: 590s to around 12 mins (even with skip upscale ON and preview 1 only)
A couple of things I noticed after testing:
Spectrum Nodes: When I manually dropped the Spectrum nodes back into SEED HUNTER, the time dropped straight down to 480–500s. Huge difference!
Memory Cleanup: 2.0 automatically clears out both VRAM and system RAM after each run, but SEED HUNTER seems to hang onto it, which slows things down on back-to-back runs.
Hope this feedback helps, and thanks a lot for building these amazing workflows!"
Are you comparing apple with apple? For example, did you use the same turbo model and steps in Seed Hunter? Also, Spectrum reduce time a lot but it comes with a noticeable quality cost.
@teiji25762 Not in my experience. I did a lot of side by side comparisons. It varies, a lot of the time the details improve with spectrum on for some reason.
Love it. Thank you so much!
Simply the best wf for minimax
Another W by the GOAT. Figured I'd try the V2V replace feature but it just didn't do anything? what needs to change for that to work right?
(use Pic as 1st frame? Keep Ref image 1? Increase steps? Change models?)
Thanks for the WF. I'm trying to muck around with video continuation but the node is throwing type not supported. The original video was created using the same WF with H264 as the encoding type. Is there a specific encode format that works or am I missing some kind of codec install?
What node is throwing that error?
@foxydits Load Video UI. It can detect the length of the input video that I want to continue but it shows "No Video with Format Type Supported and MIME Type Found". When I try to run the WF anyway, it would say file not found in the console.
@Cowgary I see. The video loader node being used just uses whatever your browser is capable of displaying. If you right click on a loaded video, you'll just see your browser's context menu pop up, not comfy's. So it comes down to your browser's decision there to not recognize the video. Privacy settings, browser settings, codecs installed.. those apply here. If you see the length and can adjust the blue bar, that means comfy can use the video. Your browser is simply just not capable/allowed to display it through its own iframe source.
This is a great WF!!
Guys,
Where is 'Enable Final Pass' ?? 😵
The red toggle that says "Enable Final Pass" near the top right of the workflow. There's pretty clear instructions on the top section that have you follow along moving from left to right.
Hi got this error:
# ComfyUI Error Report ## Error Details - Node ID: 125:12 - Node Type: SamplerCustomAdvanced - Exception Type: RuntimeError - Exception Message: RuntimeError: shape '[1, 24, 1, 1, 19, 2, 10, 2]' is invalid for input of size 19152
One of the files you're loading -- model, text encoder, video, audio vae, is mismatching. Shape errors usually involve your text encoder or vae. Make sure the files you have for minimax are correct.
Used your old LTX Seed Hunter workflows. The upscaler phase on this one is broken. I generate three videos, enable upscale pass, click the random seed button, and it starts regenerating the first three videos again. Didn't touch a single other option or make any changes.
Tried it on two different PCs. Just not working.
Generations are also slower in your WF than identical settings in different ones.
Also why some weird RIFE node no one uses instead of ComfyUI-Frame-Interpolation ?
Don't get me wrong, this is an awesome and amazing WF still. Just reporting issues as I see them.
I generate three videos, enable upscale pass, click the random seed button, and it starts regenerating the first three videos again. Didn't touch a single other option or make any changes.
This is simply impossible. ComfyUI is programmed to not repeat already performed work. The proof of this is in the fact that thousands of people use my workflow yet you're the first to accuse it of being broken in this way. Either your settings are atypical, or you changed something without realizing that comfy thinks requires regenning, or you're simply thinking it's regenning when in fact comfyUI quickly reloads the first 3 previews before progressing to upscale (it's not a regen, it's just refreshing the displays but it goes through each 1st stage preview anyway, and takes about 5-10 seconds).
The RIFE built into comfy doesn't let you set target FPS, only scale. I have users that specifically request 60 FPS, which is otherwise unobtainable.
You can complain about the speed compared to single-pass workflows, sure. Single pass workflows don't use resources like this WF does to generate options to pick from. It'll always be slower than single pass to perform latent upscales. But what isn't slower is getting a good gen at 1080p. 2-3 lower res samples to pick one that's what you wanted from your prompt, and upscaling to 1080p? Good luck single passing that and not wasting time regenning half way through when you see in the preview it looks bad/isn't following prompt.
