Відео

8s at 24p is all I can do on a 4090.

anyone else get this error when trying to install the insightface asset? any ideas how to fix this? \insightface-0.7.3-cp310-cp310-win_amd64.whl", line 104 <title>Assets/Insightface/insightface-0.7.3-cp310-cp310-win_amd64.whl at main · Gourieff/Assets · GitHub</title> ^ SyntaxError: invalid character '·' (U+00B7)

Hi mate, I cannot find a file named bigvgan_v2_44khz_128band_512x Am I missing something obvious? Thank you :)

I hadn't heard of the video styler node before so thanks for that. The previous workflow I had fed both + and - prompts into the same input, which seems weird, and suggests it doesn't take any notice of negative prompts, certainly these don't always seem to have an effect.

Awesome I finally got it working from your video! I noticed if my prompt is too long I get cache errors. Keeping it shorter like 150 words or less no problems. I have an RTX 3080ti 12GB vram

The compilation node works badly on windows because of issues overwriting files in the temp directory. So moved my ComfyUI into the windows subsystem for linux. I installed the latest pre release version of pytorch and compiled xformers wirh cuda 12.6 support. These versions are as fast without any compilation as they were on windows with compilation. They also do not gain any additional speed through compilation. Maybe there is some kind of jit compulation going on in these new versions of pytorch and xformers? Then I tested the new setup and the two caching methods TeaCache and FirstBlockCache with the new setup. I used prompt that generate pictures with a high amount of text in them because thats where you can see reduced quality first, when it starts to get letters or spelling wrong. My result is for both caches: As soon as the cache reduces render times quality is reduced also. So there is no free cake here. The only speed improvement you get for free is by moving to Linux and using the latest dependencies for pytorch and xformers. When you run comfy with or without xformers using the latest software the speed gap between xformers or pytorch attention shrinks to almost nothing. I still have to check if there is a difference in quality though. Now that we know that virtualization doesn't hurt performance the next logical step is running comfy in a docker environment.

I appreciate the depth of your descriptions, leaving no questions. I always wondered if I was crazy or had different versions of python rewriting itself on various mega-multi-layer deep folders.

I'm experiencing an issue when generating videos in ComfyUI. I'm using the latest version of ComfyUI and have tried using the models shown in the tutorial video. While I can generate videos in my other workflows, this specific one is producing blocky and pixilated results (Looks corrupt). To troubleshoot, I've tried disabling the Power Lora node and a few others, but the issue persists. Does anyone know how to resolve this issue?

My show prompt box is not displaying any text. any idea?

with this workflow, not sure what I'm missing but the video (before upscale..etc) generated like it wasn't completed... only some weird artifacts rendered. No matter what version of model I'm switching from the loader section.... my simple hunyuan workflow works so Im not sure what I missed configured on this one. Any tips?

One question. Apologies if I missed the explanation in the video, but why must the prompts submitted by so short to get effective results?

Thank you, so much, for clarifying the Hunyuan AI text2video must utilize a simple and shorter prompt. I was trying different solutions that were more complex and verbose (e.g., "ChatBot write me a prompt for...").

hello, do you think you will ever make a video showing us how to install and make a lora for hunyuan using kohya ss musubi on windows?

Yea never worked for me. Assuming its the triton thing. It works when I apply first block cache but failed with compile model+. Got a 3080ti

Wavespeed doesn't work with 1080 ti cards, but teacache does.

IDK whats wrong here. I installed it and first of all, it works, but onyl with schnell models not with dev models. i mean yeah its faster in generation but the pictres with dev are complete broken. the images looks like i would use only 3 or 4 steps on dev. i tried it with 50+ steps but everytime the same result with dev models

I can definitely confirm the Apply First Block node works without Triton, installed Wavespeed last week and got an error with the included workflow with their nodes. I removed Compile Model+ and tested it the the Apply First Block node enabled and bypassed and found generation speeds were almost twice as fast with its enabled.

I just installed the portable version comfyui 3.12 python and titron.. no issues. WaveSpeed is really great boost.

I used the 3.11 wheel and lib/includes and I just get an OOM from the block cache. It also broke a number of custom nodes from starting so I had to uninstall Triton. I hear 3.10 is the way to go for reliability. I am looking for the portable version of it.

It worked for after a whole month trying to run triton. Thanks!

