SOLACE applied to Stable Diffusion 3.5 Medium, using the model's own denoising confidence as an intrinsic reward (no external reward model at training time).

Base model: stabilityai/stable-diffusion-3.5-medium
Method: SOLACE intrinsic self-confidence reward (built on Flow-GRPO)
Code: https://github.com/wookiekim/SOLACE
Adapter type: PEFT LoRA (rank 32, applied to the SD3 transformer attention projections)

Usage

python

import torch
from diffusers import StableDiffusion3Pipeline
from peft import PeftModel

model_id = "stabilityai/stable-diffusion-3.5-medium"
lora_ckpt_path = "wookiekim/SD3.5M-SOLACE"
device = "cuda"

# Load base model and apply the SOLACE LoRA adapter
pipe = StableDiffusion3Pipeline.from_pretrained(model_id, torch_dtype=torch.float16)
pipe.transformer = PeftModel.from_pretrained(pipe.transformer, lora_ckpt_path)
pipe.transformer = pipe.transformer.merge_and_unload()
pipe = pipe.to(device)

prompt = "a photo of a cat wearing a small red hat"
image = pipe(
    prompt,
    height=512,
    width=512,
    num_inference_steps=40,
    guidance_scale=4.5,
    negative_prompt="",
).images[0]
image.save("solace.png")

Citation

bibtex

@inproceedings{kim2026solace,
    title={Improving Text-to-Image Generation with Intrinsic Self-Confidence Rewards},
    author={Kim, Wookyoung and others},
    booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    year={2026}
}

Acknowledgments

This work builds upon Flow-GRPO by Jie Liu et al.