I'm also getting the same error: I generate three videos, enable upscale pass, click the random seed button, and it starts regenerating the first three videos again. Didn't touch a single other option or make any changes.
@austregesilo_epaminondas It's always user error. If you clicked the RED manual random seed button (above final upscale, not the green one that starts first pass), didn't change anything prior such as prompt, length, MP, etc., comfyUI will not regenerate 1st stage. It will briefly look like it, where it reloads each stage's VHS node, but it'll quickly move to 2nd pass.
One of the best and quick workflows I found that can extend videos really seamlessly. Other solutions for some reason isnt working well with me and tends to degrade and shift the color alot. thanks
hi , i need help , i got this error ran out of gpu memory . help me fix this - Allocation on device 0 would exceed allowed memory. (out of memory) Currently allocated : 8.56 GiB Requested : 1.75 GiB Device limit : 12.00 GiB Free (according to CUDA): 0 bytes PyTorch limit (set by user-supplied memory fraction) : 17179869184.00 GiB
I apologize, it would be really difficult to use this workflow with less than 12-16 GB VRAM. You'll need to use startup options meant for low VRAM, such as --disable-pinned-memory and the such, turn on all VRAM saving options in the workflow, and be careful about how much you try to upscale.
@foxydits thanks for the feedback. i tried again and it worked. i had to make resolution from 0.3 to 0.2 and at upscale it worked smoothly. let me tell you am using this on my RTX 3060 12gb vram and 16gb ddr4 ram . it takes me 53 mins to generate 15 seconds clip ( i tried without lora ) . your workflow is fantastic all in one and perfection face , body consistency. next i will update you with lora turned on. let me share the video i created...
any tutorial how to use it? im such a noob
自分もわかりませんでしたが、 youtube の動画見ると書いてありますよ
Yea, it's not really like a 100% indepth tutorial but my introduction vid to the workflow is here: https://www.youtube.com/watch?v=H8JSzhkOmXA
I recommended to use:
steps:
6-8
difussion model:
10Eros_Max_h3_TURBO-hybrid_beta3_int8_convrot_skip_edges.safetensors
text encoder:
qwen3vl_32b_heretic_minimax_h3_nvfp4.safetensors
loras:
dont use any 4-step/8-step loras, they are included in diffucion model above (10Eros Max), if you use them, they will only worsen the results
how is the h3 eros model, by the way? significant improvements to NSFW motion/imagery, you'd say?
Tried this model. Looks pretty good in terms of character consistency and prompt adherence, however after the upscale the audio changes to complete gibberish instead of the preview audio which seems good.
Another gem. You are my new go-to for workflows. Every option I need and want, laid out in a way that is compact and easy to find. Thank you for all your hard work!
Muchas gracias por compartir!! funciona iniciando con --disable-pinned-memory para no saturarme la RAM
Really nice workflow. Love to see the end result without the grainy smear for once.
Just an issue tho, The audio in the preview is perfectly fine, but after the upscale, it degraded. Any way to resolve that?
I feel it's typically the result of Turbo LoRA use, which I just straight up can't recommend for anyone anymore given that I've seen how good the model can look/move without them. Also if you aren't using the most recent version of the Latent Upscaler, I would update. I believe there was a revision a couple days ago that fixed some audio latent handling.
Eventually I may need to add a 'use 1st pass audio' option.. but often times the upscale changes the motion enough that lipsyncing would be off slightly, or other sounds.
Thanks so much for this, amazing workflow! I'm struggling with loras though for some reason. I'm using the workflow with all the default models, I tried with references and those work really well, but adding in a lora makes the results really bunrt looking, even if I lower the strength down to 0.3, this is with a lora I usually run at 1.0 with references in my normal workflow. Any idea what the cause could be? I use a hybrid checkpoint usually so I guess I can try switching to that in your workflow to see if it helps
How burnt a turbo lora will make your output look depends on step count, strength (as you mention), sampler/scheduler combo, and shift value. Are you using the exact same of all of those between your other workflow and mine? Keep in mind that during the upscale step, there's even more steps that will use the LoRA. If you end up doing 8 step first pass, 4 step upscale pass, that's 12 steps with a turbo LoRA, adding to burn easily. Things you have to consider when using a 2 pass workflow.
@foxydits Sorry it's not a turbo lora, it's just a regular lora. Fore whatever reason switching to a hybrid checkpoint seems to have solved it for me though.
