import spaces import torch from diffusers import AutoencoderKLWan, WanImageToVideoPipeline, UniPCMultistepScheduler from diffusers.utils import export_to_video from transformers import CLIPVisionModel import gradio as gr import tempfile import os import subprocess from huggingface_hub import hf_hub_download import numpy as np from PIL import Image import random import warnings warnings.filterwarnings("ignore", message=".*Attempting to use legacy OpenCV backend.*") warnings.filterwarnings("ignore", message=".*num_frames - 1.*") MODEL_ID = "Wan-AI/Wan2.1-I2V-14B-480P-Diffusers" LORA_REPO_ID = "vrgamedevgirl84/Wan14BT2VFusioniX" LORA_FILENAME = "FusionX_LoRa/Wan2.1_I2V_14B_FusionX_LoRA.safetensors" # Initialize models with proper dtype handling image_encoder = CLIPVisionModel.from_pretrained(MODEL_ID, subfolder="image_encoder", torch_dtype=torch.float16) vae = AutoencoderKLWan.from_pretrained(MODEL_ID, subfolder="vae", torch_dtype=torch.float16) pipe = WanImageToVideoPipeline.from_pretrained( MODEL_ID, vae=vae, image_encoder=image_encoder, torch_dtype=torch.float16 ) pipe.scheduler = UniPCMultistepScheduler.from_config(pipe.scheduler.config, flow_shift=8.0) # Enable memory efficient attention and CPU offloading for large videos pipe.enable_model_cpu_offload() pipe.enable_vae_slicing() pipe.enable_vae_tiling() try: causvid_path = hf_hub_download(repo_id=LORA_REPO_ID, filename=LORA_FILENAME) print("✅ LoRA downloaded to:", causvid_path) pipe.load_lora_weights(causvid_path, adapter_name="causvid_lora") pipe.set_adapters(["causvid_lora"], adapter_weights=[0.75]) pipe.fuse_lora() except Exception as e: import traceback print("❌ Error during LoRA loading:") traceback.print_exc() MOD_VALUE = 32 DEFAULT_H_SLIDER_VALUE = 640 DEFAULT_W_SLIDER_VALUE = 1024 NEW_FORMULA_MAX_AREA = 640.0 * 1024.0 SLIDER_MIN_H, SLIDER_MAX_H = 128, 1024 SLIDER_MIN_W, SLIDER_MAX_W = 128, 1024 MAX_SEED = np.iinfo(np.int32).max FIXED_FPS = 24 MIN_FRAMES_MODEL = 8 # Minimum 8 frames (~0.33s) MAX_FRAMES_MODEL = 240 # Maximum 240 frames (10 seconds at 24fps) 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" 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: gr.Warning("Error attempting to calculate new dimensions") return gr.update(value=DEFAULT_H_SLIDER_VALUE), gr.update(value=DEFAULT_W_SLIDER_VALUE) def get_duration(input_image, prompt, height, width, negative_prompt, duration_seconds, guidance_scale, steps, seed, randomize_seed, progress): # Adjust timeout based on video length and complexity if duration_seconds > 7: return 180 # 3 minutes for very long videos elif duration_seconds > 5: return 120 # 2 minutes for long videos elif duration_seconds > 3: return 90 # 1.5 minutes for medium videos else: return 60 # 1 minute for short videos def export_video_with_ffmpeg(frames, output_path, fps=24): """Export video using imageio if available, otherwise fall back to OpenCV""" try: import imageio # Use imageio for better quality writer = imageio.get_writer(output_path, fps=fps, codec='libx264', pixelformat='yuv420p', quality=8) for frame in frames: writer.append_data(np.array(frame)) writer.close() return True except ImportError: # Fall back to OpenCV export_to_video(frames, output_path, fps=fps) return False @spaces.GPU(duration=get_duration) def generate_video(input_image, 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 input_image is None: raise gr.Error("Please upload an input image.") target_h = max(MOD_VALUE, (int(height) // MOD_VALUE) * MOD_VALUE) target_w = max(MOD_VALUE, (int(width) // MOD_VALUE) * MOD_VALUE) # Calculate frames with proper alignment raw_frames = int(round(duration_seconds * FIXED_FPS)) # Ensure num_frames-1 is divisible by 4 as required by the model num_frames = ((raw_frames - 1) // 4) * 4 + 1 num_frames = np.clip(num_frames, MIN_FRAMES_MODEL, MAX_FRAMES_MODEL) # Additional check for very long videos if num_frames > 120: # For videos longer than 5 seconds, reduce resolution to manage memory max_dim = max(target_h, target_w) if max_dim > 768: scale_factor = 768 / max_dim target_h = max(MOD_VALUE, (int(target_h * scale_factor) // MOD_VALUE) * MOD_VALUE) target_w = max(MOD_VALUE, (int(target_w * scale_factor) // MOD_VALUE) * MOD_VALUE) gr.Info(f"Reduced resolution