Edit: going to try the v4.4 workflow and see if that helps!

maybe add what you want to hear in the prompt "quiet room ambience, soft footsteps, distant traffic, subtle environmental sounds, natural room tone, unscored scene, documentary style audio, no musical score". Or describe the kind of music you actually want to hear... Otherwise it will try to fill the blank with something (music most of the time... as in movies)

Not sure if you figured it out, and If you mean for the node in "Input" blue group, top left, it's https://github.com/tritant/ComfyUI_Custom_Switch

画像から映像/参照への機能で問題を抱えている人もいるのに気づきました。詳細な構成については私のウォークスルー動画で説明しています。

参考画像の説明は3分20秒から始まります。もし行き詰まったら、ぜひご覧ください!

https://youtu.be/kEBgAkVHlBE?si=CoqPNj46TAw1eYru

I've done this with with a modified version of this workflow. See this post for video examples. The workflow is embedded in the images.
https://civitai.red/posts/28200204

@RandomAIUser its work! blank audio - https://github.com/anars/blank-audio/blob/master/1-second-of-silence.mp3

You could be using a wrong model version. Try downloading the exact models by name that are noted in the workflow.

Download this file,
ltx-2.3-22b-distilled-lora-dynamic_fro09_avg_rank_105_bf16.safetensors
And put it in ComfyUI/models/Lora

I had the same issue as you, and that file was missing, but I didn't know.

Also, go to the Workflow on your computer,
and in the far left it says "Model Links".
Go look at each one, and make sure that it is downloaded and in the file location that it says at the bottom of that Model Links section.
I was missing the Destilled Lora, even though ComfyUI would still run... it would just make hyper blurry videos.

Also, tip, make sure that you download the files in the links at the top, because the file names at the bottom have some typos (some of them have the old names).

Let us know what happens!

I apologize if this is stupid answer, but perhaps you're running low steps without distilled LoRA.

@Arcadeath I don't know!
I will check.
Thank you for giving me some insight, though.

@maninstrawhat You're welcome. You could always check if your model requires a distilled LoRA. I use Eros10 model and I use lltx-2.3-22b-distilled-lora-dynamic_fro09_avg_rank_105_bf16.safetensors for first pass, and ltx-2.3-22b-distilled-lora-384-1.1.safetensors for final pass.

@Arcadeath That's cool!
So... this workflow uses the ltx-2-3-22b-dev-Q4_K_M.gguf model, but you use the Eros10 model?
By the way, Arcadeath, do you know how I can check if I am using the right Lora?
Because I see the "Power Lora Loader (rgthree)" node, and it says that the toggled model is "90 video\ltx-2.3-22b-distilled-lora-dynamic_fro09_avg_rank_105_bf16.safetensors" with Strength of 0.60 , however I don't see that file in my ComfyUI folder?

So, I'm wondering if it's really loading, and where it might be at.

Also, in order to add another pass for the "final pass", could I just click the "+Add Lora" node button? And then add model, like what you are using, for instance?

Thank you for your help!
I really appreciate you taking the time to respond to me!
It makes a big difference,

Edit: I just downloaded the LoRA file in question, and put it in the Models\loras folder... so, here is to hoping that works. *shrug*

@maninstrawhat You're welcome! I think that should fix your issue, but I am not entirely sure if this workflow has first / final pass separated as the one I use (I actually want to try this workflow but didn't yet). Also, if you have 12GB VRAM + 32GB RAM you can use models better than gguf q4 (IF heh).

@maninstrawhat I just tested this workflow with and without LoRA! Without LoRA it's blurry/smeary, with LoRA it's perfect for me. Still, I don't really know what kind of LoRAs quantized models need (GGUF ones). What's your VRAM and RAM amount?

@Arcadeath I am running what it suggested: RTX 3060 12GB, and 32GB RAM.
Actually... Arcadeath, I got it working last night! :D

It seems to be what you suggested.
I think that I had the right lora selected in ComfyUI, but when I went to double check that it was actually installed, I didn't find it at all in the ComfyUI directory.
So, I downloaded it by doing a G**gle search, and put it in the models/loras folder , restarted ComfyUI... and it worked!

I was wondering... how would it even work at all without the right LoRA ?... this is all new to me.

I really appreciate your help!

Because I followed your suggestion, and it worked out!
I greatly appreciate it! :)

I'm glad that it works for you, too.
I guess next I'm going to try out adding another LoRA.

And a better GGUF model, too!

Thank you for the tips about using a better model. I want to try that now.

@maninstrawhat Hey, I'm happy it helped! So, what really happened and why it's working now: Models require lots of steps to generate a coherent output. If you increased amount of steps to 30-50 instead of 8, it would generate nice results probably too. But that would take 5 times longer sadly. So people make distilled LoRA which can reduce amount of steps (like turbo image models). It reduces quality by tiny bit, depending on models, but way way faster. Also GGUF models are way slower (quantized models), and they can't be trained upon further. My recommendation is to always get diffusion models since you can offload their size to RAM (hence the increasing RAM cost). Best way to know is if you check VRAM usage per system (I use Linux with 300MB usage), then add model GB, add LoRA, text encoder, clip, and all other files as size. If it is higher than your total VRAM + RAM then you can't technically run the model. This workflow is well optimized since it has great offload switching, model unloading when they are not used, etc, so you don't out of memory (OOM). fp8 models are bit faster on 40xx series because of native support, I got 4070ti, but they can be still ran on 3060.. Even then I think they will be faster for you than GGUF models. Also Q(number) GGUF models are ranked by size. Just because a model is 20-30 GB doesn't mean that you can't run it because you can offload parts of it to RAM. Some parts are not unloaded, some settings can be tweaked, but with 12GB VRAM and 32 GB RAM I can generate 2k videos of 15 duration with this workflow and Eros10 fp8 diffusion model. GGUF files = slow but lower size, I could run Q8 model too, but it's still worse than fp8 one in both terms of speed and quality at that point. Diffusion models need separate clip and vae files, and there are also checkpoint files which have vae and clip baked in, so you need one file instead of 3 (merged models). LoRAs are just 'addons' to existing models, like a new concept, and stacking them also increases memory usage since they are loaded with your model. So, the question is can you fit that into 48GB with system usage, apps open etc. I also have 64GB .swap file allocated on my SSD so it can use it as very slow reserve RAM.

