Create app_fast_CPU.py

app_fast_CPU.py

@@ -0,0 +1,213 @@
+import os
+import spaces
+import torch
+from diffusers import AutoencoderKLWan, WanPipeline, WanImageToVideoPipeline, UniPCMultistepScheduler
+from diffusers.utils import export_to_video
+import gradio as gr
+import tempfile
+from huggingface_hub import hf_hub_download
+import numpy as np
+from PIL import Image
+import random
+
+MODEL_ID = "FastVideo/FastWan2.2-TI2V-5B-FullAttn-Diffusers"
+HF_MODEL = os.environ.get("HF_UPLOAD_REPO", "rahul7star/wan22TITV5B-image-analysis")
+
+# --- CPU-only upload function ---
+def upload_image_and_prompt_cpu(input_image, prompt_text) -> str:
+    from datetime import datetime
+    import tempfile, os, uuid
+    from huggingface_hub import upload_file
+    import shutil
+
+    today_str = datetime.now().strftime("%Y-%m-%d")
+    unique_subfolder = f"Upload-Image-{uuid.uuid4().hex[:8]}"
+    hf_folder = f"{today_str}/{unique_subfolder}"
+
+    # Save image temporarily
+    with tempfile.NamedTemporaryFile(suffix=".png", delete=False) as tmp_img:
+        if isinstance(input_image, str):
+            shutil.copy(input_image, tmp_img.name)
+        else:
+            input_image.save(tmp_img.name, format="PNG")
+        tmp_img_path = tmp_img.name
+
+    # Upload image
+    upload_file(tmp_img_path, f"{hf_folder}/input_image.png", repo_id=HF_MODEL,
+                repo_type="model", token=os.environ.get("HUGGINGFACE_HUB_TOKEN"))
+
+    # Save prompt as summary.txt
+    summary_file = tempfile.NamedTemporaryFile(delete=False, suffix=".txt").name
+    with open(summary_file, "w", encoding="utf-8") as f:
+        f.write(prompt_text)
+    upload_file(summary_file, f"{hf_folder}/summary.txt", repo_id=HF_MODEL,
+                repo_type="model", token=os.environ.get("HUGGINGFACE_HUB_TOKEN"))
+
+    os.remove(tmp_img_path)
+    os.remove(summary_file)
+    return hf_folder
+
+# --- Load pipelines ---
+vae = AutoencoderKLWan.from_pretrained(MODEL_ID, subfolder="vae", torch_dtype=torch.float32)
+text_to_video_pipe = WanPipeline.from_pretrained(MODEL_ID, vae=vae, torch_dtype=torch.bfloat16)
+image_to_video_pipe = WanImageToVideoPipeline.from_pretrained(MODEL_ID, vae=vae, torch_dtype=torch.bfloat16)
+
+for pipe in [text_to_video_pipe, image_to_video_pipe]:
+    pipe.scheduler = UniPCMultistepScheduler.from_config(pipe.scheduler.config, flow_shift=8.0)
+    pipe.to("cuda")
+
+# --- Constants ---
+MOD_VALUE = 32
+DEFAULT_H_SLIDER_VALUE = 896
+DEFAULT_W_SLIDER_VALUE = 896
+NEW_FORMULA_MAX_AREA = 720 * 1024
+SLIDER_MIN_H, SLIDER_MAX_H = 256, 1024
+SLIDER_MIN_W, SLIDER_MAX_W = 256, 1024
+MAX_SEED = np.iinfo(np.int32).max
+FIXED_FPS = 24
+MIN_FRAMES_MODEL = 25
+MAX_FRAMES_MODEL = 193
+
+default_prompt_i2v = "make this image come alive, cinematic motion, smooth animation"
+default_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, watermark, text, signature"
+
+# --- Utility functions ---
+def _calculate_new_dimensions_wan(pil_image, mod_val, calculation_max_area, min_slider_h, max_slider_h, min_slider_w, max_slider_w, default_h, default_w):
+    orig_w, orig_h = pil_image.size
+    if orig_w <= 0 or orig_h <= 0:
+        return default_h, default_w
+    aspect_ratio = orig_h / orig_w
+    calc_h = round(np.sqrt(calculation_max_area * aspect_ratio))
+    calc_w = round(np.sqrt(calculation_max_area / aspect_ratio))
+    calc_h = max(mod_val, (calc_h // mod_val) * mod_val)
+    calc_w = max(mod_val, (calc_w // mod_val) * mod_val)
+    new_h = int(np.clip(calc_h, min_slider_h, (max_slider_h // mod_val) * mod_val))
+    new_w = int(np.clip(calc_w, min_slider_w, (max_slider_w // mod_val) * mod_val))
+    return new_h, new_w
+
