Small update. If you already downloaded before 3/7 there is a new distill LoRA being used. Download it here:
https://huggingface.co/Kijai/LTX2.3_comfy/tree/main/loras
If you are coming here and downloading for the first time it's already available.
LTX 2.3 just dropped and figured it was a good time to update a bit.
Models all updated for LTX 2.3, basically the same workflow as LTX 2
We are still using the old LTX2 distill lora as it is working just fine for now. The lora is HUGE at 7+gb total. Previously someone released a lower rank more manageable one which we use right now. Staying with this one for now until I can find a lower rank or make one myself.
Removed the audio nodes as those were creating a lot of problems and hopefully LTX has fixed the audio which to me it seems much better.
Removed the tvae previews, no longer working as 2.3 has a new VAE anyway. You just get blurred static but video comes out fine.
LTX claims to have fixed the portrait video so those should work now. Create those tikkie tokkies to your hearts content
https://huggingface.co/Lightricks/LTX-2.3 if you're looking for more info on the updates.
No examples yet but I'll drop a vid when I get the chance.
With this right now I'm only uploading the all in one zip which will contain the t2v, i2v, ta2v, ia2v, v2v as usual. I WANT ALL THE THINGS! will be the only one available for now.
I'm doing some testing to see if these settings are going to stay the same, change, if we still need lower strength on distill lora etc. I'll update all workflows as needed. For now all tested and working by me.
Description
FAQ
Comments (119)
whys my video coming out all fuzzy
NVM i had the old spatial upscaler
@Ragamuffin20 Thanks , your troubleshoot help me a lot!
@qbaner np, i figured somone was gonna run into the same issue
VAEoaderKJ Error: Invalid VAE: None If a VAE has occurred on the checkpoint loader node, the checkpoint does not contain a valid VAE. That's the problem...
I updated the node and it got better
I got the same error when loading LTX2.3_video_vae.safetensors with VAELoader_KJ. I already updated the latest version of KJNodes. Don't know why and how to solve it?
(but LTX2.3_audio_vae.safetensors can be loaded with VAELoader_KJ)
The solution is to reinstall KJnodes pack and make sure you have installed the dependencies.
@etherloth Doesn't work. I have uninstalled, reinstalled, updated but I am getting this same error too.
@Artmagnet I also updated the ComfyUI-GGUF node pack - check if that solves it.
@Artmagnet update Kijai Nodes and Comfyui
@OzzyOsman @etherloth I've done both things. Its still giving me this VaeLoader error.
@OzzyOsman Which unet_name are you using? I downloaded ltx 2-3-22b-dev-Q4_K_S. I have Rtx 5060 TI 16 gb
I am getting the same error. Use audio VAE for LTX-2. It seems to work with i2v workflow.
@creatorjulie743 all workflows work for me and lots of others. This seems to be the major issue with a lot of people is this VAE.
Seems like already uninstalled and reinstalled, updated and all that.
Common issues have been that, bad vae downloads, and node packs conflicting with the vae kj nodes.
Redownload the vae files. Check the vae node, should be a tag on it on the right corners make sure that says kjnodes. Another person had that node and the title was KJ but the tag was feaernworksnodes or something like that.
@Urabewe @OzzyOsman I got to work guys. I GOT IT TO WORK!!
Works great. Thank you.
I2V works like a charm, thank you! Will there be a version with an end frame input?
I have that planned, yes
Please help me understand what I'm doing wrong. When I use a GGUF model, the execution time for my scenario is 118 seconds. Running the same scenario using a large .safetensors model (43 GB) takes 77 seconds. I have 16 GB of VRAM
GGUF is really only if the fp8 won't run on your system. I get nothing but OOM.
GGUF is only small in file size it will take slightly longer due to basically being "changed" at inference.
The weights get quantized in a gguf and are "un-quantized" on the fly once inference starts so there is a bit of extra time for that.
Strange i have the same models like in hour workflow but i keep getting error on firsrt Sampler
# ComfyUI Error Report ## Error Details - Node ID: 172 - Node Type: SamplerCustomAdvanced - Exception Type: RuntimeError - Exception Message: mat1 and mat2 shapes cannot be multiplied (1024x4096 and 32x4096) TIPS: If you have any "Load CLIP" or "*CLIP Loader" nodes in your workflow connected to this sampler node make sure the correct file(s) and type is selected.
Google didnt help. Any ideas? Comfy UI and nodes Up to Date
For me the fix was using Q4 "DEV" ltx-2-3-22b-dev-Q4_K_M.gguf linked in the workflow.
The issue seemed to be with the "distilled" version I had downloaded separately.