I do have another issue now though, I've generated a video just fine now using references for appearance and audio reference for the character's voice, but when I try to extend the video I'm getting an error:
- Exception Message: RuntimeError: shape mismatch: value tensor of shape [5934, 32] cannot be broadcast to indexing result of shape [6062, 32]
I if I disable the audio reference it does generate just fine though, is that just a limitation or I should I be able to use audio reference while extending a video?
Thanks so much for the help!
@kaileykk Dang, I can't replicate that error no matter what I try. And believe me I've tried since you're not the first to say you get that error when using Video Continuation. So you can use Continuation if you disable your <Audio 1>/2/3 references? I'm able to use the Continue option with all 3 audio references turned on, even Force Custom Audio option... I just tried. So I am just super confused why that would make it work for you. Is there any other info you can give me? I hope it doesn't have to do with checkpoints, since I am only using fl2va.
@foxydits I can't really think of anything, I'm using comfy kitchen, sol, rife, chunking (vram saver) and Low vram attention, so it oculd well be one of those options or a combination of them. I'll have to experiment more tomorrow and see if I can find any combination that works, I'll let you know if I do find anything though, if you haven't yourself by then.
Thanks again for the workflow though, despite that error! It's the first workflow I've used that the upscaler works for me, no idea why but I've had issues with latent upscaling and with video continuation, so I'm so glad to have a workflow thats mostly working for me, cheers!
@kaileykk Thanks for troubleshooting so much. Another user solved your error by fixing the audio VAE they had selected. So last thing is, I would just double check that your audio VAE is the right file: minimax_h3_audio_vae_fp32.safetensors
@foxydits I double checked and my audio vae is correct. Everything in the workflow works for me except the specific combination of doing video extend with an audio reference enabled. I've tried with different video lengths and audio reference lengths too.
I noticed that the error of the tensors not matching "RuntimeError: shape mismatch: value tensor of shape [5934, 32] cannot be broadcast to indexing result of shape [6062, 32]" I think (not 100% sure) but I think its that the vidoe and audio latents are different lengths, or at least the length that is expected isn't what it gets. I notice if I add a second audio reference the numbers will jump up to past 10k but are still offset. No idea if that's helpful or not.
I've tried disabling all optimizations, and changing the checkpoint I use back to fl2va bur it doesn't change anything. The error happend the instant one of the preview gens start.
Anyway I'll keep poking around but I'm mostly out of ideas for now
@foxydits Sorry for tagging again, but I managed to fix my issue. I did try the new version you just uploaded but that still wasn't working. I eneded up doing a completely fresh comfyui portable install and moved over the models and nodes to the new install and it's working now. I had been using that same comfyui portable install for a long time and just updating it every few days, so I guess it had just become bloated and something internally must have broken. I know it's not an ideal solution but if anyone else is still having the issue it's a good last resort!
Tested the new version of the workflow for a few hours and so far I get better character consistency and prompt adherence with the older version. Same settings on both and same models. Might be related to the attention changes in the new one.
The only difference that would (POTENTIALLY) make a visual quality change is the swap from PlagueKind's SLA node to Kijai's far more popular/revered Sol-Attention. Otherwise the workflows are identical in terms of model handling/processing. You may want to verify they are indeed the same settings... easy stuff to miss like the max megapixels setting on each image load node, and perhaps 'match' is selected instead of 'max' on the h3 ref node, etc.
@foxydits Checked everything thanks. I have already set the max megapixels to 0 for no limit and it is max not match on both. Everything is the same. Only difference is Sol instead of SLA. I am also using a hybrid model for both instead of F2LF model, because it gives a lot better results in likeness and accuracy. Maybe the hybrid does not like Sol attention
@josharkness449 It's possible. I'm going to do some A/B tests with/without SOL now. Perhaps the default settings aren't so good. I haven't myself used the hybrid model since fl2va produces perfect results for my productions (don't fix something that ain't broke, kind of reasoning).
I'm seeing the same thing. I have 2 characters. It seems to get the one character right all the time on this latest version but will not even come close on the 2nd. For the old version 1.0, I can get 100% accurate representations on every generation for both characters. It's pretty odd.
@freewellgear Yes, very odd considering there's no functional difference between 1.0 and 1.2 in terms of how references pass to samplers when vid continuation is off. My guess: placebo, coincidence, etc. If your settings are the same, the same seed will produce the same result on both workflow versions.
