Guiders support for Wan

[a-r-r-o-w]

This PR shows the changes required to make existing pipelines compatible with guiders.

Example results with Wan 1.3B:

Guidance methods
APG	CFG	CFG-Zero*	PAG	SLG
AdaptiveProjectedGuidance.mp4	ClassifierFreeGuidance.mp4	ClassifierFreeZeroStarGuidance.mp4	PerturbedAttentionGuidance.mp4	SkipLayerGuidance.mp4

^{A frog doing his taxes while he sits on a swing!}

code

import argparse
from pathlib import Path

import torch
from diffusers import AutoencoderKLWan, WanPipeline
from diffusers.hooks import apply_first_block_cache, FirstBlockCacheConfig
from diffusers.guiders import AdaptiveProjectedGuidance
from diffusers.guiders import ClassifierFreeGuidance
from diffusers.guiders import ClassifierFreeZeroStarGuidance
from diffusers.guiders import PerturbedAttentionGuidance
from diffusers.guiders import SkipLayerGuidance
from diffusers.hooks import LayerSkipConfig
from diffusers.utils import export_to_video


def main(args):
    output_dir = Path(args.output_dir)
    output_dir.mkdir(parents=True, exist_ok=True)
    
    model_id = "Wan-AI/Wan2.1-T2V-1.3B-Diffusers"
    vae = AutoencoderKLWan.from_pretrained(model_id, subfolder="vae", torch_dtype=torch.float32)
    pipe = WanPipeline.from_pretrained(model_id, vae=vae, torch_dtype=torch.bfloat16)
    pipe.to("cuda")

    if args.cache_threshold is not None:
        pipe.transformer.enable_cache(FirstBlockCacheConfig(threshold=args.cache_threshold))

    cfg = ClassifierFreeGuidance(guidance_scale=5.0)

    slg = SkipLayerGuidance(
        guidance_scale=4.0,
        skip_layer_guidance_scale=2.0,
        skip_layer_guidance_start=0.01,
        skip_layer_guidance_stop=0.2,
        skip_layer_config=LayerSkipConfig(indices=[11, 16]),
    )

    cfgz = ClassifierFreeZeroStarGuidance(
        guidance_scale=5.0,
        zero_init_steps=1,
    )

    apg = AdaptiveProjectedGuidance(
        guidance_scale=12.0,
        adaptive_projected_guidance_momentum=0.2,
        adaptive_projected_guidance_rescale=15.0,
        eta=1.0,
        guidance_rescale=0.0,
        use_original_formulation=False,
    )

    pag = PerturbedAttentionGuidance(
        pag_applied_layers="blocks.11",
        guidance_scale=4.0,
        pag_scale=1.5,
    )

    prompt = "A frog doing his taxes while he sits on a swing!"
    negative_prompt = "Bright tones, overexposed, static, blurred details, subtitles, style, works, paintings, images, static, overall gray, worst quality, low quality, JPEG compression residue, ugly, incomplete, extra fingers, poorly drawn hands, poorly drawn faces, deformed, disfigured, misshapen limbs, fused fingers, still picture, messy background, three legs, many people in the background, walking backwards"

    cfg_methods = [cfg, slg, cfgz, apg, pag]

    for guidance in cfg_methods:
        video = pipe(prompt=prompt, negative_prompt=negative_prompt, guidance=guidance, generator=torch.Generator().manual_seed(42)).frames[0]
        if args.cache_threshold is not None:
            output_path = output_dir / f"{guidance.__class__.__name__}_cache_threshold_{args.cache_threshold:.3f}.mp4"
        else:
            output_path = output_dir / f"{guidance.__class__.__name__}.mp4"
        export_to_video(video, output_path, fps=16)


def get_args():
    parser = argparse.ArgumentParser()
    parser.add_argument("--output_dir", type=str, default="basic_comparison_wan")
    parser.add_argument("--cache_threshold", type=float, default=None)
    return parser.parse_args()


if __name__ == "__main__":
    args = get_args()
    main(args)

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.