@daxnx Try asking ChatGPT or Google Gemini. But from the error message it seems that you have incorrect model/s loaded into your text encoder nodes (Load CLIP or CLIP Loader). ComfyUI tells you as much.
Well, the issue was with the checkpoint. I searched the internet and found that someone had the same problem. Later that day, I found out that Unsloth had updated their checkpoint. I redownloaded the q6 version and it worked.
Well the good news is unlike the other one that was giving so many fits, T2V works here. Only puzzling thing is this:
Your example did output. Audio is a bit wonky and the video is a tad fuzzy, but no idea if that's related to whatever it's gagging on here.
got prompt
Found quantization metadata version 1
Using MixedPrecisionOps for text encoder
Requested to load LTXAVTEModel_
loaded completely; 11201.91 MB loaded, full load: True
CLIP/text encoder model load device: cpu, offload device: cpu, current: cpu, dtype: torch.float16
gguf qtypes: F32 (2672), BF16 (28), Q4_K (1214), Q6_K (326), Q5_K (204)
model weight dtype torch.bfloat16, manual cast: None
model_type FLUX
lora key not loaded: diffusion_model.audio_caption_projection.linear_1.alpha
lora key not loaded: diffusion_model.audio_caption_projection.linear_1.lora_A.weight
lora key not loaded: diffusion_model.audio_caption_projection.linear_1.lora_B.weight
lora key not loaded: diffusion_model.audio_caption_projection.linear_2.alpha
lora key not loaded: diffusion_model.audio_caption_projection.linear_2.lora_A.weight
lora key not loaded: diffusion_model.audio_caption_projection.linear_2.lora_B.weight
lora key not loaded: diffusion_model.caption_projection.linear_1.alpha
lora key not loaded: diffusion_model.caption_projection.linear_1.lora_A.weight
lora key not loaded: diffusion_model.caption_projection.linear_1.lora_B.weight
lora key not loaded: diffusion_model.caption_projection.linear_2.alpha
lora key not loaded: diffusion_model.caption_projection.linear_2.lora_A.weight
lora key not loaded: diffusion_model.caption_projection.linear_2.lora_B.weight
Requested to load LTXAV
loaded partially; 1335.29 MB usable, 0.00 MB loaded, 13900.45 MB offloaded, 1352.27 MB buffer reserved, lowvram patches: 0
0%| | 0/8 [00:00<?, ?it/s, Model Initializing ... ]ERROR lora diffusion_model.adaln_single.linear.weight shape '[36864, 4096]' is invalid for input of size 100663296
ERROR lora diffusion_model.audio_adaln_single.linear.weight shape '[18432, 2048]' is invalid for input of size 25165824
12%|██████▏ | 1/8 [00:54<06:21, 54.47s/it, Model Initialization complete! ]ERROR lora diffusion_model.adaln_single.linear.weight shape '[36864, 4096]' is invalid for input of size 100663296
ERROR lora diffusion_model.audio_adaln_single.linear.weight shape '[18432, 2048]' is invalid for input of size 25165824
25%|█████████████████████ | 2/8 [00:26<01:19, 13.30s/it]ERROR lora diffusion_model.adaln_single.linear.weight shape '[36864, 4096]' is invalid for input of size 100663296
ERROR lora diffusion_model.audio_adaln_single.linear.weight shape '[18432, 2048]' is invalid for input of size 25165824
38%|███████████████████████████████▌ | 3/8 [00:38<00:57, 11.58s/it]ERROR lora diffusion_model.adaln_single.linear.weight shape '[36864, 4096]' is invalid for input of size 100663296
ERROR lora diffusion_model.audio_adaln_single.linear.weight shape '[18432, 2048]' is invalid for input of size 25165824
50%|██████████████████████████████████████████ | 4/8 [00:49<00:46, 11.62s/it]ERROR lora diffusion_model.adaln_single.linear.weight shape '[36864, 4096]' is invalid for input of size 100663296
ERROR lora diffusion_model.audio_adaln_single.linear.weight shape '[18432, 2048]' is invalid for input of size 25165824
62%|████████████████████████████████████████████████████▌ | 5/8 [01:01<00:35, 11.71s/it]ERROR lora diffusion_model.adaln_single.linear.weight shape '[36864, 4096]' is invalid for input of size 100663296
ERROR lora diffusion_model.audio_adaln_single.linear.weight shape '[18432, 2048]' is invalid for input of size 25165824
75%|███████████████████████████████████████████████████████████████ | 6/8 [01:13<00:23, 11.74s/it]ERROR lora diffusion_model.adaln_single.linear.weight shape '[36864, 4096]' is invalid for input of size 100663296
ERROR lora diffusion_model.audio_adaln_single.linear.weight shape '[18432, 2048]' is invalid for input of size 25165824
88%|█████████████████████████████████████████████████████████████████████████▌ | 7/8 [01:25<00:11, 11.90s/it]ERROR lora diffusion_model.adaln_single.linear.weight shape '[36864, 4096]' is invalid for input of size 100663296
ERROR lora diffusion_model.audio_adaln_single.linear.weight shape '[18432, 2048]' is invalid for input of size 25165824
100%|████████████████████████████████████████████████████████████████████████████████████| 8/8 [01:36<00:00, 12.12s/it]
VAE load device: cuda:0, offload device: cpu, dtype: torch.bfloat16
0 models unloaded.