@foxydits I'm getting similar errors as well. for REF2VA, I tested using 4 references, 1st one is a location, 2-4 are characters. the 4th character would not follow the reference image at all and each gen would result in a totally different character. I will test the older version to see if I can get better picture reference.
@12EVO I found out that the 4th picture node is not connected.
@josharkness449 Ugh, I hate that you're right. That's such a dumb thing to happen. I have literally zero clue how something like that could've happened. Not like I changed the image load ref areas. Thank you for pointing this out. I will need to update the version now to fix this and let people know -_-
@josharkness449 i just discovered this as well. it was driving me nuts that every gen the 4th character would just be some random person. @foxydits otherwise great workflow as always!
@josharkness449 well that might explain it. My reference image was in image 4.
@freewellgear sorry about this. i honestly don't know how it happened. and it's no surprise so few people reported it, many users don't seem to use more than 1-2 refs.
hi, one question, since im not sure if i did something wrong. the Minimax H3 Latent Upscaler i downloaded from the link is different from the one in the tutorial video. for example, it has no mode selection, and this leads to an error when i try to run the upscaler
wha? You installed this one correct? https://github.com/LBH-123-AI/Comfyui_Minimax_h3_latent_Upscaler
That's the one linked here on civit/in the workflow. If you directly installed that one, and open up the workflow, it will definitely be the correct one with a mode dropdown/etc.
@foxydits problem solved. The frontend was stuck in an older version, so the node wasnt loading like it should.
Really great workflow! Creating and upscaling is working super nice, but I'm having an issue when I try to use seamless video continuation. Upscale pass was also on, not sure what I'm missing. Can anyone help?
- Node ID: 125:12
- Node Type: SamplerCustomAdvanced
- Exception Type: RuntimeError
- Exception Message: RuntimeError: shape mismatch: value tensor of shape [2093, 96] cannot be broadcast to indexing result of shape [299, 96]
So, weird... You can use the workflow just fine, but when Seamless Vid Continue is toggled on you get an error on the 1st pass sampler?.. Hmm. I've just spent the last few minutes trying to recreate this without any 'luck'.
Is there anything more you can tell me about the circumstances of your video? If the WF gens normally without Vid Ext, then something in the handling of that extra element is causing this for you, so it's not the typical "make sure your text encoder is correct" type of solution.
Just played around a bit and it DOES work if I select that the OG video has no audio, with audio it throws the error, if that helps, but I'll play around a bit more. Many thanks for the reply!
Sorry, I'm so dumb, I mistakenly selected a video vae for the audio. Works flawlessly now. 😅
@dttk0009 Okay, well, maybe that's what others are doing too. Thanks for letting me know.
When gen 15 seconds 0.3mp to 1mp getting OOM at upscale. 5070ti 16 GB VRAM, 32gb RAM, please advise.
This model is a memory hog. Make sure the VRAM optimization options are both turned on, and that your latent upscale node is fully updated with its enable_temporal_chunking and force_unload both on. That gives you the best chances. But the reality is, 15 seconds is the model's max, and 1.0 MP is 768p, so the latent space is taking up a lot of memory leaving very little room for the model to be loaded at the same time. It's a difficult ask. You may be better off running a single pass (it's an option in this workflow) at 1.0 MP, since at least then the latent isn't being manipulated so much.
@foxydits With all these chunking things and SLA attention i set instead of spectrum thing, with enabled kitchen attention, i was able to make only 10 seconds of video upscale from 0.3-1mp. About those things like low vram attention, will increasing of head_chunks help me out? Or in Chunk feed forward node- i'll try to change chunks from 4 to 8 to see if any changes happen, but what about seq_threshold? Should i leave it at 2048? Also, i have to use second upscale step as i'm using video continuation mode
@foxydits I think i'll try to generate with upscale to target MP, with 1mp, 8 seconds 0.3 to 1mp gives 35 sec/it at upscaling
i get an error when i want to use a custom audio reference. then i toggle "enable OG video has audio". when i try to start i get this error:
RuntimeError: shape mismatch: value tensor of shape [938, 32] cannot be broadcast to indexing result of shape [1010, 32]
without custom audio it works
Most likely the same issue I had. Have a look at your audio VAE, do you have the correct one selected?
