eDMD / README.md
zttzzt's picture
init commit
d08411b
|
Raw History Blame Contribute Delete
2.68 kB
metadata
license: other
license_name: per-directory
license_link: LICENSE.md
base_model:
  - stabilityai/stable-diffusion-3-medium-diffusers
  - black-forest-labs/FLUX.2-klein-base-4B
  - black-forest-labs/FLUX.2-klein-base-9B
base_model_relation: adapter
library_name: diffusers
pipeline_tag: text-to-image
tags:
  - lora
  - text-to-image
  - distillation
  - few-step

eDMD: Endpoint Distribution Matching for Few-Step Diffusion Distillation

teaser

4-step students of three text-to-image teachers, as LoRA. Paper: arXiv:2611.17194. Code: Westlake-AGI-Lab/eDMD.

Directory Base model HPSv3++ (DrawBench-1000) License
sd3_medium/ stabilityai/stable-diffusion-3-medium-diffusers 11.17 Stability AI Non-Commercial Research Community License
flux2_klein_4b/ black-forest-labs/FLUX.2-klein-base-4B 12.70 Apache 2.0
flux2_klein_9b/ black-forest-labs/FLUX.2-klein-base-9B 12.43 FLUX Non-Commercial License

Each directory holds the LoRA (edmd_<directory>_lora.safetensors) with its teacher's license (LICENSE.md).

Diffusers

import torch
from diffusers import Flux2KleinPipeline

pipe = Flux2KleinPipeline.from_pretrained("black-forest-labs/FLUX.2-klein-base-4B", torch_dtype=torch.bfloat16)
pipe.load_lora_weights("zttzzt/eDMD", subfolder="flux2_klein_4b", weight_name="edmd_flux2_klein_4b_lora.safetensors")
pipe.to("cuda")
image = pipe("a red fox in fresh snow, morning light", num_inference_steps=4, guidance_scale=1.0).images[0]

The students sample in 4 steps without guidance, on their teacher's grid:

Directory Pipeline Call
sd3_medium/ StableDiffusion3Pipeline num_inference_steps=4, sigmas=[1.0, 0.75, 0.5, 0.25], guidance_scale=1.0
flux2_klein_4b/, flux2_klein_9b/ Flux2KleinPipeline num_inference_steps=4, guidance_scale=1.0

With the eDMD code

python scripts/infer.py --weights zttzzt/eDMD/flux2_klein_4b/edmd_flux2_klein_4b_lora.safetensors --prompt "a red fox in fresh snow"

Each file's metadata carries the student's sampler (steps, grid, size), which the eDMD code reads for sampling and for the paper's evaluation (python -m edmd.cli.evaluate generate --checkpoint <file> --eval drawbench1k ...).