loaded partially; 0.00 MB usable, 0.00 MB loaded, 13900.45 MB offloaded, 1352.27 MB buffer reserved, lowvram patches: 0
0%| | 0/3 [00:00<?, ?it/s, Model Initializing ... ]ERROR lora diffusion_model.adaln_single.linear.weight shape '[36864, 4096]' is invalid for input of size 100663296
ERROR lora diffusion_model.audio_adaln_single.linear.weight shape '[18432, 2048]' is invalid for input of size 25165824
33%|████████████████▎ | 1/3 [00:27<00:54, 27.36s/it, Model Initialization complete! ]ERROR lora diffusion_model.adaln_single.linear.weight shape '[36864, 4096]' is invalid for input of size 100663296
ERROR lora diffusion_model.audio_adaln_single.linear.weight shape '[18432, 2048]' is invalid for input of size 25165824
67%|████████████████████████████████████████████████████████ | 2/3 [00:55<00:27, 27.71s/it]ERROR lora diffusion_model.adaln_single.linear.weight shape '[36864, 4096]' is invalid for input of size 100663296
ERROR lora diffusion_model.audio_adaln_single.linear.weight shape '[18432, 2048]' is invalid for input of size 25165824
100%|████████████████████████████████████████████████████████████████████████████████████| 3/3 [01:23<00:00, 27.96s/it]
Requested to load AudioVAE
loaded completely; 1883.55 MB usable, 693.46 MB loaded, full load: True
Requested to load VideoVAE
0 models unloaded.
I asked chatgpt what this error is all about, this is the answer I got: The log shows that the LTX 2.3 GGUF model itself loads correctly, but there is a LoRA compatibility issue. The errors appear because the LoRA being loaded was trained for a different model architecture or version, so its weight shapes don’t match the layers in LTX 2.3. When ComfyUI tries to apply the LoRA, it cannot reshape the tensors, so it prints errors like “invalid shape” and “lora key not loaded.” However, this is not a VRAM problem and not a fatal error—the base model, VAE, and generation process still run normally. The LoRA is simply ignored or only partially applied, meaning the user should either remove the LoRA or use one specifically trained for LTX 2.x.
@denjay5157 yes, if doing t2v bypass the detailer Lora if you're using that it makes everything fuzzy. Didn't realize until later
Lora keys are also from the distill Lora. The default is 7gb which is huge there is a much smaller one now I will link later which will fix all that.
@Urabewe I would also appreciate you linking that! Ty
Try replacing the lora with this one and see if the error goes away https://huggingface.co/drbaph/LTX-2.3-FP8/tree/main/LoRA
@denjay5157 worked perfectly, ty
@denjay5157 thanks for linking that! Actually working on the workflows right now for an update with new samplers and a lot more. Just found these resized loras which I was really hoping to find. If not I was just going to make one myself. 7GB for a lora is gigantic. Kijai also just posted one about 5hrs ago that is 2.6-ish GB I'm testing that one.
I tried using Comfy for the first time and try this workflow, but I got an error on the DualCLIPLoader that says "Expecting value: line 1 column 1 (char 0)"
Is a fresh, new installation of ComfyUI portable and I only installed ComfyUI Manager, then the 2 nodes that you specify on the I2V worklfow + the requirements for each one.
There's a question that I have, and maybe that is what's giving me the problem.
On the notes at left, Dual CLIP Connector there's a file that we have to download, and that we have to place it in "clip/text encoders" does this mean "text 'encoders folder', that is inside 'clip'" or "either on the 'clip' folder or the 'text encoders' folder"?
I'm pretty much new on Confy, hopefully you guys can help me here.
You place text encoders into text_encoders folder. However, just for your information, if you make a subfolder in this folder, or in any models folder, you can place corresponding models in that subfolder. ComfyUI will see them in that subfolder. There is no need to create 'clip' subfolder in 'text_encoders' folder, but it may come in handy if you want to sort your models. For example, keep checkpoints for, say, SDXL and Pony separate.