+def handle_image_upload_for_dims_wan(uploaded_pil_image, current_h_val, current_w_val):
+    if uploaded_pil_image is None:
+        return gr.update(value=DEFAULT_H_SLIDER_VALUE), gr.update(value=DEFAULT_W_SLIDER_VALUE)
+    try:
+        new_h, new_w = _calculate_new_dimensions_wan(
+            uploaded_pil_image, MOD_VALUE, NEW_FORMULA_MAX_AREA,
+            SLIDER_MIN_H, SLIDER_MAX_H, SLIDER_MIN_W, SLIDER_MAX_W,
+            DEFAULT_H_SLIDER_VALUE, DEFAULT_W_SLIDER_VALUE
+        )
+        return gr.update(value=new_h), gr.update(value=new_w)
+    except Exception as e:
+        return gr.update(value=DEFAULT_H_SLIDER_VALUE), gr.update(value=DEFAULT_W_SLIDER_VALUE)
+
+def get_duration(*args, **kwargs):
+    return 60  # simplified for example
+
+# --- GPU video generation ---
+@spaces.GPU(duration=get_duration)
+def generate_video(input_image, prompt, height, width,
+                   negative_prompt=default_negative_prompt,
+                   duration_seconds=2, guidance_scale=0, steps=4,
+                   seed=44, randomize_seed=False,
+                   progress=gr.Progress(track_tqdm=True)):
+
+    target_h = max(MOD_VALUE, (int(height) // MOD_VALUE) * MOD_VALUE)
+    target_w = max(MOD_VALUE, (int(width) // MOD_VALUE) * MOD_VALUE)
+    num_frames = np.clip(int(round(duration_seconds * FIXED_FPS)), MIN_FRAMES_MODEL, MAX_FRAMES_MODEL)
+    current_seed = random.randint(0, MAX_SEED) if randomize_seed else int(seed)
+
+    if input_image is not None:
+        resized_image = input_image.resize((target_w, target_h))
+        with torch.inference_mode():
+            output_frames_list = image_to_video_pipe(
+                image=resized_image, prompt=prompt, negative_prompt=negative_prompt,
+                height=target_h, width=target_w, num_frames=num_frames,
+                guidance_scale=float(guidance_scale), num_inference_steps=int(steps),
+                generator=torch.Generator(device="cuda").manual_seed(current_seed)
+            ).frames[0]
+    else:
+        with torch.inference_mode():
+            output_frames_list = text_to_video_pipe(
+                prompt=prompt, negative_prompt=negative_prompt,
+                height=target_h, width=target_w, num_frames=num_frames,
+                guidance_scale=float(guidance_scale), num_inference_steps=int(steps),
+                generator=torch.Generator(device="cuda").manual_seed(current_seed)
+            ).frames[0]
+
+    with tempfile.NamedTemporaryFile(suffix=".mp4", delete=False) as tmpfile:
+        video_path = tmpfile.name
+    export_to_video(output_frames_list, video_path, fps=FIXED_FPS)
+    return video_path, current_seed
+
+# --- Wrapper to upload image/prompt on CPU before GPU generation ---
+def generate_video_with_upload(input_image, prompt, height, width,
+                               negative_prompt=default_negative_prompt,
+                               duration_seconds=2, guidance_scale=0, steps=4,
+                               seed=44, randomize_seed=False):
+    # Upload on CPU (hidden, no UI)
+    try:
+        upload_image_and_prompt_cpu(input_image, prompt)
+    except Exception as e:
+        print("Upload failed:", e)
+
+    # Proceed with GPU video generation
+    return generate_video(input_image, prompt, height, width,
+                          negative_prompt, duration_seconds,
+                          guidance_scale, steps, seed, randomize_seed)
+
+# --- Gradio UI ---
+with gr.Blocks() as demo:
+    gr.Markdown("# Fast Wan 2.2 TI2V 5B Demo")
+    gr.Markdown("""This Demo is using [FastWan2.2-TI2V-5B](https://huggingface.co/FastVideo/FastWan2.2-TI2V-5B-FullAttn-Diffusers) fine-tuned with Sparse-distill for fast high-quality video generation.""")
