Anyone tested this yet?
Does anyone know if this is worth downloading?
I tried 640*480 - 81Frames - T2V one step takes took 129.09s and there are 20 steps to generate - 4060 TI 16GB - Model used hunyuanvideo1.5_480p_t2v_cfg_distilled_fp8_scaled.safetensors 8.33 GB
Anyone can speed up?
Updated: it works well, took 90s to generate 48 frames.
I have the 5090, I was able to generate 5 seconds video in 284s at 24fps using the 720p model at 848*480
Whenever I tried to extend the frames from 121 to 245 for a 10s video, it took 498 seconds but the frames were still consistent and coherent the whole time
Then I tried 365 frames to see how it would go, for 15s video: 15 minutes generation time
Tried 1600 frames but OOM on my 5090 so now will try 1000 frames as a test
please share the workflow. Two hours of hot sex with comfy didn't give me any results :)
please share the workflow
Download the videos and open them in ComfyUI, there are the workflows
Two hours of hot sex with comfy didn't give me any results :)
Oh
thanks a lot, everything is working. It didn't work with the standard templates.
I found gguf models they work too
https://huggingface.co/jayn7/HunyuanVideo-1.5_I2V_720p-GGUF
https://huggingface.co/jayn7/HunyuanVideo-1.5_T2V_480p-GGUF
please share the workflow.
They are available within ComfyUI itself, in settings => Browse templates
The hunyuan documentation talks about the lightning loras for speeding it up, does anyone know where they are?
The hunyuan documentation talks about the lightning loras for speeding it up, does anyone know where they are?
I dont think there is any loras (..yet)...
But rather distilled models, cfg distilled, that lets you run interference faster with cfg set to 1 (but with same ish steps)
(but use the comfy ones here, they to work with comfy)
"We will soon update more models on our HuggingFace page, including step distillation"
When this is out, it might allow fewer steps as well
@Kijai
any chance the tiny vae could be made comfy compatible? not sure if its doable, maybe its only suitable for the lightx own codebase
Since the vae process is a little slow-ish with this model, curious if the tiny vae would be lots faster ;-)
https://huggingface.co/lightx2v/Autoencoders/blob/main/lighttaehy1_5.safetensors
The hunyuan documentation talks about the lightning loras for speeding it up, does anyone know where they are?
we don't provide lighting loras. you can control the light using the prompt as illustrated in the prompt handbook.
https://github.com/Tencent-Hunyuan/HunyuanVideo-1.5/blob/main/assets/HunyuanVideo_1_5_Prompt_Handbook_EN.md
The low step loras ("lightning") seems to be available already, at least for T2V 480p. The other formats are incoming soon I bet ;-)
https://huggingface.co/Comfy-Org/HunyuanVideo_1.5_repackaged/tree/main/split_files/loras
I have the 5090, I was able to generate 5 seconds video in 284s at 24fps using the 720p model at 848*480
Whenever I tried to extend the frames from 121 to 245 for a 10s video, it took 498 seconds but the frames were still consistent and coherent the whole time
Then I tried 365 frames to see how it would go, for 15s video: 15 minutes generation timeTried 1600 frames but OOM on my 5090 so now will try 1000 frames as a test
Did you set any parameters when using it? My 4090 graphics card has a very slow generation speed, and the default workflow for 720P Wensheng videos takes 20 minutes. Do I need to set additional parameters to speed up the process
Did you set any parameters when using it? My 4090 graphics card has a very slow generation speed, and the default workflow for 720P Wensheng videos takes 20 minutes. Do I need to set additional parameters to speed up the process
The CFG distilled models allows CFG set to 1 and is a lot faster.
Also there might be some low-step loras coming from LightX
They already made one for T2V 480p, that allows video inference with only 4 step
https://huggingface.co/Comfy-Org/HunyuanVideo_1.5_repackaged/tree/main/split_files/loras
They already made one for T2V 480p, that allows video inference with only 4 step
https://huggingface.co/Comfy-Org/HunyuanVideo_1.5_repackaged/tree/main/split_files/loras
I'm having too low contrast, faded results with this. Similar to early Wan2.2 lightning lora. Most notable if you try to render anything in low light or midnight.