@creatorjulie743 Yeah I kind of expected that to be the case. It felt weird to have a text_encoders folder only to create another one in clips. Still, even if I put the file there I still get the error...
@archtosisbimbosis108 Well, cannot help right now. I am getting an error of my own. But tell me, exactly what file you are talking about?
@creatorjulie743 I'm getting the "Expecting value: line 1 column 1 (char 0)" error on the DualCLIPLoader node. When I start running the workflow It goes trough like 3 nodes or so for half a second, then the error appears. The node glows in a magenta color (instead of red).
I don't think the problem is in any file that I was required to download, I guess is something that I failed to install or update, but not sure what.
@archtosisbimbosis108 are you two using comfy manager to install and update? That thing can be a mess sometimes if you are.
So,if you downloaded the models I linked and getting that error. Especially on a fresh install, I would say that's some sort of node/dependency problem but it's hard to say without an error log.
If you want put an error log on pastebin and I can take a look in a little bit. But more than likely this is a node problem
@Urabewe I'm installing the nodes with CMD git pulls and then the requirements, also with CMD. Then I use Manager and hit "Update All". I can do another clean install and try again without Manager. I heard that updating with Manager can mess up things, but I thought that only happend when you have a lot of different nodes.
I'll reinstall Comfy and try again, hopefuly It will work this time.
Thanks for the help Urawebe.
@archtosisbimbosis108 Since I came out with the LTX-2 workflows originally I have learned that a lot of people have installs with custom nodes that won't even be in use but causing conflicts, nodes that never update until deleted, mysterious crashes with no error codes...
The workflow, though not fully optimized, works. I know that for a fact. It's just a matter of finding out what is conflicting on your server/system.
@Urabewe I tried again without Manager, still no luck. This is what appears in my CMD when I run the I2V workflow:
https://pastebin.com/eP0i8hKb
This is what I did to install everything:
1- Download and extract the portable version of Comfy from Github
2- Run Comfy trough "run_nvidia_GPU", check that it works, then close it
3- Run "update_comfyui"
4- Go to both links for the nodes that you put in the workflow (thanks for that btw)
5- Clone the nodes in "custom_nodes", then install the requirements as the GitHub for the nodes say
6- Place the files in their proper folders in "models"
7- Run comfy, place an image, hit Run, then I get the error
If it's important: I run this on a Windows 10 LTSC ('cus I hate 11), I have a 4090 (24 GB) Nvidia card and 32 RAM.
@archtosisbimbosis108 I'm not sure at that point. Where is the clip stuff located? clip or text encoders? whatever folder you have it in.... put it in the other! see if that works, seriously...
@Urabewe are you two using comfy manager to install and update?
-----------------------------------
Hell, no. Manager is good to check out what nodes are missing or installed, and even then there can be some nuances. I am using git clone and then install only those packages from the requirements.txt that are not installed yet.
The problem in my case was that one of my packages was not meeting the requirements of LTX 2.3. Updated it, and everything started working just fine.
By the way, I am running the default LTX-2.3 workflow that comes with ComfyUI. I got it working with all the standard safetensor models, including Gemma-3-12B-it, but I don't know what to make out of it yet. Need to experiment a bit more.
@archtosisbimbosis108 When you update with Manager, it pulls new versions of all packages, some of which may conflict with already installed packages. I personally check requirements.txt file and install via pip only those packages that are not installed yet, and only those versions that are indicated.
In fact, at my latest update of ComfyUI I got lazy and just installed the whole requirments.txt file, and it pulled one package that conflicted with the rest. So had to downgrade it a version to get everything working again.
Run pip check in your activated environment and see what it says. It should say that there are no conflicts.
@Urabewe It was on text_encoders the first time, I moved it to clip and I still got the same error...
@archtosisbimbosis108 Of course you did. Because the problem is not in where you put the file. If it's a text encoder, then it MUST be in the .../ComfyUI/models/text_encoders folder. Putting it anywhere else outside this path will get you a missing model error, while putting it into a subfolder within this path makes no sense, unless you want to keep your models separate. Asking you again, exactly what file are you talking about? If you want some help, you need to provide some information about the problem, not just stating that you have it.