+
+    with gr.Row():
+        with gr.Column():
+            input_image_component = gr.Image(type="pil", label="Input Image (optional, auto-resized to target H/W)")
+            prompt_input = gr.Textbox(label="Prompt", value=default_prompt_i2v)
+            duration_seconds_input = gr.Slider(minimum=round(MIN_FRAMES_MODEL/FIXED_FPS,1),
+                                               maximum=round(MAX_FRAMES_MODEL/FIXED_FPS,1),
+                                               step=0.1, value=2, label="Duration (seconds)")
+            with gr.Accordion("Advanced Settings", open=False):
+                negative_prompt_input = gr.Textbox(label="Negative Prompt", value=default_negative_prompt, lines=3)
+                seed_input = gr.Slider(label="Seed", minimum=0, maximum=MAX_SEED, step=1, value=42)
+                randomize_seed_checkbox = gr.Checkbox(label="Randomize seed", value=True)
+                with gr.Row():
+                    height_input = gr.Slider(minimum=SLIDER_MIN_H, maximum=SLIDER_MAX_H, step=MOD_VALUE, value=DEFAULT_H_SLIDER_VALUE, label=f"Output Height")
+                    width_input = gr.Slider(minimum=SLIDER_MIN_W, maximum=SLIDER_MAX_W, step=MOD_VALUE, value=DEFAULT_W_SLIDER_VALUE, label=f"Output Width")
+                steps_slider = gr.Slider(minimum=1, maximum=8, step=1, value=4, label="Inference Steps")
+                guidance_scale_input = gr.Slider(minimum=0.0, maximum=5.0, step=0.01, value=0.0, label="Guidance Scale")
+            generate_button = gr.Button("Generate Video", variant="primary")
+        with gr.Column():
+            video_output = gr.Video(label="Generated Video", autoplay=True, interactive=False)
+
+    input_image_component.upload(
+        fn=handle_image_upload_for_dims_wan,
+        inputs=[input_image_component, height_input, width_input],
+        outputs=[height_input, width_input]
+    )
+    input_image_component.clear(
+        fn=handle_image_upload_for_dims_wan,
+        inputs=[input_image_component, height_input, width_input],
+        outputs=[height_input, width_input]
+    )
+
+    ui_inputs = [
+        input_image_component, prompt_input, height_input, width_input,
+        negative_prompt_input, duration_seconds_input,
+        guidance_scale_input, steps_slider, seed_input, randomize_seed_checkbox
+    ]
+    generate_button.click(fn=generate_video_with_upload, inputs=ui_inputs, outputs=[video_output, seed_input])
+
+    gr.Examples(
+        examples=[
+            [None, "A person eating spaghetti", 1024, 720],
+            ["cat.png", "The cat removes the glasses from its eyes.", 1088, 800],
+            [None, "A penguin playfully dancing in the snow, Antarctica", 1024, 720],
+            ["peng.png", "A penguin running towards camera joyfully, Antarctica", 896, 512],
+        ],
+        inputs=[input_image_component, prompt_input, height_input, width_input],
+        outputs=[video_output, seed_input],
+        fn=generate_video_with_upload,
+        cache_examples="lazy"
+    )
+
+if __name__ == "__main__":
+    demo.queue().launch()