@maninstrawhatAlso, to add different LoRAs to first and final passes just copy paste lora loader with those two tiny nodes connected to and connect model to both of it. Setter of lora loader 2 will automatically update. You can double click that main "Processing" node and on bottom left in new window you can make another setter. Processing Is actually like "folder" node to organize background stuff. Inside you can see how models are loaded for first and final pass. In this workflow they are called earlier, then they got fed into  both first and final pass models at once. There you can feed new getter separately into final pass (Follow the lines and where they go and try to understand them, search for first get_m model. It's best ComfyUI learning experience. In this specific one when you're in processing, zoom out a lot, and long curvy line in top middle is model load for final pass. It goes down right into it (when that node is green during 3/3 processing in terminal). Click on nodes and you can open their documentation on the right. Some can be double clicked to expand since they work like back end organizers only).

I can send it after weekend once I got more time. You can see at very beginning how model's are "SET", you need to set another one to different name, and "GET" it separately into first and final pass, you need to reconnect the line with new getter into final pass node.

Follow terminal and what node is green is best way of understanding what happens when.

Tried the v.43 workflow as well. getting the same results

how without ANY info whatsoever..

Do you have the Destilled LoRA in your ComfyUI/models/loras folder?
Because I found out that I did not, and after downloading it and manually inserting it into the folder, the workflow worked.

Before I did this it was generating very blurry videos.

This is the name of the file I was missing:
ltx-2.3-22b-distilled-lora-dynamic_fro09_avg_rank_105_bf16.safetensors

I would check on the far left side of this workflow, in the "Model Links" section, and make sure that you have every file it lists present inside of your ComfyUI install folder, and in the listed location.
It says at the bottom of that section where the file needs to go.
Be careful to use the model names at the top, though, because at the bottom it has some typos.

NEVERMIND I just found how to do it, great workflow, amazing and simple to edit!

In each subgraph there are two nodes called ManualSigmas with a bunch of numbers in them. Each of those numbers is a step. The easiest way to change it so you have control of the number of steps it to replace the Sigmas node with a 'Basic Scheduler' node. Use 'linear_quadratic' as the scheduler and adjust the number of steps to whatever you want.

@RandomAIUser Thanks for the tips, it is improved after testing combinations. I added the LTXVScheduler node and use 16 steps on first sampler, and this sampler alone the quality of video surpass the final from the original workflow. Then on second sampler I used the Basic Scheduler(linear_quadratic) combined with the Sigma Rescale node(set to start at .75) for the 3 step.

@R3G4L Hey there. I see something interesting is happening here. Can you share your workflow?

@R3G4L I have had really good luck with setting the first stage to 2-3 steps and 8 steps for the second. I know that is totally backwards to what they are supposed to be. But the quality (and prompt adherence) after the second stage seems better for some reason. It almost looks like it starts over with the second pass. Making me wonder what a single stage workflow would look like. But I haven't tried it.

couldn't wait, so I set it up myself =). In the end, for the first pass, I set CFG to 3, 30 steps, denoising in the Base Scheduler node to 1, plus various LoRAs. For the second pass, I set the Distillation LoRA to strength 0.6, denoising in the Base Scheduler node to 0.45, and 9 steps. The face stays consistent without any hocus pocus in most cases. And the quality seems to have improved. I'm still testing it, but overall it's OK.

sometimes, but adding this helps:

her mouth opens and closes naturally with each syllable as she says "blah blah blah"

generally only need to say it the first time, but if there's a lot of movement between dialogs or multi characters speaking, sometimes you need to add it at each problem area.

did u find a solution to this?

I got rid of that missing custom node. One of these commands did it. Not sure which.
Python Package > + > "pip install" opencv-contrib-python

Python Packages > kornia package > downgrade it to the version below 8.0

install triton-windows

Package Commands > Install Bitsandbytes (ROCm)

I'm failing to install "comfyui_layerstyle" nodepack though. Got any tips how to install it?
Probably related to the opencv-... (it needs opencv-contrib-python) but comfy seems to auto-install other opencvs that conflict with it. Couldn't solve that one yet. Any ideas?

The issue is that ComfyUI uses its own virtual environment (.venv) with a different Python version, but soundfile was installed into your system Python. ComfyUI cannot see packages installed outside its virtual environment.

To fix this, open a command prompt and run the following command. Replace the path with the actual path to your ComfyUI installation:

"C:\Your\Path\To\ComfyUI\.venv\Scripts\python.exe" -m pip install soundfile

For example, if your ComfyUI is installed in D:\ComfyUI, the command would be:

"D:\ComfyUI\.venv\Scripts\python.exe" -m pip install soundfile

After running this command, restart ComfyUI Desktop completely. The missing nodes (JWInteger and JWIntegerToFloat) should then load correctly.

Same here

i manually replaced all comfy-various node to alternative, especially those float to int. conversion, comfy-core have similar function. Also can use those hot/common node like rgthree kjnode....
donn't why author use so many different custom node while some of them are not necessary...