Create wan2fast_1.py

wan2fast_1.py

@@ -0,0 +1,175 @@
+import os
+
+# ✅ Patch for NVML-related crash in ZeroGPU
+os.environ["PYTORCH_NO_NVML"] = "1"
+
+# ✅ Ensure proper PyTorch version for CUDA 12.6 in Spaces
+os.system('pip install --upgrade --pre --extra-index-url https://download.pytorch.org/whl/nightly/cu126 "torch<2.9"')
+
+import torch
+from diffusers import AutoencoderKLWan, WanPipeline, UniPCMultistepScheduler
+from diffusers.utils import export_to_video
+import gradio as gr
+import tempfile
+import spaces
+from huggingface_hub import hf_hub_download
+import numpy as np
+import random
+
+# MODEL_ID
+MODEL_ID = "Runware/Wan2.2-T2V-A14B"
+
+# Load model and scheduler (no .to("cuda") yet)
+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.scheduler = UniPCMultistepScheduler.from_config(pipe.scheduler.config, flow_shift=8.0)
+
+# Configuration
+MOD_VALUE = 32
+DEFAULT_H_SLIDER_VALUE = 768
+DEFAULT_W_SLIDER_VALUE = 1344
+
+IS_ORIGINAL_SPACE = os.environ.get("IS_ORIGINAL_SPACE", "True") == "True"
+
+LIMITED_MAX_RESOLUTION = 640
+LIMITED_MAX_DURATION = 2.0
+LIMITED_MAX_STEPS = 4
+
+ORIGINAL_SLIDER_MIN_H, ORIGINAL_SLIDER_MAX_H = 128, 1536
+ORIGINAL_SLIDER_MIN_W, ORIGINAL_SLIDER_MAX_W = 128, 1536
+ORIGINAL_MAX_DURATION = round(81 / 24, 1)
+ORIGINAL_MAX_STEPS = 8
+
+if IS_ORIGINAL_SPACE:
+    SLIDER_MIN_H, SLIDER_MAX_H = 128, LIMITED_MAX_RESOLUTION
+    SLIDER_MIN_W, SLIDER_MAX_W = 128, LIMITED_MAX_RESOLUTION
+    MAX_DURATION = LIMITED_MAX_DURATION
+    MAX_STEPS = LIMITED_MAX_STEPS
+else:
+    SLIDER_MIN_H, SLIDER_MAX_H = ORIGINAL_SLIDER_MIN_H, ORIGINAL_SLIDER_MAX_H
+    SLIDER_MIN_W, SLIDER_MAX_W = ORIGINAL_SLIDER_MIN_W, ORIGINAL_SLIDER_MAX_W
+    MAX_DURATION = ORIGINAL_MAX_DURATION
+    MAX_STEPS = ORIGINAL_MAX_STEPS
+
+MAX_SEED = np.iinfo(np.int32).max
+FIXED_FPS = 24
+FIXED_OUTPUT_FPS = 18
+MIN_FRAMES_MODEL = 8
+MAX_FRAMES_MODEL = 81
+
+default_prompt_t2v = "cinematic footage, group of pedestrians dancing in the streets of NYC, high quality breakdance, 4K, tiktok video, intricate details, instagram feel, dynamic camera, smooth dance motion, dimly lit, stylish, beautiful faces, smiling, music video"
+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"
+
+def get_duration(prompt, height, width, negative_prompt, duration_seconds, guidance_scale, steps, seed, randomize_seed, progress):
+    return int(duration_seconds) * int(steps) * 2.25 + 5
+
+@spaces.GPU(duration=get_duration)
+def generate_video(prompt, height, width, 
+                   negative_prompt=default_negative_prompt, duration_seconds=2,
+                   guidance_scale=1, steps=4,
+                   seed=42, randomize_seed=False, 
+                   progress=gr.Progress(track_tqdm=True)):
+
+    if not prompt or prompt.strip() == "":
+        raise gr.Error("Please enter a text prompt. Try to use long and precise descriptions.")