@dttk0009 just checked it, right audio vae is selected
I have the same exact problem. The workflow functions perfectly fine with a raw generation with reference pictures and audio, but when I try using the "seamless video continuation" option with reference audio, I get the same error as you
@Mitch_Connor_420_69 Make sure you have the newest 1.2.2 version that I uploaded an hour ago. But yeah, nobody's been able to give me any method to recreate this error. I've tried every permutation I could think of from several users who experienced this, and it never errors for me or many others I've checked with. For a couple of people it was because they were using weird model files or didn't properly select them in the model loading section (such as selecting the video vae for the audio vae slot). But it remains a mystery unless it's fixed in my newest version which did correct an audio related bug that MAY be related.
@foxydits Thanks for the reply, updating the workflow wasn't the issue but I did figure it out.
Basically I had made a new Comfy install for Minimax when it came out, but I set it up to use the nodes, checkpoints and LORAs found in my old install. Eliminating the node callout from the "extra model paths" file and freshly installing the nodes into the new setup solved the problem
@Mitch_Connor_420_69 wow I'm glad you figured that out, I would have never even considered the idea as a solution...
So good.
Hi ones question is it better genning everything once with seamless, an then enable pass usocale, or upscale it everytime each gen?
Work gud I like <3
glad you like, and thanks for the tips 'n stuff!
我添加了一段10秒的视频，希望再生成一段连续的10秒视频，我生成出来的视频无法和前10秒的视频保持连续，生成的视频有一小段重复，而且画面会跳动一下
就是从22帧开始跳动了一下
Make sure the dimension of the original video and the new one you're generating is the same.
For your white haired character, what image generator did you use?
Just wondering, can we do latent upscale with seamless video extend?
I am getting massive error when trying seamless extension with upscale, if i disable upscale it will work though
# ComfyUI Error Report ## Error Details - Node ID: 122 - Node Type: ImpactSwitch - Exception Type: TypeError - Exception Message: TypeError: int() argument must be a string, a bytes-like object or a real number, not 'NoneType' ## Stack Trace ``` File "E:\ComfyUI\ComfyUI_windows_portable\ComfyUI\execution.py", line 511, in execute required_inputs = await asyncmap_node_over_list(prompt_id, unique_id, obj, input_data_all, "check_lazy_status", allow_interrupt=True, v3_data=v3_data_lazy)
Great workflow! However my 'Select Latent to proceed to Finalization' node seems to be broken, all 3 toggles are permanently set to 'yes'. Which then throws an error so I cannot upscale any vids. How can I fix this?
Toggle all of them to No first. Then toggle one of them to Yes.
@teiji25762 They do not toggle, even when clicked.
@jamessumble512 Turn off Nodes 2.0 in comfy settings, then restart. Make sure your rgthree plugin is updated, make sure comfy is updated. They do work, it's just configuration on your comfy preventing it. Nodes 2.0 is the usual culprit.
قمت بتجربته
سير عمل مذهل
سريع جدا
جوده عالية
شكرا
Where is the load image crop node from? manager can't find it
Look at the bright red note in the workflow on the left side titled "CUSTOM NODE LINKS"
Noticing that my outputs keep on getting more and more oversharpened/grainy as I proceed further into the video. Is there any way to avoid this? I've been starting with 1 image and 1 audio reference, upscaling the desired output, then leaving everything as is except for adding in the new video to continue from and changing up the prompt to the next scene. Should I be removing the image reference once the first gen in the sequence is completed?
I also should mention that I'm using a turbo lora which may be the reason too. I noticed you mentioned in your video that they can sometimes produce a "burnt" look
If you have high quality character refs / maybe a pic of the scene/location, make sure to load that in while you gen and use good prompting principles to assign the character(s)/location to those images. If you're continuing a video of a girl running, have a high quality pic of that girl in <Picture 1> (or a ref sheet), and assign in your prompt <Subject 1> is the girl in <Picture 1>, and in your detailed_description section describe: The video begins with <Subject 1> running, and blah blah blah.. your prompt goes on. There will be no degradation if the model has a high quality reference.
@foxydits Ah I see. Instead of a reference image I have more of a "frame 1" type of image loaded. That might explain it
This is a really nice workflow. But please consider adding automatic stitching of clips. So you can make a long video like this with just one prompt:
prompt for shot 1
---
prompt for shot 2
---
prompt for shot 3
etc
and it will continue to do all shots after another in one go. I think it's really useful because you can queue big jobs at once of a video of a few minutes long, instead of generating a video, and then manually putting it into the workflow and creating the next part etc.