@Urabewe Well. Frankly, I am confused myself now. I finally got to loading this workflow and see some strange model folder structure. According to the instructions in the workflow, model gemma_3_12B_it_fp4_mixed.safetensors (text encoder) must to into some 'clip' folder. But there is no 'clip' folder in the default ComfyUI model folder hierarchy. Usually this model is placed into the 'text_encoders' folder. It IS in this folder in my setup and node DualCLIPLoader sees it with no problems. The same with the ltx-2.3_text_projection_bf16.safetensors model. The instructions say to put it into clip/text_encoders folder, but I have it in the standard text_encoders folder and the DualCLIPLoader node also sees it with no problems. I am currently downloading the rest of 2.3 models. I will see how it works when I have them and will then tell you where to put these models. But, as far as can see now, both of these models should be places into the ...ComfyUI/models/text_encoders folders. I don't know WTF that 'clip' folder is.
@creatorjulie743 Yeah that was a bit confusing how I put it. It's clip OR text encoders depending on which one you have, some will have one or the other some both. Depending on how old your install is, portable vs desktop and many other factors.
@archtosisbimbosis108 Here is ChatGPT says about your error message:
The problem is not the VAE. It is one of the two CLIP/text-encoder files being loaded by DualCLIPLoader. In your workflow, that node is set to load:
gemma_3_12B_it_fp4_mixed.safetensors
ltx-2.3_text_projection_bf16.safetensors
LTX2.3 I2V GGUF 12GB-set
The traceback shows the crash happens here:
nodes.py -> load_clip
comfy.sd.load_clip(...)
comfy.utils.load_torch_file(...)
load_safetensors(...)
then json.loads(...) fails on the safetensors header
That means ComfyUI opened one of those .safetensors files and found data that is not a valid safetensors file header. In practice, that almost always means one of these:
the file is corrupted or incomplete
the file is actually an HTML page / text file renamed as .safetensors
the file was downloaded from a Hugging Face blob page instead of the raw file
the file size is 0 bytes or otherwise wrong
Your enclosed workflow itself contains Hugging Face links in its notes that use .../blob/main/... for at least the text encoder files, which is a red flag if those were downloaded incorrectly via browser instead of as the actual raw model files.
So, check if you indeed have the right text encoder files. gemma_3_12B_it_fp4_mixed.safetensors must be around 10 GB, and ltx-2.3_text_projection_bf16.safetensors about 2.5 GB. And check that you have exactly these model file names.
@Urabewe Well, I myself am using a manual install, and even though there is 'clip' folder in my model folder hierarchy, this is not where I put my text encoders. Right now, there is indeed model gemma_3_12B_it_fp4_mixed.safetensors in 'clip' folder in my setup, but I am using gemma_3_12B_it.safetensors from 'text_encoders' folder. Although the DualCLIPLoader node does see the models in 'clip' folder too. Anyway, will see how it all works after I get all the 2.3 models.
Alright. Got it running. Both text encoders ( gemma_3_12B_it_fp4_mixed.safetensors and ltx-2.3_text_projection_bf16.safetensors) go into the '...ComfyUI/models/text_encoders' folder.
A couple of small problems. Audio VAE LTX23_audio_vae_bf16.safetensorsthrows an exception. Using audio VAE for LTX-2. Seems to work. Also, link is for wrong LoRA, for LTX-2. The right link for LTX-2.3 LoRA is this: https://huggingface.co/Lightricks/LTX-2.3/resolve/main/ltx-2.3-22b-distilled-lora-384.safetensors. But works even with the wrong LoRA, just bitches about it and, apparently, not using it.
@creatorjulie743 @Urabewe I finally made it work, It seems that something went wrong when I was downloading the files that go into "text_encoders", and they only weighted like 84 KB. I downloaded them again and now it works.
Thanks to both of you!, really excited to try this.
@creatorjulie743 there has been an update with the workflows you can get the new distill Lora from the workflow description at the top. Ltx2.3 distill at 2.5gb instead of 7gb. Lower rank still works just as well from kijai.
Old Lora was throwing key errors but was being used otherwise my outputs would have been broken due to low step counts and no cfg.
@archtosisbimbosis108 glad you got it working, lots of people having issues with clip and vae and sadly I can't just go to their computers and be like "nah it's this" I gotta guess and hope they are doing things correctly lol so, thanks for sticking with it and I hope you have fun!
@Urabewe I can see it. Like this guy. The issue was so simple. How in the world can not see that you downloaded 55 KB instead of 10 GB? Beyond me.