+
+    if IS_ORIGINAL_SPACE:
+        height = min(height, LIMITED_MAX_RESOLUTION)
+        width = min(width, LIMITED_MAX_RESOLUTION)
+        duration_seconds = min(duration_seconds, LIMITED_MAX_DURATION)
+        steps = min(steps, LIMITED_MAX_STEPS)
+
+    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)
+
+    # ✅ Move to GPU inside @spaces.GPU function
+    pipe.to("cuda")
+
+    with torch.inference_mode():
+        output_frames_list = 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_OUTPUT_FPS)
+    return video_path, current_seed
+
+# Gradio UI
+with gr.Blocks(css="body { max-width: 100vw; overflow-x: hidden; }") as demo:
+    gr.HTML('<meta name="viewport" content="width=device-width, initial-scale=1">')
+    gr.Markdown("# ⚡ InstaVideo")
+    gr.Markdown("This Gradio space is a fork of [wan2-1-fast from multimodalart](https://huggingface.co/spaces/multimodalart/wan2-1-fast), and is powered by the Wan CausVid LoRA [from Kijai](https://huggingface.co/Kijai/WanVideo_comfy/blob/main/Wan21_CausVid_bidirect2_T2V_1_3B_lora_rank32.safetensors).")
+
+    if IS_ORIGINAL_SPACE:
+        gr.Markdown("⚠️ **This free public demo limits the resolution to 640px, duration to 2s, and inference steps to 4. For full capabilities please duplicate this space.**")
+
+    with gr.Row():
+        with gr.Column():
+            prompt_input = gr.Textbox(label="Prompt", value=default_prompt_t2v)
+            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, interactive=True)
+                randomize_seed_checkbox = gr.Checkbox(label="Randomize seed", value=True, interactive=True)
+                with gr.Row():
+                    height_input = gr.Slider(
+                        minimum=SLIDER_MIN_H, 
+                        maximum=SLIDER_MAX_H, 
+                        step=MOD_VALUE, 
+                        value=min(DEFAULT_H_SLIDER_VALUE, SLIDER_MAX_H), 
+                        label=f"Output Height (multiple of {MOD_VALUE})"
+                    )
+                    width_input = gr.Slider(
+                        minimum=SLIDER_MIN_W, 
+                        maximum=SLIDER_MAX_W, 
+                        step=MOD_VALUE, 
+                        value=min(DEFAULT_W_SLIDER_VALUE, SLIDER_MAX_W), 
+                        label=f"Output Width (multiple of {MOD_VALUE})"
+                    )
+                duration_seconds_input = gr.Slider(
+                    minimum=round(MIN_FRAMES_MODEL / FIXED_FPS, 1),
+                    maximum=MAX_DURATION,
+                    step=0.1,
+                    value=2,
+                    label="Duration (seconds)",
+                    info=f"Clamped to model's {MIN_FRAMES_MODEL}-{MAX_FRAMES_MODEL} frames at {FIXED_FPS}fps."
+                )
+                steps_slider = gr.Slider(minimum=1, maximum=MAX_STEPS, step=1, value=4, label="Inference Steps")
+                guidance_scale_input = gr.Slider(minimum=0.0, maximum=20.0, step=0.5, value=1.0, label="Guidance Scale", visible=False)
+            generate_button = gr.Button("Generate Video", variant="primary")
+        with gr.Column():
+            video_output = gr.Video(label="Generated Video", autoplay=True, interactive=False)
+
+    ui_inputs = [
+        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, inputs=ui_inputs, outputs=[video_output, seed_input])
+
+    example_configs = [
+        ["a majestic eagle soaring through mountain peaks, cinematic aerial view", 896, 512],
+        ["a serene ocean wave crashing on a sandy beach at sunset", 448, 832],
+        ["a field of flowers swaying in the wind, spring morning light", 512, 896],
+    ]
+    if IS_ORIGINAL_SPACE:
+        example_configs = [
+            [example[0], min(example[1], LIMITED_MAX_RESOLUTION), min(example[2], LIMITED_MAX_RESOLUTION)]
+            for example in example_configs
+        ]
+    gr.Examples(
+        examples=example_configs,
+        inputs=[prompt_input, height_input, width_input], 
+        outputs=[video_output, seed_input], 
+        fn=generate_video, 
+        cache_examples="lazy"
+    )
+
+if __name__ == "__main__":
+    demo.queue().launch()