[wip] I need to figure out why images are being generated as oversaturated/very red, think I am missing a video processing step after generation

Hi @luke14free , what would be the difference between using the normal pipeline with a boundary_ratio = 1.0 and num_frames=1 and this?

hey @asomoza it is pretty much the same, except you don't also load transformer (the high noise one). you can 100% make it work with existing pipelines. just thought it would have been cleaner to have a separate pipeline, but in case it's not i'll close

@luke14free not completely sure (haven't tested it yet) but I think you can pass a transformer=None so it doesn't load. You can keep it open if you want and see if there's some interest in having a specific pipeline for this but mostly we don't create additional pipelines when you can do the same with the base one.

indeed you can, I tested it myself and it's what I am doing here. I think you are right it makes little sense to have this, I'll close it anyways

vae = AutoencoderKLWan.from_pretrained(
    "Wan-AI/Wan2.2-T2V-A14B-Diffusers", 
    subfolder="vae", 
    torch_dtype=torch.float32
)

transformer_low_noise = WanTransformer3DModel.from_pretrained(
    "Wan-AI/Wan2.2-T2V-A14B-Diffusers",
    subfolder="transformer_2",
    torch_dtype=torch.bfloat16,
)

self.pipe = WanPipeline.from_pretrained(
    "Wan-AI/Wan2.2-T2V-A14B-Diffusers",
    vae=vae,
    transformer=transformer_low_noise,
    boundary_ratio=None,
    transformer_2=None,
    torch_dtype=torch.bfloat16
)

output = pipe(
    prompt=input_data.prompt,
    negative_prompt=input_data.negative_prompt,
    height=height,
    width=width,
    num_frames=1,  # Required for text-to-image to create proper temporal dimension
    guidance_scale=input_data.guidance_scale,
    num_inference_steps=input_data.num_inference_steps,
).frames[0]

for anyone interested, this is how you do it with the existing pipelines :)