DFloat11/Wan2.2-I2V-A14B-2-DF11

DFloat11 Compressed Model: Wan-AI/Wan2.2-I2V-A14B

This is a DFloat11 losslessly compressed version of the original Wan-AI/Wan2.2-I2V-A14B model. It reduces model size by 32% compared to the original BFloat16 model, while maintaining bit-identical outputs and supporting efficient GPU inference.

🔥🔥🔥 Thanks to DFloat11 compression, Wan-AI/Wan2.2-I2V-A14B can now generate a 5-second 720P video on a single 24GB GPU, while maintaining full model quality. 🔥🔥🔥

📊 Performance Comparison

Model	Model Size	Peak GPU Memory (5-second 720P generation)	Generation Time (A100 GPU)
Wan-AI/Wan2.2-I2V-A14B (BFloat16)	~56 GB	O.O.M.	-
Wan-AI/Wan2.2-I2V-A14B (DFloat11)	19.47 + 19.44 GB	29.12 GB	42 minutes
Wan-AI/Wan2.2-I2V-A14B (DFloat11 + CPU Offloading)	19.47 + 19.44 GB	20.01 GB	44 minutes

🔍 How It Works

We apply Huffman coding to the exponent bits of BFloat16 model weights, which are highly compressible. We leverage hardware-aware algorithmic designs to enable highly efficient, on-the-fly weight decompression directly on the GPU. Find out more in our research paper.

🔧 How to Use

1. Install or upgrade the DFloat11 pip package (installs the CUDA kernel automatically; requires a CUDA-compatible GPU and PyTorch installed):

   pip install -U dfloat11[cuda12]
   

2. Install the latest diffusers package from source:

   pip install git+https://github.com/huggingface/diffusers
   

3. Save the following code to a Python file i2v.py:

   import time
   import torch
   import numpy as np
   import argparse
   from diffusers import WanImageToVideoPipeline
   from diffusers.utils import export_to_video, load_image
   from dfloat11 import DFloat11Model
   
   parser = argparse.ArgumentParser(description='Image to Video generation using Wan2.2-I2V model')
   parser.add_argument('--cpu_offload', action='store_true', help='Enable CPU offloading')
   parser.add_argument('--image_path', type=str, default="https://huggingface.co/datasets/YiYiXu/testing-images/resolve/main/wan_i2v_input.JPG", help='Path or URL to the input image')
   parser.add_argument('--width', type=int, default=1280, help='Output video width')
   parser.add_argument('--height', type=int, default=720, help='Output video height')
   parser.add_argument('--prompt', type=str, default="Summer beach vacation style, a white cat wearing sunglasses sits on a surfboard. The fluffy-furred feline gazes directly at the camera with a relaxed expression. Blurred beach scenery forms the background featuring crystal-clear waters, distant green hills, and a blue sky dotted with white clouds. The cat assumes a naturally relaxed posture, as if savoring the sea breeze and warm sunlight. A close-up shot highlights the feline's intricate details and the refreshing atmosphere of the seaside.", help='Prompt for video generation')
   parser.add_argument('--negative_prompt', type=str, default="色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走", help='Negative prompt for video generation')
   parser.add_argument('--num_frames', type=int, default=81, help='Number of frames to generate')
   parser.add_argument('--guidance_scale', type=float, default=3.5, help='Guidance scale for generation')
   parser.add_argument('--num_inference_steps', type=int, default=40, help='Number of inference steps')
   parser.add_argument('--seed', type=int, default=42, help='Random seed for generation')
   parser.add_argument('--output', type=str, default='i2v_output.mp4', help='Output video path')
   parser.add_argument('--fps', type=int, default=16, help='FPS of output video')
   
   args = parser.parse_args()
   
   image = load_image(args.image_path)
   
   pipe = WanImageToVideoPipeline.from_pretrained("Wan-AI/Wan2.2-I2V-A14B-Diffusers", torch_dtype=torch.bfloat16)
   
   DFloat11Model.from_pretrained(
       "DFloat11/Wan2.2-I2V-A14B-DF11",
       device="cpu",
       cpu_offload=args.cpu_offload,
       bfloat16_model=pipe.transformer,
   )
   DFloat11Model.from_pretrained(
       "DFloat11/Wan2.2-I2V-A14B-2-DF11",
       device="cpu",
       cpu_offload=args.cpu_offload,
       bfloat16_model=pipe.transformer_2,
   )
   
   pipe.enable_model_cpu_offload()
   
   max_area = args.width * args.height
   aspect_ratio = image.height / image.width
   mod_value = pipe.vae_scale_factor_spatial * pipe.transformer.config.patch_size[1]
   height = round(np.sqrt(max_area * aspect_ratio)) // mod_value * mod_value
   width = round(np.sqrt(max_area / aspect_ratio)) // mod_value * mod_value
   image = image.resize((width, height))
   
   generator = torch.Generator(device="cuda").manual_seed(args.seed)
   
   start_time = time.time()
   output = pipe(
       image=image,
       prompt=args.prompt,
       negative_prompt=args.negative_prompt,
       height=height,
       width=width,
       num_frames=args.num_frames,
       guidance_scale=args.guidance_scale,
       num_inference_steps=args.num_inference_steps,
       generator=generator,
   ).frames[0]
   print(f"Time taken: {time.time() - start_time:.2f} seconds")
   
   export_to_video(output, args.output, fps=args.fps)
   
   max_memory = torch.cuda.max_memory_allocated()
   print(f"Max memory: {max_memory / (1000 ** 3):.2f} GB")
   

4. To run without CPU offloading (40GB VRAM required):

   PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True python i2v.py
   

   To run with CPU offloading (22.5GB VRAM required):

   PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True python i2v.py --cpu_offload
   

   Setting PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True is strongly recommended to prevent out-of-memory errors caused by GPU memory fragmentation.

📄 Learn More

• Paper: 70% Size, 100% Accuracy: Lossless LLM Compression for Efficient GPU Inference via Dynamic-Length Float
• GitHub: https://github.com/LeanModels/DFloat11
• HuggingFace: https://huggingface.co/DFloat11