High praise and hugely appreciated!

I'm extremely old fashioned and handwrite EVERYTHING. The only AI I use is local open source stable diffusion models. Handwriting prompts forces you to learn and adapt!

I'll try it. But so far, anything that 'draws' the audio output from the VHS load video node when the loaded video has no audio, results in an error from the node itself. I'd use a different video loading node but that one is just so damn useful, with its ability to select start frames/frame cap/set FPS/etc.

From the official documentation:
"video editingAn existing source video is directly modified"
So something on the lines of: "[video editing] Replace the girl in <Video 1> with the blonde haired girl from <Picture 1>" then go on to describe the actions in the video, reaffirming through description that the girl from the video looks like your reference. Sometimes if reference looks too similar to video, the model doesn't bother trying since it thinks 'close enough'

Depends on what you're going for. Action scenes? Sure. Talking head? Not so much. That's why they're included in my wf as toggable options! And why I list this as an 'advanced' workflow; I expect people to know the trade-offs!

Glad to hear it! Keep it up

Lol interesting review but I'll take it. I'm dropping a new version today that is pretty exciting!

@Wallyhojo Couldn't agree more.

Yeah, when I made the workflow there were no LoRAs. I'll update today with a spot to start clicking in LoRAs. For now you can just add your own "Load LoRA" node and connect it somewhere in the model link pipeline.

It really depends on what you're trying to do. I'm building a textpad file full of little prompt nuggets for ref2va that have worked for me.

@udizjxvo 
[video editing] The target video is an edited version of <Video 1>. Replace the girl in <Video 1> with the exact girl in <Picture 1> fully_preserved. is a good starting point. It helps to describe the new deepfaked character too, and the background. Like if I were replacing a girl on the beach with an anime waifu, I would do:
[video editing] The target video is an edited version of <Video 1>. Replace the girl in <Video 1> with the exact 2D animate blonde haired girl in <Picture 1> fully_preserved. Preserve <Video 1>'s background environment in the target video. The girl dances around on the beach playfully, her blonde wavy hair flowing in the breeze as waves crash behind her.
This will motivate the model to maintain the video reference's environment/setting.

@udizjxvo If ref character and video character look too alike, I've noticed the model ignores editing it entirely. I think it says "good enough" lol. That's the only time this method hasn't worked for me. As long as your prompt spells out the differences; how your reference character looks different than the one in the video.

It's not required, it's just a speed-up helper for high resolution/high step count gens. you can disable it. it requires triton to be installed to work. if it's not working, it typically means you don't have triton installed or have an old version.

If you don't tell me what the error says I can't really help.
On their github it says:
"The node adds no third-party Python dependency. It uses PyTorch and ComfyUI modules already present in a normal ComfyUI installation."
So the only thing that could be an issue unless it's installed wrong is that your pytorch is out of date. Check your run logs for outdated cuda, too.. should be cu130 (cuda 13.0)

Ah, I see now, the most recent comfyUI update broke the node and the author hasn't updated it yet. You can use EasyCache in the meantime, it's built into the workflow.

@Sam110177 I also see that Spectrum updated the node 3 hrs ago so I'm sure they addressed the missing variable due to Comfy's update. Should be set now.

The mem efficient node is in my current version of the WF that will be updated today. Minimax h3 cache just drags behind the others. The only reason EasyCache is included is because it's comfy core and doesn't require a special node. Spectrum and T8 Block are better. I an considering switching the option to t8 since Kijai says it handles audio better.

@drfaker911219 I can do that, but in my experience testing it, I felt it wasn't that good. Like, I feel it made videos look weird in an uncanny sort of way. But sure, as an option, why not.

@dragonite9000263 <- is correct. It's comfyUI's fuckup.

You would use the ref2va workflow, load 10-30 sec of your character's voice into <Audio 1>, make sure the "Use <Audio 1> for T2V/I2V" toggle is OFF (it will force your audio latent to be 1:1 with the audio file).
In your prompt, under the subject_definitions section:
"<Subject 1> is the girl in <Picture 1>
<Audio 1> is the voice-timbre reference for <Subject 1>"

In the summary section where you are writing out your shots, at that point all you would need to do is write:
"<Subject 1> speaks towards the camera: <d>Are you gonna finish that muffin?</d>"

If your sampler/scheduler/step count are all good, and you're not using too many speedups / cache nodes, you should get a fairly convincing cloned voice.