@creatorjulie743 shit happens I just have tons of people and tons of error logs and not enough time for all of them to decipher and guide each person through the troubleshooting process and it's all fixing their comfy installs or in this case redownloading models
I fixed the audio vae issue with kj nodes because of the comments but now im getting this massive repeating error (Error(s) in loading state_dict for VideoVAE: size mismatch for decoder.conv_in.conv.weight: copying a param with shape torch.Size([1024, 128, 3, 3, 3]) from checkpoint, the shape in current model is torch.Size([256, 128, 3, 3, 3]). size mismatch for decoder.conv_in.conv.bias: copying a param with shape torch.Size([1024]) from checkpoint, the shape in current model is torch.Size([256]). size mismatch for decoder.up_blocks.0.res_blocks.0.conv1.conv.weight: copying a param with shape torch.Size([1024, 1024, 3, 3, 3]) from checkpoint, the shape in current model is torch.Size([256, 256, 3, 3, 3]). size mismatch for decoder.up_blocks.0.res_blocks.0.conv1.conv.bias: copying a param with shape torch.Size([1024]) from checkpoint, the shape in current model is torch.Size([256]). size mismatch for decoder.up_blocks.0.res_blocks.0.conv2.conv.weight: copying a param with shape torch.Size([1024, 1024, 3, 3, 3]) from checkpoint, the shape in current model is torch.Size([256, 256, 3, 3, 3]). size mismatch for decoder.up_blocks.0.res_blocks.0.conv2.conv.bias: copying a param with shape torch.Size([1024]) from checkpoint, the shape in current model is torch.Size([256]). size mismatch for decoder.up_blocks.0.res_blocks.1.conv1.conv.weight: copying a param with shape torch.Size([1024, 1024, 3, 3, 3]) from checkpoint, the shape in current model is torch.Size([256, 256, 3, 3, 3]). size mismatch for decoder.up_blocks.0.res_blocks.1.conv1.conv.bias: copying a param with shape torch.Size([1024]) from checkpoint, the shape in current model is torch.Size([256]). size mismatch for decoder.up_blocks.0.res_blocks.1.conv2.conv.weight: copying a param with shape torch.Size([1024, 1024, 3, 3, 3]) from checkpoint, the shape in current model is torch.Size([256, 256, 3, 3, 3]).)
Your LTX workflows are the best. I'm getting much better results than ComfyUI's default template, even better than using the full 46gb dev model. I don't know how you fine-tune it, but it works!
Little changes especially in i2v with the image2vid node. Lots of examples on reddit where the face morphs and gets all wrinkles is because of that being too high.
This was just an update to get 2.3 rolling. A major update will be coming soon. May require the ltx node pack on that not sure if I'll use it or not since it seems to double gen times.
Clean up the workflows for better organization, tweak sigmas and samplers and maybe bring in a regular k sampler instead of this advanced shit most users can't even decipher, tweak the way we do things with the dev and distill Lora. Then there is fflf workflows to make and many other things.
Might be a week or two but the next workflows will be a true actual version change I'd say.
When doing 20 seconds,I got result in the end where it have the box colors https://civitai.com/images/123429409
I'm getting it too. Even if I change it to 18 or 15 seconds, I still get it. Sometimes it looks like a bright watermark, almost looks like a logo or title. The only way I don't see it is if I change the video length to 10 seconds and below.
@denjay5157 I was trying this https://civitai.com/models/2295882/ltx-23-basic-gguf-720p-workflow, it fixed for me
@Cyberfolk Thanks! If you're keen with First to Last Image to video using LTX 2.3, you can try these workflows https://huggingface.co/RuneXX/LTX-2.3-Workflows/tree/main
Been noticing that too. I updated my workflows in a hurry to get them out for 2.3
I will be going over them, cleaning them up, comparing to official and adjusting as needed.
I'm also looking at the desktop app and their other pipeline code to see how they are doing things differently in desktop app and hopefully bring that into comfy.
Right now desktop app is much better quality than the comfy version. So not sure if their API is just different weights or if they are doing something hidden under the hood.
@Urabewe thanks ! appreciate it
x = torch.nn.functional.linear(input, weight, bias)
RuntimeError: mat1 and mat2 shapes cannot be multiplied (1024x4096 and 32x4096)
Prompt executed in 12.76 seconds
i dont know what the error is.
That's usually text encoder. Make sure everything is updated. Comfy, kjnodes, and gguf nodes.
Updating with manager doesn't always work. You might have to git pull manually in the directory or completely uninstall and clone repo again
I also have this problem. I have already created a post on Reddit. I will provide a link to the post so that you will be informed if a solution appears.
https://www.reddit.com/r/comfyui/comments/1rnhd96/is_it_only_me_who_has_this_problem_ltx_23_gguf/?utm_source=share&utm_medium=web3x&utm_name=web3xcss&utm_term=1&utm_content=share_button
"RuntimeError: The size of tensor a (3520) must match the size of tensor b (466048) at non-singleton dimension 2" Anyone getting this error deleting the ComfyUI_smZNodes in the CustomNodes folder and restarting Comfyui did the trick.
the unet loader is mismatched and not loading the gguf. any ideas or suggestions?
update gguf nodes more than likely. manager doesn't always work, git pull in the directory itself or uninstall and re-clone the repo manually.
git pull is easiest and should work. go to the custom node folder, open cmd pointing at that folder and type "git pull" press enter.