This won't work if your portable comfy is using 3.12 python. Triton bricks it since it uses a depreciated ImpImporter. I tried using the cp312 whl file and it still fails- preventing comfy from opening. To fix you need to delete the triton folders under lib/sitepackages

i stuck at ksampler it showing 0/8 and the it crashed i use image motion guider shape '[1, 10, 17, 30, 16, 1, 2, 2]' is invalid for input of size 337280 i new to this and i need help thank you

I'm a little surprised. I used the "flux-dev-f16.gguf" because it had the best quality. And I notice maybe 10-20% speed up. But if I use the flux1-dev.sft model and the "load diffusion model" node as usual, I have significantly more speed. and instead of 97% VRAM only 87%. my workflow: 217s and the one with gguf: 398s all on rtx 3090ti. But thanks , it may helps with Hunyuan.

I collected a dozen links on installing Triton on Windows that I was about to slog through, including using Visual Studio to compile the whl file, installing WSL, and a bunch of horror stories of it ruining the entire environment. This is much more straightforward. Thank you for taking the time to explain it so clearly!

It is the tutorial I was waiting for. It took almost one day to me to figure out how to install this triton. So Big Hug and Big tanks to you guys and thanks for the tutorial

I am slightly confused: i would expect to install these commands inside the virtual environment of comfy, so using venv

I am sick and tired of this triton thing, I have a 4060, run 2 comfyUI, a manual one I don't fidget with ever because I don't wanna break it, but I do have a portable version which runs python 3.12.7 and pytorch version: 2.5.1+cu124, I tried and I've bloody trying to install this .whl file from the release page corresponding to my version and it just won't work, what am I doing wrong, all the videos show it like it were such a straight forward installation process. I just ran flux with the 'compile model' node plugged in and I get this error. backend='inductor' raised: ImportError: DLL load failed while importing cuda_utils: The specified module could not be found. Set TORCH_LOGS="+dynamo" and TORCHDYNAMO_VERBOSE=1 for more information You can suppress this exception and fall back to eager by setting: import torch._dynamo torch._dynamo.config.suppress_errors = True I get a different error when using this compile node with hunyuan or LTX, but all I know for a fact is that this god damn triton is an absolute pain to install, especially for the windows users. Just for the record since the newer version didn't work the last time I took your advice and installed this github.com/woct0rdho/triton-windows/releases/download/v3.1.0-windows.post5/triton-3.0.0-cp312-cp312-win_amd64.whl

Great video! Do you have any suggestion for preventing Hunyuan from creating multiple shots in one video? Sometimes it’ll make 2 or 3 very short (and not very cohesive) shots in 1 gen.

6th like first comment Thanks for this extra-ordinary approach and your kindness.

Thank you very much..!

no panix yet again same another comes. flux not get missed. it is just sdxl what was sd

I can't get it working. I hope you can help. When installing the 2.5.1 whl, I right clicked in the python_embeded folder and opened terminal, then typed python.exe -m pip install torch-2.5.1+cu121-cp311-cp311-win_amd64.whl. It installed, but when i opened comfyui, it said Python was still on 2.3.1+cu121. So I tried it again but typed CMD in the nav bar. It then said "This app can't run on you PC". So i reopened by right clicking the Embeded/terminal folder again. The terminal opened up fine and I reinstalled the 2.5.1 whl. It said it was already installed. So I started Comfi and I got that message error that "This app can run on your PC". Comfi won't open.

Do you have any suggestions for how to install on Runpod? I am not sure where to install facexlib and insightface.

Can this work on Mac at all and also can you do image to video instead of text video?

A super good tutorial and above all, thank you for the free and beautiful workflow

So is this creating a prompt from the video or image rather than using the actual input video as a blueprint for the generated one? Is there anyway to do image to vid or vid to vid?

You are amazing! I'm worried about saving up for a rtx 5090 while I've been using a rtx 4070ti. I keep telling myself I need more vram to actually produce anything but you blow that idea out of the window! You're doing everything I dream of doing with a 8 gb vram card. I should be ashamed of myself. Thanks for all of the hard work you're doing bro!

got this error with every mp4. clip. HunyuanVid mp4 could not be loaded with cv.

The workflow looks great. Thank you. Keep on!

Sub'd

excellent content. thank you, sir

Can huayuan video create long videos like 30 sec or 1 minute?

6th like 1st comment Thanks for your kindness

Ouch, that workflow is huge and wants to install loads of stuff and some of it throws warnings and also errors about downgrading. It could break ComfyUI. It would be nicer to split it into minimalistic parts.

Nice explanation to a great design. I always prefer the on/off switches but you added another gem with the model loaders section off to the left along with the switches, too. Plus the strategically placed ones around the modules, a concept others should use!

Great tutorial! my only critique is to have a longer pause between sentences XD

hello Sneaky, I m Stuck at PulidModelLoader Error(s) in loading state_dict for IDEncoder: i am not able unsderstand why am i getting this error

Please A1111

Have you tried it in combination with Redux?

Thanks for taking the time to explain this !!