app.py

#!/usr/bin/env python
"""
Gradio demo for Wan2.1 FLF2V – First & Last Frame → Video
"""

import os
# Persist HF cache on /mnt/data so it survives across launches
os.environ["HF_HOME"] = "/mnt/data/huggingface"

import numpy as np
import torch
import gradio as gr
from diffusers import WanImageToVideoPipeline, AutoencoderKLWan
from diffusers.utils import export_to_video
from transformers import CLIPVisionModel
from PIL import Image
import torchvision.transforms.functional as TF

# ----------------------------------------------------------------------
# CONFIG
# ----------------------------------------------------------------------
MODEL_ID       = "Wan-AI/Wan2.1-FLF2V-14B-720P-diffusers"
DTYPE          = torch.float16
MAX_AREA       = 1280 * 720
DEFAULT_FRAMES = 81

# ----------------------------------------------------------------------
# PIPELINE LOADING
# ----------------------------------------------------------------------
def load_pipeline():
    # 1) load CLIP image encoder in full precision
    image_encoder = CLIPVisionModel.from_pretrained(
        MODEL_ID, subfolder="image_encoder", torch_dtype=torch.float32
    )
    # 2) load VAE in reduced precision
    vae = AutoencoderKLWan.from_pretrained(
        MODEL_ID, subfolder="vae", torch_dtype=DTYPE
    )
    # 3) load the WanImageToVideo pipeline, balanced across GPU/CPU
    pipe = WanImageToVideoPipeline.from_pretrained(
        MODEL_ID,
        vae=vae,
        image_encoder=image_encoder,
        torch_dtype=DTYPE,
        device_map="balanced",            # auto-offload large modules to CPU
    )
    # 4) reduce VAE peaks & enable CPU offload for everything else
    pipe.enable_vae_slicing()
    pipe.enable_model_cpu_offload()
    return pipe

# create once, at import time
PIPE = load_pipeline()

# ----------------------------------------------------------------------
# IMAGE PREPROCESSING UTILS
# ----------------------------------------------------------------------
def aspect_resize(img: Image.Image, max_area=MAX_AREA):
    ar  = img.height / img.width
    # ensure multiple of patch size
    mod = PIPE.vae_scale_factor_spatial * PIPE.transformer.config.patch_size[1]
    h   = round(np.sqrt(max_area * ar)) // mod * mod
    w   = round(np.sqrt(max_area / ar)) // mod * mod
    return img.resize((w, h), Image.LANCZOS), h, w

def center_crop_resize(img: Image.Image, h, w):
    ratio = max(w / img.width, h / img.height)
    img = img.resize(
        (round(img.width * ratio), round(img.height * ratio)),
        Image.LANCZOS
    )
    return TF.center_crop(img, [h, w])

# ----------------------------------------------------------------------
# GENERATION FUNCTION
# ----------------------------------------------------------------------
def generate(
    first_frame: Image.Image,
    last_frame:  Image.Image,
    prompt:      str,
    negative:    str,
    steps:       int,
    guidance:    float,
    num_frames:  int,
    seed:        int,
    fps:         int,
    progress=    gr.Progress()
):
    # seed
    if seed == -1:
        seed = torch.seed()
    gen = torch.Generator(device=PIPE.device).manual_seed(seed)

    # initial progress
    progress(0, steps, desc="Preprocessing images")

    # resize / crop
    first_frame, h, w = aspect_resize(first_frame)
    if last_frame.size != first_frame.size:
        last_frame = center_crop_resize(last_frame, h, w)

    # callback to update progress bar on each denoising step
    def progress_callback(step, timestep, latents):
        progress(step, steps, desc=f"Inference step {step}/{steps}")

    # run the pipeline (streams progress via callback)
    result = PIPE(
        image=first_frame,
        last_image=last_frame,
        prompt=prompt,
        negative_prompt=negative or None,
        height=h,
        width=w,
        num_frames=num_frames,
        num_inference_steps=steps,
        guidance_scale=guidance,
        generator=gen,
        callback=progress_callback,
    )

    # assemble and export to video
    frames     = result.frames[0]  # list of PIL images
    video_path = export_to_video(frames, fps=fps)

    # return video and seed used (Gradio will auto-download the .mp4)
    return video_path, seed

# ----------------------------------------------------------------------
# GRADIO UI
# ----------------------------------------------------------------------
with gr.Blocks(theme=gr.themes.Soft()) as demo:
    gr.Markdown("## Wan 2.1 FLF2V – First & Last Frame → Video")

    with gr.Row():
        first_img = gr.Image(label="First frame", type="pil")
        last_img  = gr.Image(label="Last frame",  type="pil")

    prompt   = gr.Textbox(label="Prompt", placeholder="A blue bird takes off…")
    negative = gr.Textbox(label="Negative prompt (optional)", placeholder="ugly, blurry")

    with gr.Accordion("Advanced parameters", open=False):
        steps      = gr.Slider(10, 50, value=30,   step=1,   label="Sampling steps")
        guidance   = gr.Slider(0.0, 10.0, value=5.5, step=0.1, label="Guidance scale")
        num_frames = gr.Slider(16, 129, value=DEFAULT_FRAMES, step=1, label="Number of frames")
        fps        = gr.Slider(4, 30, value=16,    step=1,   label="FPS (export)")
        seed       = gr.Number(value=-1, precision=0, label="Seed (-1 = random)")

    run_btn   = gr.Button("Generate")
    video_out = gr.Video(label="Result (.mp4)")
    used_seed = gr.Number(label="Seed used", interactive=False)

    run_btn.click(
        fn=generate,
        inputs=[ first_img, last_img, prompt, negative,
                 steps, guidance, num_frames, seed, fps ],
        outputs=[ video_out, used_seed ]
    )

# no special queue args needed
demo.launch()