There is another workflow on civitai called "h3 multishot seamless chain" that has it like that. You could look at how they did it and use it for inspiration. But your workflow is better in other ways so thats why i will use yours.
Sorry, there are workflows that specialize in that, but mine isn't meant for such inane tasks. That is never how a filmmaker would need to operate. My workflow is for AI film production, and people making serious work don't typically want to step away and let 3+ gens tag off of each other. You would put care into each gen, watching the seed hunting and picking the best one to continue with, and use a proper video editor to master the visual and audio.
The WF does automatically stitch whatever you extend though. So if your OG video is 15 seconds, and your gen is 15 seconds, what pops out is 30 seconds. That's as far down that rabbit hole as I want to go, since no serious filmmaker needs more functionality than that, as you could then easily chain that to a 45 second video, 60 second video. It's just not gonna do it in one long prompt chain like you suggest.
@foxydits Ok it's up to you, every single generation you do is "serious film making"? instead of just experimenting for fun a lot of the time? It's only an option. I'll probably try to figure out how to do it myself based on how other workflows did it. Thank you for making such a great workflow though, it is really useful.
@Kambai I do experiment and have fun a lot. But this workflow is a tool designed to meet the needs of high level AI filmmaking first and foremost. The video continuation is a useful tool to extend a single shot past 15 seconds, which has a use-case in filmmaking (like with my Effy's Earth Vacation Vlog series being a long hand-held selfie style video). Most films have cuts and so you produce shots gen-by-gen and use a video editor to compile them into a short film. Each gen uses references for characters and scenes and even past screenshots from previous gens as blocking references. There's no quality control when you just chain 4+ prompts together and have it tie everything manually. Quality work is produced in modular pieces, with high control and precision, in my opinion.
Why don't you use Film Net for frame interpolation? Isn't it better than RIFE?
Yes, but a lot slower and people complain whenever I put it in my workflows. I literally switched to it, got 5 complaints the next day, and then switched it back on my last workflow for minimax lol. People just want speed and the ability to choose their interpolation target FPS. And, so far the quality hasn't been too bad. RIFE used to suck but not so much anymore. I think others are still better though.
@foxydits Haha! yeah this one is also working fine so far. And Thanks for this amazing workflow. For motion control/ transferring motion from a video reference to image, do we need to switch to ref2va model?
@AnomalyXOXO No, the fl2va model IS the ref2va in disguise. There's virtually no identifiable difference outside that fl2va is slightly higher quality looking and sounding. I encourage you to A/B test though; run it with fl2va, and then same seed ref2va and see which you like better consistently.
@foxydits THAT...is interesting.
What was updated from 1.2.1?
version notes:
- Fixed 4th Image: It wasn't connected to anything so your 4th reference image wouldn't apply. Also corrected an audio connection bug that could lead to errors depending on your configuration.
hi you disabled the sol? was that thing that was causing ghost people walking behind in distance and object shifts from the background?
No. I just disabled it by default 'cause too many people were asking questions about how to install it (requires custom node download / installing triton-windows). I use it myself. I just have to balance what's enabled by default with what I think the beginner/intermediate crowd can handle. I'm already asking a lot with this behemoth of a workflow.
Is there a way to automatically stitch the videos together for the continuations? Or must that always be done manually?
Wait no I see it's linked to the upscaling block, just needed to selectively bypass the upscaling and leave the extension bookkeeping block.
@zomgawesomeness the workflow by default stitches the original video to the beginning of the new gen. So if you load a 15 second OG video, gen a 10 second video, you get a 25 second video final output. You can reduce this by dragging the blue bar's left handle in the video trimmer, because that'd become the new start to the OG video.
Hi, is there any alternative to the latent selector node that you use?Not that I know of. Why, you can't install impact pack for some reason?
Hi, yeah, I can probably install it. But I wasn't sure about taking the risk, since there have been issues with the Impact Pack in the past—viruses and the like. Is it safe to use? That’s why I was asking if there was another node I could use that works just as well.
@Bananaai Viruses from Impact pack? It's a super popular node. Thousands of people use it.