Is there a simple workflow for 480p with 4step lora?
Is there a simple workflow for 480p with 4step lora?
Just add the node LoraLoaderModelOnly and connect that between the main model loader and the Sampler.
And if you want to speed up even more, tiny vae is now supported. The vae decode becomes blazing fast
Video below made in 4 steps, a little over 1 minute to generate (around 1:32) at 832x480p
Vae decode in 2-3 seconds or so
(workflow for 4-step lora in embedded in the video, so just drop it into comfyui)
(and if you run it through the Hunyuan upscaling it will look more detailed etc)
And if you want to speed up even more, tiny vae is now supported. The vae decode becomes blazing fast
Thanks! Also...your screenshot just shows standard vae node, but it doesn't work right? lighttaehy1_5 seem to require custom nodes?
I get error trying to install lightvae nodes. Standard vae is slow but acceptable
lighttaehy1_5 seem to require custom nodes?
Use the regular vae loader. But support was added in ComfyUI very recently, so you must update comfy to the very latest
Video below made in 4 steps, a little over 1 minute to generate (around 1:32) at 832x480p
Vae decode in 2-3 seconds or so(workflow for 4-step lora in embedded in the video, so just drop it into comfyui)
(and if you run it through the Hunyuan upscaling it will look more detailed etc)
Hello, I noticed that in the ComfyUI workflow of this video, is using the 720p Hunyuan/hunyuanvideo1.5_720p_t2v-Q8_0.gguf model with a 480p LoRA mounted to generate a 480p video, which seems a bit counterintuitive. Additionally, I'd like to confirm: is it true that this process involves 6 steps of sampling without CFG distillation, meaning there are actually 6*2=12 forward passes actually?
Hello, I noticed that in the ComfyUI workflow of this video, is using the 720p Hunyuan/hunyuanvideo1.5_720p_t2v-Q8_0.gguf model with a 480p LoRA mounted to generate a 480p video, which seems a bit counterintuitive. Additionally, I'd like to confirm: is it true that this process involves 6 steps of sampling without CFG distillation, meaning there are actually 6*2=12 forward passes actually?
The resolution in my workflow was set to 832x480 .. .so thats 480p ish
Usually the larger models supports both 720p and 480p. So thats why i often end up loading the bigger res one so i can alter between size without having to change model
With the low step lora its 4 steps, but I might have set it to 6 just to squeeze out a bit of extra details. And with low step loras the CFG is set to 1.
" meaning there are actually 6*2=12 forward passes actually". That I dont know. Could be, a little over my "pay grade" ;-) i'm not that technical to the inner workings
I wish there was a VAE similar to wan2.1. Official hunyuan vae is too huge, and lightvae is noticeably worse, at low resolution at least, which i usually use, because even 480p is too much for my system in terms of generation time.
STEP distilled version for I2V (480p) is out, with only 8 steps needed..
https://huggingface.co/Comfy-Org/HunyuanVideo_1.5_repackaged/blob/main/split_files/diffusion_models/hunyuanvideo1.5_480p_i2v_step_distilled_fp8_scaled.safetensors
https://huggingface.co/Comfy-Org/HunyuanVideo_1.5_repackaged/blob/main/split_files/diffusion_models/hunyuanvideo1.5_480p_i2v_step_distilled_fp16.safetensors
8 Steps, less than 2 mins to generate, Image 2 Video:
STEP distilled version for I2V (480p) is out, with only 8 steps needed..
https://huggingface.co/Comfy-Org/HunyuanVideo_1.5_repackaged/blob/main/split_files/diffusion_models/hunyuanvideo1.5_480p_i2v_step_distilled_fp8_scaled.safetensors
https://huggingface.co/Comfy-Org/HunyuanVideo_1.5_repackaged/blob/main/split_files/diffusion_models/hunyuanvideo1.5_480p_i2v_step_distilled_fp16.safetensors8 Steps, less than 2 mins to generate, Image 2 Video:
Happen to have the workflow for these as the one I have something was not right with these?