Xixixixihahahaha/RealAlign-SD-1.5

RealAlign — SD-1.5 LoRA Checkpoints

LoRA weights for Stable Diffusion v1.5, aligned with RealAlign from the paper "When Preference Labels Fall Short: Aligning Diffusion Models from Real Data" (ICML 2026).

Resource	Link
📄 Paper	arXiv:2605.19839
🌐 Project page	cwyxx.github.io/RealAlign
🚀 Code	github.com/Cwyxx/RealAlign
🤗 Dataset	RealAlign-Dataset

Summary

RealAlign aligns text-to-image diffusion models using real data as the preference signal: instead of human-annotated preference pairs, it treats a high-quality reference image as the preferred ("win") sample and a perturbed/inpainted version as the non-preferred ("lose") sample. These LoRA adapters are the result of fine-tuning SD-1.5 with RealAlign's two-stage procedure:

1. Stage 1 — Diffusion-DRO (inverse RL / distributionally-robust objective), LoRA + LoRA-init.
2. Stage 2 — Diffusion-DPO with LoRA-init, warm-started from the Stage 1 LoRA.

Checkpoints

Each file is a separate LoRA, trained on preference pairs from a different curation source:

File	Training source
HPDv3.safetensors	HPDv3 (real-photo references)
Civitai-top.safetensors	Civitai top SFW images
Pick-a-pic-v2.safetensors	Pick-a-Pic v2 (top subset)

• Format: diffusers-style UNet LoRA (unet.*.lora_A/lora_B.weight), fp32.
• LoRA rank: 4, applied to the UNet self- and cross-attention projections (to_q, to_k, to_v, to_out).

Note: The Pick-a-Pic v2 LoRA is included here, but the corresponding image dataset is not released on the Hub because the source data may contain NSFW content (see the dataset card).

Usage

import torch
from diffusers import StableDiffusionPipeline

pipe = StableDiffusionPipeline.from_pretrained(
    "stable-diffusion-v1-5/stable-diffusion-v1-5",
    torch_dtype=torch.float16,
).to("cuda")

# Pick the LoRA trained on the source you want
pipe.load_lora_weights(
    "Xixixixihahahaha/RealAlign-SD-1.5",
    weight_name="HPDv3.safetensors",
)

image = pipe("a photo of an astronaut riding a horse on the moon").images[0]
image.save("out.png")

Citation

@article{chen2026preference,
  title={When Preference Labels Fall Short: Aligning Diffusion Models from Real Data},
  author={Chen, Weiyan and Deng, Weijian and Xiao, Yao and Tu, Weijie and Dong, ZiYi and Radwan, Ibrahim and Lin, Liang and Wei, Pengxu},
  journal={arXiv preprint arXiv:2605.19839},
  year={2026}
}