Can you tell me what the problem might be? When I merge the videos, the resulting video has no sound—the audio is only present in the last 5 seconds. Even though the audio is there originally, and the song from the previous video continues.
First of all, thanks for sharing this workflow! the first pass previews work fine for me (around 2 minutes per gen for a 15s clip with 1:1 ratio). However, the upscale to 1.0 mp (4 steps) took me 2 hours to complete. Is that normal for a 4070 with 64gb ram?
Im not the author but no. even my setup (9060xt 16GB) can upscale faster then that.
That would only happen if your system is overwhelmed. Make sure you're not trying to upscale like 10+ seconds to 2 MP unless you have a huge beefy system. This model's a resource hog. On my 3090 with 24 GB VRAM, 64 GB system ram, I can do about 6-8 seconds of 1.6 MP, and 15+ seconds of 1.0-1.2 MP without my GPU stalling out like yours did. Gotta find what your ceiling is.
Hey brother! Enjoying your workflow as always but I came across a possible "bug"? Im trying to just directly upscale my 1st sampler preview by just keeping the "enable upscale pass" node on, but whenever I run it the upscaled video is a bunch of weird blocks and colors. I tried switching the "Reroll upscale pass" seed to randomize but that still didn't fix it. Any clue? Or is the workflow not meant to take the preview directly into the upscaler? Any help would be appreciated.
No bug, you can absolutely just leave Upscale Pass on and go from 1st pass directly to 2nd pass with just preview #1 turned on. Basically that's just like a normal 2-pass workflow at that point, and I use that configuration quite often. My strong suspicion is that you've just got a misconfigured option selected somewhere. Make sure your "Selected Latent" is #1, if you select latent #2 when it's turned off you will get random stuff like that in upscale pass.
@foxydits Hmm Ill give it a try with the new update. The only modification I had made was switching the base model to something else and no use of a Lora. Ill keep experimenting tho.
@foxydits Nope still getting that issue. Im running 1.5 right out of the "box" and the first gen with upscale enabled runs well, then the second time I run it with a new prompt, the 1st pass looks good for 0.4mp, but the final one is an artifact mess with random music and overbaked filter on everything. Do I need to generate a new fixed seed every time or keep both on random? What exact model and turbo (if any) do you use? Trying to figure out what could be the cause.
@123sirako123521 No need for any seed changes. Something is wrong elsewhere, some comfy/python configuration isn't playing nice with my workflow. You're the only one reporting this on any version, and as you say, it works once and then breaks. It's not like the workflow changes in between runs... so without having any clear indication of what could be wrong, all I've got as a suggestion would be to make sure you've run your update_comfyui_and_python_dependencies.bat file (assuming you're running Portable like you should be). The other thing is update your GPU driver if you haven't in a while. who knows what wacky shit is going on behind the scenes to make the wf work once, and then never again, for just you.
Thanks a lot for your wf ! I also used the LTX ones you did. Regarding video stitching I think there is a memory issue with Kijai node, may be use an alternative for next versions ? TIA !
Which Kijai node? You talking about the Image batch stitching of OG + new gen? I've gotten 100% ram hang ups from it myself but don't know of any alternatives to bring two batches of images together without forcing ppl to download yet another node.
@foxydits Yes, this one. There are other custom nodes doing it I believe (the one I know takes the latent, not the images though), but indeed it would mean one more dl of a custom node.
Is there any way to get around color and contrast degrading after 3rd-4th video?
Originally I assumed that if you use references for your characters that the model would 'self-correct' -- but it's simply not true as I use this more. The only way to do it is to pin a last-frame for the video (use the '<Picture 2> is the last frame option' or do it via prompting) so the model is forced to come to a solid keyframe that is high quality and not overcooked. I am also working on this issue.
@foxydits So like, save last frame and use it as reference and mention it in prompt as first frame pic? Would it affect the video continuation be seamless?
However you might generate a final keyframe for the video, yeah. Minimax's goal is to get to that last frame, and it'll update the contrast/color/quality to do it. It sort of forces the video to reset its quality. The main problem being, you have to generate a high quality keyframe of the last shot in the video. There are plenty of ways to do it, but it takes effort as you might imagine.
奇怪的事，为什么我禁止了背景音乐，还是会有，而且还有歌声，我一段设置是8秒，当我选择5秒，就不会有音乐
Does this use multi image refrence alone to stitch the videos, or does it use the same kind of thing as WAN SVI?