20-25 with Spectrum or EasyCache is good. It'll seem really slow at first but then around 10 steps in you'll start seeing the cache node skipping unnecessary steps.

wowee, good catch. that's what i get for last minute additions. thank you. i will fix asap.

There isn't a good upscaler, simply put. RTX upscaler, LTX latent upscaler.. bleh. Just kinda waiting for minimax's 2k upscaler release.

If the gen length doesn't match the audio file duration closely, that can happen. It can be like 0.5 sec off but it really should be close. It works really well for me most of the time, with both t2v/i2v, and ref2va.

I don't know how comfy desktop works with this (you might just ask chatGPT), but for me comfy manager has an option to choose preview methods, and I start my comfy with the flag --preview-method latent2rgb

I get black screens from kijai's vae. I gave up lol. I use the official.

No longer available?.. what are you smokin' dude?

@Fealow I don't know anything about the way cloud gpus' handle custom nodes, but Sol-Attn is completely downloadable from github. It's Kijai's newest node, and he's still working on it because this is a cutting edge brand new model that only released like 7 days ago. When you put reviews like this out there saying "the workflow is moot because the custom nodes aren't available", you imply that would be true for everyone.. but it's not. Considering thousands of people use this workflow, it's clear this idea of nodes not being available is only a problem for you.

Hey since you listed out in numbers I'll reply to each.
1. I wasn't satisfied with RTX or LTX's upscaler, so I didn't add it. I spent a while trying to make the LTX latent upsampler worth the time, but it just wasn't. I went deep, looking at pixels, and found that yes the resolution was suddenly 1080p, but it was no more detailed or higher quality, and the motion/detail actually would go down. I will consider adding it as a subgraph now that my workflow already uses LTX nodes anyway for my custom audio hack.
2. I'm not sure what you're saying here actually. Sorry. Are you saying that video references can't have audio in my current workflow? I have a little note explaining the video-audio ref situation at the bottom right of the workflow.
3. That's true, I should add that.

@josharkness449 Oh, that has nothing to do with my workflow. That's your comfy settings. I had a similar thing at one point a long time ago and the solution was to turn off intermediate files in the settings. I don't know the setting name exactly, but you should be able to find it. Maybe in VHS settings. ask chatGPT if you can't find it.

Hmm. Very strange. It is indeed all just rgthree-comfy nodes. Nothing magical or outside the normal. I guess you can update comfyUI / python dependencies if you haven't, and when you boot it, look at the logs. Find where it tries to import rgthree-comfy and let me know what it says, there might be something like a "FAILED TO IMPORT" message.

@bernt8379944 Indeed, it is confusing. You're the first person in many thousands who have used my workflows to have reported issues with rgthree like this. It is a very popular and well-regarded essential node-pack for many workflows. I simply have no idea what could be going wrong for your installation compared to the others, EXCEPT that typically nodes not loading even though the pack is installed typically means python dependencies are wrong.

Yes, it gives me black vids too. Minimax generative audio is still rather bad. It's one of the main reasons why I included that hack (toggled on with the "Force <Audio 1> latent" option on the top left of the workflow) that lets you inject your own custom audio, which should come out clear. Just make sure your audio file is of similar length to your clip for that one though.

One warning about global sage, (--use-sage-attention) is that it breaks some models, like Krea2.

Sorry about that, it's because it should be easily installable through the comfy-UI manager which doesn't require much effort. You'd just go to comfy manager -> missing nodes and it should pull up WhatDreamsCost's node suite. If not, here you go:
https://github.com/WhatDreamsCost/WhatDreamsCost-ComfyUI

Nothing that should affect memory usage was changed. The only things changed are node cleanups, the forced custom audio option now working for ref2va as well, and Sol-Attn being plugged in instead of unlinked on accident. So it's likely being cause by some other factor.

Nothing that affects speed was changed between 1.5 and 1.6 besides Sol-attn being correctly linked (see below). My guess is you updated comfy or something is different without you noticing it between the time you used 1.5 and the 1.6 update.

The only thing I can think of between the versions is that in 1.5 the Sol-Attn wasn't connected to anything so it never worked, which was fixed in 1.6.

If you are using ref2va, it could be as simple as your input images/ref video being large (2+ MP) and that the minimax ref2va node is set to 'max' instead of 'match' which uses the full resolution references. That can lead to slower results. But the reality is, my workflow didn't change so much as everything else did; comfy updates, the nodes being updated, settings you're maybe not aware of.

I am adding it to the v1.8 version releasing in a few minutes.

@josharkness449 There's a note inside the workflow that gives links to all the nodes. It's bright red on the left side.