Hunyuan Video

I'm on a 3090. Personally, I tend to use FP8 everything (hunyuan 720 distill e4m3fn, llama here), except for clip (some unknown old file I had lying around, just 240mb), and vae at bf16 (tiled) - even with teacache and sage attention, you're still looking at minimum 16gb to near saturation for a decent (say 600x800) 5 second video, so I'd rather not waste that on a few extra bits. I haven't really had much luck in the past with lower quantisations and gguf, but I might give those a shot too should the opportunity present itself.

There are some workflows which lower vram further, but they're virtually all based on kjnodes which I dislike. The vanilla nodes don't require much fiddling - I use rgthree's lora loader, welltop's teacache, bleh's sage attention, and the rest is just standard extras from the sample workflow I copied (video manipulation, pythongosss etc).

I also have a 3080 and had the same issue as you. Try following the steps here - https://blog.comfy.org/p/running-hunyuan-with-8gb-vram-and Worked for me!

I use a 2080ti and it works for me, so yes. BUT, only if you have the right settings. You're not going to get high resolutions unless you're willing to wait an hour for a 3second video unfortunately. Try these tweeks to the default workflow.

1. Use the 720p model or the FP8 model with the "Fast Video" lora.
2. Change the step-count to 12-15 from 20 (I've noticed too much noise if I go lower but I've read you can)
3. Change the video length to around 50.
4. Make sure you're using the FP8 model, clip and vae.
5. Change the resolution to 304x304. (Biggest one.)

Follow these 5 steps and I guarantee you, you'll get a video. Even though the resolution sounds horrible, the videos actually come out very well. It's like a mini HD video. Make sure you put in quality-of-life prompts at the end. Much like you would for images. I've just used 3 at the end of my video prompts and have had great results. They are: [Highest resolution, highest quality, sharp focus]

These settings for me on 2080ti make short videos that aren't bad. They take about 3mins each. This at least allows me to experiment and learn the model until I'm able to upgrade my GPU. You'll most likely get better results with a 3080 over my 2080ti but don't expect much more than I've listed here as both of our cards only have 12GB VRAM and that's by far the biggest bottleneck when making these videos.

Based off your results, you can always experiment with higher resolutions, longer lengths, more steps, or the FP16 models/clips/vae.

(Pro tip, sometimes cmd prompt will get stuck if you're trying something beyond your systems capabilities or have multiple programs running while waiting for the video to render. Open your CMD that's running comfyi and press enter.)

Hope this helps and gl with your videos!

https://comfyanonymous.github.io/ComfyUI_examples/hunyuan_video/

I'm wondering the same thing but for the best speeds I think you should use FP8 and if you are unsure about which model then look at which one is being used the most in the gallery. That should give you an idea about which models are preffered.

worked, waited 26 hours for a 6s video

workflow?

workflow?

try the Wan2.1 model. Its a little faster than hunyuan

@yofoton174609 didnt know, thank you, anyways, is a laptop1050so im expecting massive heat and at least 4 hours per generation xD

@goodstrk got one form reddit, how I can upload comfyui workflows here? sorry im new to this, still stuck with forge and automatic1111 (those dont destroy my poor laptop as comfyui does when I try to do something)

@yofoton174609 work on anything with 4gb of vram? im interested as even flux take 2 hours per image + upscale

@dingapriya23 i found one, but my setup cant handle it lol. thank you though!

@yofoton174609 What's the best currently out for img2video? Cuz I've tried the LTX workflow for img2video, and honestly, it seemed to completely ignore the image, and the output was like a bad 90's shockwave flash animation made by a child.

@Lazman I don't think LTX is so good

@yofoton174609 Yea, lol. AI moves fast, and it's been two months since I wrote that. I found out since that Hunyuan works much better, and I'm hearing good things about WAN, but I haven't tried it myself yet.

This video will help you out :)
https://youtu.be/0jdFf74WfCQ

runpod didnt match my expectations. The pods comfyui are outdated and not updateable or there is no access through http worst case not any connection option turns into ready mode but even then the pod(running)time is charged. Would like to know ur experiences

@sikasolutionsworldwide709 IYou can access it through a browser, wdym?... I personally like it, I don't use it too often since my 3060 TI is good enough for SDXL, so I am mainly going to use it for training, Hunyuan or sometimes even heavy image process. Maybe be only $0.44 an hour but still adds up for long training or video. A40 is jussstt over 3x faster than my 3060 TI.

I've had problems with my pod, like it wouldn't start or suddenly I don't have a GPU, so I found my solution. What I did was just create a network volume(persistent/permanent and 0.07cent/GB per month) to mount to the deployed pod. First Run the Ultimate Comfyui invoke ai template, let it do its work, now you have a permanent comfyui installation on your network volume, now all you need to run is the Run Pytorch 2.4 template. But you'll need to make a script to get the requirements installed again on the venv and activate it, then after that you only need a script to just activate venv and run main.py with --listen. Now you update the Comfyui get all your models etc. and you can deploy and terminate pods when you want (you will since you can't "stop" a pod with network volume mounted, but Run Pytorch starts fast, so you can terminate when you are not using it, and you have everything you need on the network volume, just be sure to add the 8188 port when deploying or edit it later.

@Felldude  BF16 it is. What workflow would you recommend?

@rgoobty69 thx for going so deep into details. Runpod was introduced to me via a small tutorial. In this tutorial the established pod has had the option to run comfyui through a browser this option was available by clicking the connect button other options for example were Jupiter notebook and terminal access. The option U presented seems to be the best but I need time to follow the steps as I don't have the know how yet. So for Hunyuan I am still sticking with my 4070. I am still waiting for runpod support reply.

放到c盘试试

I have the same errors

3060ti g6x 8g create video need 6min

C 盘

我下到d盘了