可以在comfyui里面用吗
thanks
I just get black outputs when i try using these. How are they supposed to be used?
https://raw.githubusercontent.com/Comfy-Org/workflow_templates/refs/heads/main/templates/video_hunyuan_video_1.5_720p_i2v.json
https://raw.githubusercontent.com/Comfy-Org/workflow_templates/refs/heads/main/templates/video_hunyuan_video_1.5_720p_t2v.json
That doesn't help. The models do not work with the "official" workflow. The outputs are just black.
How many steps and what is the sr type ? hy15_1080p_sr_cfg_distiled_fp8_e4m3_lightx2v.safetensors
also with default wf
Just another great release from Lightx2v in the end. I don't mean to be ungrateful and I appreciate the work, but for real, anything you release is an absolute mess that barely works. The Wan LighX2V loras were, and still are a mess, the LightVAE repo was a horror show and did not work for weeks until fixed, now this...? Come on.
@Scratchproof
tbh it's not meant to work with comfyui atm I think based on Hy1.5 docs , so kinda depends on comfy side
It's friday , give it a few days
@Scratchproof tbh it's not meant to work with comfyui atm I think based on Hy1.5 docs , so kinda depends on comfy side
It's friday , give it a few days
Yeah that's fair, I don't mean to sound like a dick. Just maybe at least if we had a model card or something, you know.
Couldn't be happier with LightX myself ;-)
Its one of the best thing that ever happened to local open-source video generation, without a doubt.
With the loras i can generate videos in insanely few steps, its like magic how it even is possible ;-)
Generate a video that looks great in few steps in a couple of minutes, vs waiting 20 minute for a 5 second video ... it makes all the difference in the world ;)
So a huge thank you to LightX for sure
Anyone ungrateful for their work should go try generate videos with 50 steps again and wait 20 minutes for a 5 second clip, and see how much fun that is ;-)
Anyone ungrateful for their work should go try generate videos with 50 steps again and wait 20 minutes for a 5 second clip, and see how much fun that is ;-)
They would use 20-30* steps
I just get black outputs when i try using these. How are they supposed to be used?
You can always use these for now
https://huggingface.co/Comfy-Org/HunyuanVideo_1.5_repackaged Both fp16 and fp8
Maybe ComfyUI will update to support the LightX models too, although they are probably intended for LightX own codebase
https://huggingface.co/lightx2v/Hy1.5-Quantized-Models/discussions/2#6921bbee921df8e6e4f81d4f
They would use 20-30* steps
most models have 50 steps as recommended
Although, 20-30 steps can be ok
But anyways, lightx makes life so much easier, only needing a few steps
There are no step distilled models yet, you aren't missing out anything in ComfyUI from not being able to use these models here, they are same models just with different quantization, they would use same step count anyway.
Comfy native distill model in 8 steps is fine to me ,just a toy but enough in most cases
Comfy native distill model in 8 steps is fine to me
yes, I'm also getting really good results with 8-12 steps euler_a / beta sampler destilled model (Image 2 video)
Good enough for testing ideas etc for sure, before rendering out a full step one with upscale
Newest still black,no need to try
@Kijai thanks for your 4 step lora, i2v can run but got bad color. https://huggingface.co/Comfy-Org/HunyuanVideo_1.5_repackaged