@Urabewe that was the case for me, thanks
Hi friend. Great job. Question: Why does an image appear superimposed on the final video, matching my previous image? It creates a hazy effect. Could you help me? Thanks a lot.
This is excellent work. Thank you.
I didn't see an abliterated gemma being used but it stills seems to be NSFW or is it? Thanks.
Ltx is uncensored I'm pretty sure but doesn't really do it well. Don't do NSFW stuff so I don't really know. If it does it would need a Lora.
Abliterated gemmas are out there but I never see a big enough difference to warrant it when using as a text encoder
@Urabewe Thanks again!
I'm trying to use the I2V workflow but get a load of errors like this:
VAELoaderKJ
Error(s) in loading state_dict for VideoVAE: size mismatch for decoder.conv_in.conv.weight: copying a param with shape torch.Size([1024, 128, 3, 3, 3]) from checkpoint, the shape in current model is torch.Size([256, 128, 3, 3, 3]). size mismatch for decoder.conv_in.conv.bias: copying a param with shape torch.Size([1024]) from checkpoint, the shape in current model is torch.Size([256]).
I have all the correct models indicated by the workflow notes and have tried updating the KJNodes package. Anyone had this?
Some are having problems with the vae files. Try redownloading. If you are using comfy manager to update the node pack it may not always update. Git pull inside of the kjnodes folder or delete and reclone.
Some have conflicting node packs that screw with the vae node. Make sure the vae node says it is from kjnodes and not some other node pack
Also getting weird vae errors.
On V2V, does this extend the length of an existing video? how do you control the extend frames?
V2v does extend the video. Feed it a few seconds and it will clone video and audio.
Just set the time I suggest setting for 1 second longer than you want due to some files being a bit longer than what they say. A 5 second video isn't always 5 seconds sometimes it's 5.4 seconds.
Frames are calculated automatically from the source video and length in seconds. They are calculated in a way to keep from getting a few seconds of static at the end where the mask goes beyond frames.
Everything should be automatic just supply a video, length, and caption.
I'm on RTX 5060ti / 16 Gb and still running out of NVRAM with the original workflow settings. Adjusting the picture and video size to smaller values, changing frames to 63, changing upscale from 0.8 to 0.5 does not help. What am I missing?
How much system ram do you have?
@Urabewe 64, anyway, the error says I'm out of NVRAM. When generating, the error appears at stage 2, and the busy RAM does not climb over 27-30 Gb.
Also, I'm on Linux.
I'm no expert, but is it the resolution? Try generating at a lower resolution and up-scaling afterwards
@KurtStaubitzer Thank you for your kindness! I do have changed to 480 px and 43 frames, also I have reduced the movements complexity, yet no positive result. Moving on.
That's weird. I can run the original (ComfyUI default) i2v workflow with no problems on my 3060 with 12GB VRAM. I do have 128GB RAM though and it gets used heavily, sometimes up to 75-80%.
I managed to run the Q3_K_M GGUF version on my AMD rig, wich uses a 7900XT and only had 32GB of RAM.
Ksampling is fairly fast, there's probably something wierd with VAE decoding, takes much more time than it should. the video took around an hour overall to complete, far worse than it should be, but still impressive it works at all.
if you're running on AMD (using ROCM7) and have VAE slowdowns, you need to disable cuDNN. you can use https://github.com/sfinktah/comfy-ovum to turn it off and on automatically which should speed things up.
@slln the link is broken, but i gess you are referring to the "ovum-cudnn-wrapper" custom node.
I needed to manually exclude LTXVAudioVAEEncode and LTXVAudioVAEDecode, but after that it worked indeed much better.
Thanks, i didn't know this one.
any ideas why output is blurry like this? https://civitai.com/posts/27197809
I use latest latent upscaler in proper folder
you are probably using incompatible lora, try disabling or reducing strength
Per Claude's analysis: ● Now I can see the exact bug. The Decoder.__init__ pre-computes output_channel (the starting/largest channel count) by
iterating blocks in reverse — but it only accounts for res_x_y and compress_all multipliers. The LTX 2.3 decoder uses
compress_space and compress_time blocks which also change channel counts, but those are not counted in the
pre-computation.