to {target_w}x{target_h} for long video generation") print(f"Generating {num_frames} frames (requested {raw_frames}) at {target_w}x{target_h}") current_seed = random.randint(0, MAX_SEED) if randomize_seed else int(seed) resized_image = input_image.resize((target_w, target_h), Image.Resampling.LANCZOS) # Clear GPU cache before generation if torch.cuda.is_available(): torch.cuda.empty_cache() try: with torch.inference_mode(): # Generate video with autocast for memory efficiency with torch.autocast("cuda", dtype=torch.float16): output_frames_list = 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), return_dict=True ).frames[0] except torch.cuda.OutOfMemoryError: torch.cuda.empty_cache() raise gr.Error("Out of GPU memory. Try reducing the duration or resolution.") except Exception as e: torch.cuda.empty_cache() raise gr.Error(f"Generation failed: {str(e)}") # Clear cache after generation if torch.cuda.is_available(): torch.cuda.empty_cache() with tempfile.NamedTemporaryFile(suffix=".mp4", delete=False) as tmpfile: video_path = tmpfile.name # Export using imageio if available, otherwise OpenCV used_imageio = export_video_with_ffmpeg(output_frames_list, video_path, fps=FIXED_FPS) # Only try FFmpeg optimization if we have a valid video file if os.path.exists(video_path) and os.path.getsize(video_path) > 0: try: # Check if ffmpeg is available subprocess.run(['ffmpeg', '-version'], capture_output=True, check=True) optimized_path = video_path + "_opt.mp4" cmd = [ 'ffmpeg', '-y', # Overwrite without asking '-i', video_path, # Input file '-c:v', 'libx264', # Codec '-pix_fmt', 'yuv420p', # Pixel format '-profile:v', 'main', # Compatibility profile '-level', '4.0', # Support for higher resolutions '-movflags', '+faststart', # Streaming optimized '-crf', '23', # Quality level '-preset', 'medium', # Balance between speed and compression '-maxrate', '10M', # Max bitrate for large videos '-bufsize', '20M', # Buffer size optimized_path ] result = subprocess.run(cmd, capture_output=True, text=True) if result.returncode == 0 and os.path.exists(optimized_path) and os.path.getsize(optimized_path) > 0: os.unlink(video_path) # Remove original video_path = optimized_path else: print(f"FFmpeg optimization failed: {result.stderr}") except (subprocess.CalledProcessError, FileNotFoundError): print("FFmpeg not available or optimization failed, using original export") return video_path, current_seed # Gradio Interface with gr.Blocks() as demo: gr.Markdown("# Fast 4 steps Wan 2.1 I2V (14B) FusionX-LoRA") gr.Markdown("Generate videos up to 10 seconds long! Longer videos may use reduced resolution for stability.") with gr.Row(): with gr.Column(): input_image_component = gr.Image(type="pil", label="Input Image (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), # 0.3s (8 frames) maximum=round(MAX_FRAMES_MODEL/FIXED_FPS, 1), # 10.0s (240 frames) step=0.1, value=2, # Default 2 seconds label="Duration (seconds)", info=f"Video length: {MIN_FRAMES_MODEL/FIXED_FPS:.1f}-{MAX_FRAMES_MODEL/FIXED_FPS:.1f}s. Longer videos may take more time and use more memory." ) 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=DEFAULT_H_SLIDER_VALUE, label=f"Output Height (multiple of {MOD_VALUE})") width_input = gr.Slider(minimum=SLIDER_MIN_W, maximum=SLIDER_MAX_W, step=MOD_VALUE, value=DEFAULT_W_SLIDER_VALUE, label=f"Output Width (multiple of {MOD_VALUE})") steps_slider = gr.Slider(minimum=1, maximum=30, 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) gr.Markdown("### Tips for best results:") gr.Markdown("- For videos longer than 5 seconds, consider using lower resolutions (512-768px)") gr.Markdown("- Clear, simple prompts often work better than complex descriptions") gr.Markdown("- The model works best with 4-8 inference steps") 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, inputs=ui_inputs, outputs=[video_output, seed_input]) gr.Examples( examples=[ ["peng.png", "a penguin playfully dancing in the snow, Antarctica", 896, 512], ["forg.jpg", "the frog jumps around", 448, 832], ], inputs=[input_image_component, prompt_input, height_input, width_input], outputs=[video_output, seed_input], fn=generate_video, cache_examples="lazy" ) if __name__ == "__main__": demo.queue(max_size=3).launch()