roundman/DiffusionRVGE-SD3.5M-MultiReward

DiffusionRVGE: SD3.5-Medium multi-reward LoRA

Rank-32 LoRA for Stable Diffusion 3.5 Medium, trained jointly on five rewards with DiffusionRVGE (Reinforcement Learning via Regression-Based Value Gradient Estimation). This is the joint model of the paper's multi-reward comparison.

Usage

import torch
from diffusers import StableDiffusion3Pipeline

pipe = StableDiffusion3Pipeline.from_pretrained(
    "stabilityai/stable-diffusion-3.5-medium", torch_dtype=torch.bfloat16
).to("cuda")
pipe.load_lora_weights("roundman/DiffusionRVGE-SD3.5M-MultiReward")

image = pipe(
    "a photo of a red bicycle next to a blue umbrella",
    num_inference_steps=40,
    guidance_scale=1.0,  # the reported numbers use no classifier-free guidance
    height=512,
    width=512,
).images[0]
image.save("sample.png")

The LoRA scale alpha / r = 2 is already folded into the lora_B weights, so load it at the default scale of 1.0. Accept the SD3.5 Medium license before downloading the base model.

Training

• Base model: stabilityai/stable-diffusion-3.5-medium, 512 × 512.
• LoRA: rank 32, alpha 64, on to_q, to_k, to_v, to_out.0, add_q_proj, add_k_proj, add_v_proj, to_add_out of every attention block.
• Rewards: PickScore, CLIPScore and HPSv2.1 (combined, on Pick-a-Pic prompts), GenEval, and OCR.

License

The weights are a derivative of Stable Diffusion 3.5 Medium and are released under the Stability AI Community License.