The math:
- LTX 2.3 config has compress_space {multiplier:2} and compress_time {multiplier:2} in the decoder
- These DepthToSpaceUpsample blocks change channels by their stride factors (4x for space, 2x for time)
- Current code ignores them → computes output_channel = 256
- Checkpoint was built with those accounted for → actual starting channel = 1024
This is a bug in ComfyUI's causal_video_autoencoder.py that hasn't been patched yet.
what bug exactly?
tl;dr - In the V2V workflow, is there a way to have only the new segment of video go through upscale and output? I just want the new video output with the source video cut out, and I'll join them later.
Basically what I'm trying to do is extend a video again and again indefinitely, while maintaining faces and voices. It OOMs on upscale due to running the source video that I already have as well as the new segment. I could just cut both durations in half, but even then, some portion of every video segment is redundant, and that's a new mess, as well as generation time being increased on video I don't need. I've tried some latent trimming nodes in different places, but as you can probably tell, I don't really know what I'm doing, and it barfs when I do that. Do you know of a way I can cut out the source video just before the upscaling stage, but keep all the context from it?
there would be ways yes but i haven't put it in nor would i plan to. This would be done by putting nodes in before the upscale that would separate the two then upscale just the new segment.
Possible yes, but I haven't ever looked into it so you'd have to figure that one out
So I've been using the Comfy template of LTX 2.3 successfully, and just plugged in your workflow (thank you btw). I think I have everything in its right place, no missing nodes, no errors when I run a generation, but when I click "Run", it cycles through the workflow in like 2 seconds and says "Job Completed", but there's no output. Any idea what this could mean?
*edit: I figured it out. If anyone else runs into this just make sure you READ THE INSTRUCTIONS IN THE WORKFLOW AND PUT YOUR TEXT ENCODERS IN THE FOLDER HE TELLS YOU TO. lol
Solid advice!! Hey, at least you got it working! Have fun
@Urabewe Thank you so much! It's an incredible workflow. Especially coming from WAN 2.2
Anyone having an error with VAE, UPDATE ComfyUI-KJNodes
These workflows are great. Having a good time figuring out how to make a music videos.
One request: Being able to start a clip at 0:30.5 instead of 0:30 or 0:31
The custom node ComfyUI-Logic won't import :/
I'm using the Easy-Install fork, and everything is up to date
I can't seem to make the lipsync work. I even changed the audio to stereo, prompted the exact words the character should say, changed the video length to match the audio length, etc. Any help, please?
Are you using the crop audio node? That needs to be connected to the original audio node. I will have that fixed up with a few more fixes very soon.
I'm just about finished with some tweaks and it seems everything has settled down a bit. New upsampler, previews, better connections with crop guides and other intermediate nodes
@Urabewe Hey, yes, everything was connected. Managed to solve it. Elevenlabs audio comes too clean, for whatever reason, adding some background noise makes lip sync work.
@zexeor Did you change anything else than the background noise to get it working?
@nikoe No, that was enough to make it work
OMG this workflow is great. Thanks. 1st time using LTX and it looks great (image to video). I am coming from Wan2.2! So much FASTER!!! I have an RTX5060ti and 32gb of memRAM. Just had to "Update All" in the manager and it worked great
is this supposed to merge the original video with the new extended video? im using the V2V and its making a great extended video but its output is just the new video.
Been having a blast with this one. really well made. question, is there a way to avoid the garbled text/logos/graphics that almost always appear at the end of my outputs? I've tried all the negative prompting I can think of but it doesn't seem to help. Thanks!
It doesn't work. First there were problems with the BAE, then with the sampler, and then finally just a black square. Nothing makes sense. There's no sound either.
So there seems to be a lot of confusion about why something doesn't work in ComfyUI. Your problem is not the workflow it is your install of Comfy or it is the models you downloaded.
If you post a few error codes or at least a description of what's going on I might be able to help.
Did you install all the links that were provided in the workflow?
For audio update the custom nodes
I'm getting good results with 11 second videos. Img to video. Is doing "Last Frame Loop" sime on this? My chatbot mentioned doing this for longer videos and changing scenes? I guess I'll look on youtube what that's all about.
Hey gays, according to a thread on runninghub, I changed the first sigmas to 1.0, 0.99375, 0.9875, 0.98125, 0.975, 0.909375, 0.881203, 0.863321, 0.841251, 0.820089, 0.655, 0.381875, 0.0 . It's worth a try.
NO SLIDER Clip duration (in seconds)
can this be used on wan2gp?
Can it work without swap? Dont want to kill SSD.