Bonus TIPS:

I did everything,i even added a new code which i found as a solution on github for comfyui-art-venture,this time it broke my reActor but i fixed it after deleting the code :D

Sageattention doesn't work as a node. Follow the guide and install (preferably use a conda enviroment for comfyui)

https://github.com/thu-ml/SageAttention?tab=readme-ov-file#install-package

Now if you have more than one gpu and one is capable of sage v2 ( i have a 1660 and 3090) then you will get an error.

what i did to fix that (because im on ubuntu and gpu's index weirdly i guess) was to modify setup.py
line 108 "for i in range(device_count):"
to
"for i in range(device_count-1):"

No quotes obviously

@tangentplum598 I'm at the last step and its saying this,

AttributeError: module 'pkgutil' has no attribute 'ImpImporter'. Did you mean: 'zipimporter'?

@Santaonholidays does this help https://stackoverflow.com/questions/77364550/attributeerror-module-pkgutil-has-no-attribute-impimporter-did-you-mean#:~:text=Due%20to%20the%20removal%20of%20the%20long%2Ddeprecated%20pkgutil.ImpImporter%20class%2C%20the%20pip%20command%20may%20not%20work%20for%20Python%203.12.

It would be helpful to know the enviroment you are installing in

@tangentplum598 I was using the comfyui desktop app terminal doesn't that work? :D

@Santaonholidays what

@tangentplum598 Nvm,i don't know how to make a conda .env

Installation of this was really hard. I had to redo the whole python cuda and triton installation and use a version of comfy that has embedded python 3.11, that was my main issue is my comfy used 3.10. Install system wide python 3.11, cuda 12.4, visual studio with the right packages, and triton 3.1. Build triton wheel then build sage and it should work if everything is installed like that right. Need a 30 series+ nvidia video card minimum.

@tenstrip My Brain exploded xD

What is about the other attns in the hunyuan wrapper llike sdpd, comfy, fast att? Will they not work?

@sikasolutionsworldwide709 I tried that and it does work but only if youre lucky,i tried an ice cream to make it drip and it was a static pixelated picture

@Santaonholidays The i2v lora is not perfect almost half of the gens are usually just bad and after messing with settings for a few days it seems like lowering the lora to 0.7 or 0.5 can help but you need the prompt to make the model line the motion up for the input image perfectly or it just peels into some completely random noise. And definitely don't use any Lora's that aren't trained on motion they ruin it more. This is just a scuffed shortcut because the real i2v model release will have much more contextual data and training to base off an input image.

@sikasolutionsworldwide709 They work but sage cut a massive amount off time for me on a 3090, 50 step 3 seconds went down to under 2 minutes at 540p it's definitely the only one that's really optimizing things.

@tenstrip thats why i need to fix it to be cool like you c:

Reduce the total frame rate or lower the resolution

@fayer1688 Thanks, I was able to generate 576x576, that seems to be fine.

@yajukun Don't set it to a square resolution, crop the image to a rectangle, use a rectangular resolution, and you can set the height higher. I tried 896x496 num_frames 101 and it worked fine

@fayer1688 Oh, let me try that! Thanks!

@fayer1688 I just tried 896x496 @101 frames and got OOM error? Not sure if it's becuase I am using 3 Loras including the new img2vid one? Maybe I could do 101 is I only use x2 Loras? Seems like 73 frames works with 3 Loras tho.

@yajukun Don't use teacache, it will use a lot of extra video memory

@fayer1688 Oh, interesting! I will try to bypass it and test it out again. Thanks!

same here

non-stop now even on the original django workflows. Something changed since the first i2v examples.. Even when I seemingly randomly get it to work, all I get is blackscreens anyway.

Hi, I think we have seen a similar issue before, it had to do with having to downgrade transformers. Take a look at the comments section for this article: https://stable-diffusion-art.com/hunyuan-video-ip2v/

Were you able to get it working? I have a 4090 not 3090 but same 24GB VRAM and I can run it with the 544p img2vid Lora plus 2 other Loras. Per @fayer1688 suggestion, I use 896x496. If you are getting OOM issues maybe try use the smaller 240p img2vid Lora and a smaller image and see if that runs?

@yajukun Not exactly, I just completly reinstalled Comfy and now it works. It's slow as hell, but it works. It's just most of the time the output is very fragmented and I don't know why.

