ChenHe727/EdgeDiffusion_distilled

EdgeDiffusion - Distilled

A pruned and distilled Stable Diffusion 1.5 UNet (647.2M params, ~25% smaller than original 858.5M).

Pipeline

1. Iterative Pruning: 4 rounds of ~7% Taylor-importance pruning (858.5M → 647.2M)
2. Knowledge Distillation: 15K steps with Realistic Vision v5.1 as teacher (feature + noise MSE loss)

Files

File	Description
pruned_unet.safetensors	Pruned + distilled UNet weights
pruned_unet.config.json	Model config (contains model_config for rebuilding)
pruned_rebuild.py	Script to rebuild the pruned UNet architecture

Usage

# 1. Download all 3 files to the same directory, then:
from pruned_rebuild import create_unet_from_safetensors
from diffusers import StableDiffusionPipeline
import torch

# 2. Rebuild the pruned UNet
unet = create_unet_from_safetensors(
    "pruned_unet.safetensors",
    "pruned_unet.config.json"
)

# 3. Load into a standard SD 1.5 pipeline
pipe = StableDiffusionPipeline.from_pretrained(
    "runwayml/stable-diffusion-v1-5",
    unet=unet,
    torch_dtype=torch.float16,
    safety_checker=None,
)
pipe = pipe.to("cuda")

# 4. Generate
image = pipe("a beautiful sunset over mountains", num_inference_steps=30).images[0]
image.save("output.png")

Requirements

pip install diffusers transformers safetensors torch accelerate