Damn i wish i had the patience to let 3 videos gen at once.
You can disable previews #2 and #3 with their option toggles, don't have to let all 3 gen. I often just keep preview #1 turned on when I'm sitting at my computer.
The only thing this uses to stitch the videos is the last 5-39 frames of your previous video added into the minimax latent via the Add Guide node. Then it's just a matter of including the OG video at the beginning of the final output video that drops out.
The Final Upscale Pass alters the preview video to a noticeable degree. Is it possible to denoise less or something like that? Or is there an option I've overlooked / overcooked that would alter the preview clip? Either way, it's still an awesome WF combined with minimax_h3_fused_refdelta_r1024_turbo08_mystic07_int8_convrot at 4-steps :-)
I would suspect the final upscale would indeed alter a lot if your 1st pass is only 4 steps. If you're doing production quality work (original fl2va model + 0.4-0.5 MP 1st pass + not using Turbo LoRAs), 1st pass should be somewhere between 8-12 steps, with a final pass around 4-8 steps. With a high quality fully denoised 1st pass, you don't get big changes in final upscale, just details. The lower the steps in 1st, the bigger the changes in 2nd. Hope that makes sense.
@foxydits yeah, it does. Tbh, I could have taken a step back (no pun intended) to consider and come to the same conclusion, but I was so impressed with the 4-step output at the time!
@tommo5608 this excellent workflow has a single pass option as well, you can simply tune it to produce high resolution videos in the first pass and not use upscale at all. You can tune everything there. This is by far the best Minimax workflow I've ever seen, I uninstalled every other I had after finding this one.
Any chance to implement https://github.com/lihaoyun6/ComfyUI-H3VAE_TRT ?
Thank you very much. This is a work or art workflow for MinMaxH3. It's the only one I'm using and it works really good !
Great workflow! I have an issue though, after extending 2 x 15 sec clips, using the resulting 30 sec clip as the source vid to extend results in the latest 3rd clip produced to have delayed audio. like it doesn't matter if i take the full 30 seconds or just the last few seconds of the source clip, the audio for any generated video is delayed (but the audio is fine when previewing in the first step)
That would come from some misconfiguration in the stitching at the final output process. Make sure you have only one option selected in your Video Continuation group for audio.. "The OG video has audio" or "does not have audio".. if both or none are selected it causes this issue. Especially given this worked for you before, it's very clearly just a misconfig and you have to hunt down what setting needs to be changed.
To be clear, I'm about to update this workflow and it fundamentally changes how vid continuation works, so keep your eyes open for that.
I'm getting an error "can't access property "workflow", res is undefined" with this workflow. And it doesn't even start so I don't know where is the error coming from. Other minimax workflows work just fine. Any clue?
You're the first to report that one. No idea. Have you tried closing it and just re-importing the workflow? Sounds like it just got corrupted or something.
@foxydits yeah it's really weird. I'll try redonwloading the workflow.
@foxydits btw the new version works like a charm so far, much appreciated!
I'm really confused; I seem to have an issue that neither the interwebs nor GPT can solve.
I'm not a stranger to ComfyUI, and I generally like to make my own workflows, but R2V is a different beast.
I tried your latest workflow, but even the basic templates didn't seem to work properly; it basically outputs the reference video 1:1 instead of using the reference images to replace the character.
The first attempt with this workflow worked flawlessly, but every attempt after that, even after reloading your workflow (or any other), just gives me the almost 1:1 reference video back, and after a few days of fiddling, I am at a loss. Could it be comfy being a bleep? Could a GPU driver be working against me?
By now, I don't know anymore.
I2V and T2V work fine, image generation works fine, just.. not R2V.
Hardware is fine: 48-core Threadripper, 128 GB RAM, 5090, all the good stuff.
If there is anyone who knows their shiznit, and might be able to shed some light, please do.
One thing I just seem to have found out is, if I render an image, and then go back to this workflow, re-load all the images, the odds of it working fine is higher. still not perfect, but at least im not getting input-ref-video-results ALL the time. Will try more once I am back from work.
Okay, it seems Minimax doesn't always properly unloads/clears cache?
(could be a driver, comfy/minimax being a bleep) but the odds of getting a successful result are much higher when I do points to the previous reply
有什么办法让无缝视频每段的人物声线都一样吗