@Jellon So today I tried another members Hunyuan WF that had missing nodes, when I installed them it started uninstalling my triton and sage, borked my comfy. So I too had to reinstall today. I did noticed that my intial reinstall ran this img2vid workflow w/o popup errors but all my output looks like ghosts that are glitching out. Is that what your output looks like? I remember I missed that config.ini security setting in the Manager and wasn't sure if that borked my setting up of sage, triton, hunyuan nodes? So I removed all the nodes and reinstalled everything. All my crap seems to work ok now. I just tested 896x496 @73 frames (using the 544p img2vid Lora and x2 other Loras), I finish it in about 1:40 with teacache and about 2:45 w/ teacache bypassed. I can also confirm that @fayer1688 was correct, by turning off teacache I was able to extend my frames out to 101 w/o getting OOM errors.

@yajukun I don't think so. A PC restart seems to have fixed the issue for now without me doing anything. I restarted comfy a couple of times already, but that didn't help. Weird issue. I hope it doesn't come up again.

@Jellon Something that I just noticed in my cmd window when I queue my prompts, if the Request to Load HyVideoModle = False, the output will be borked no matter how many time I generate. It will not give me a GUI error, it will allow me to keep generating garbage. I adjust Lora strengths and it reloaded the model, then I'm good to go. Glad you got it working.

For anyone else getting this error remove --lowvram from the startup batch file.

However I am unable to get satisfactory output from any kijai based workflows. I either get OOM's or garbled output at lower res/frames. I can get up to 720x720 at 129 frames on a 4080 Super on a 5800X3D with only 32GB of DDR4. Yet on kijai I get OOM's eveb at 480x480.

If someone can work out how to get a gguf and custom sampler version of this workflow it would be a godsend.

same thing is happening to me. did you find the problem?

Try this fix https://github.com/kijai/ComfyUI-HunyuanVideoWrapper/issues/118

One of the fixes fixed my problem.

AMD users will have an issue using SDPA with torch version < 2.5.1 giving black screen, and not choosing bf16 in the Hunyuan model loader which also can give black screen.

I initially had trouble running Kija's nodes on my 12GB 4080 too until I set the HunYuanVideo Model Loader to auto_cpu_offload and turn off upcast_rope. In addition I enabled quantization for the text encoder (bnb_nf4) by installing bitsandbytes on Windows.

If I didn't enable quantization my 32 GB RAM would be entirely in use by the llm and clip models and my VRAM would fill up entirely. With quantization turned on the llm and clip models only take around ~15 GB and some VRAM. When it's loading the main video model that memory gets released back to the pool.

Of course resolution and number of frames can still cause OOM's, reduce those until you use around 90/95% of VRAM when the video is being generated.

@funscripter627 Thanks for the tips! The first couple config changes got me a little leeway but it's still running OOM very often. I installed bitsandbytes using the pip installer in comfyui manager and am getting an error "module 'bitsandbytes' has no attribute 'nn'". Is there a parameter that I have to configure for defining this? Thanks.

@radiantResistor No problem man. I'm unfortunately unfamiliar with that specific error. Maybe try to ask ChatGPT. I think I installed it by opening my venv and executing the pip command. I'm using bitsandbytes-0.45.1 if that helps.

For me the error was gone after replacing the timm folder in mentioned path with version 1.0.13

I was able to get sage attention installed on windows. It won't work if you're running comfy in Stability Matrix because there's a bug in how a python package is loaded. You need to install triton which as a bunch of dependencies then it should install okay. https://github.com/woct0rdho/triton-windows Just carefully follow the directions under "install from wheel". It's worth the pain; a real boost in speed and capabilities.

@radiantResistor using Stability Matrix 🤣

Maybe i got it run, but now i stuck on vram allocation. Have only 12G.

It means there is a res mismatch.The image size and the sampler size is different. Do a image resize for your input image and connect the width and height outputs from the resizer node to the sampler inputs.

@jtsanborn Thanks I´ll try.

same

it's not the same thing at all...Here it's i2v, hunyuangp is for t2v and it can't be transposed to comfyui.

All that is is fast hy compiled + sageattn in its own gradio env. They make it seem like something new you can easily recreate it in comfy.