README.md

---
base_model:
- stabilityai/stable-diffusion-3.5-medium
language:
- en
license: apache-2.0
pipeline_tag: text-to-image
library_name: diffusers
tags:
- lora
- flow-matching
- distillation
- stable-diffusion
- rtdmd
---

<div align="center">

<img width="70%" height="70%" alt="logo" src="https://cdn-uploads.huggingface.co/production/uploads/64b500fdf460afaefc5c64b3/l1JM1Si5PDCgvJR5SSiqf.png" />

<h2> Reinforcing Few-step Generators via Reward-Tilted Distribution Matching </h2>

<p><b>Reward-Tilted DMD &nbsp;·&nbsp; Ambient-Consistent Distillation &nbsp;·&nbsp; Hybrid Policy Gradient</b></p>

[![Paper](https://img.shields.io/badge/paper-arXiv-A42C25?style=for-the-badge&logo=arxiv&logoColor=white)](https://huggingface.co/papers/2605.26108)
[![Github](https://img.shields.io/badge/Harahan%2FRTDMD-000000?style=for-the-badge&logo=github&logoColor=white)](https://github.com/Harahan/RTDMD)
[![Hugging Face Collection](https://img.shields.io/badge/RTDMD_Collection-fcd022?style=for-the-badge&logo=huggingface&logoColor=000)](https://huggingface.co/collections/Harahan/rtdmd)

[![License: Apache 2.0](https://img.shields.io/badge/License-Apache%202.0-blue.svg)](https://opensource.org/licenses/Apache-2.0)
[![Python](https://img.shields.io/badge/Python-3.10%2B-blue.svg)](https://www.python.org/)

</div>

<div align="center">

[Yushi Huang](https://harahan.github.io/)<sup>1, 2,</sup>\*<sup>†</sup>, [Xiangxin Zhou](https://zhouxiangxin1998.github.io/)<sup>2,</sup>\*, Ruoyu Wang<sup>2, 3,</sup>\*<sup>†</sup>, [Chi Zhang](https://icoz69.github.io/)<sup>3</sup>, [Jun Zhang](https://eejzhang.people.ust.hk/)<sup>1</sup>, [Tianyu Pang](https://p2333.github.io/)<sup>2,</sup>‡

<sup>1</sup>The Hong Kong University of Science and Technology &nbsp;&nbsp;
<sup>2</sup>Tencent Hunyuan &nbsp;&nbsp;
<sup>3</sup>Westlake University

\* Equal contribution &nbsp;·&nbsp; † Work done during internship at Tencent Hunyuan &nbsp;·&nbsp; ‡ Corresponding author

</div>

---

## 📖 Abstract

This repository contains the 4-NFE LoRA checkpoints distilled from **Stable Diffusion 3.5 Medium** using the framework proposed in the paper [Reinforcing Few-step Generators via Reward-Tilted Distribution Matching](https://huggingface.co/papers/2605.26108).

We propose **Reward-Tilted Distribution Matching Distillation (RTDMD)**, a two-stage framework that unifies distribution-matching distillation with reward-guided RL for few-step flow generators. Minimizing the KL divergence to a *reward-tilted teacher distribution* decomposes naturally into a **distribution-matching** term and a **reward-maximization** term — instantiated as **Ambient-Consistent DMD (AC-DMD)** for the cold start and a **hybrid policy gradient** (SubGRPO + final-step reward back-propagation) for the RL stage.

---

## 🍭 Method Overview

<div align="center">
  <img src="https://cdn-uploads.huggingface.co/production/uploads/64b500fdf460afaefc5c64b3/GSQ5Q9bF6SAiUqyFR4ZKs.png" alt="RTDMD method overview" width="70%">
  <br/>
  <em>RTDMD overview. Trajectories: teacher (blue), few-step generator (green), fake score (yellow).</em>
</div>

---

## 📦 Contents

This repository hosts the 4-NFE LoRA checkpoints distilled from **Stable Diffusion 3.5 Medium** with [RTDMD](https://github.com/Harahan/RTDMD).

```
.
├── cold_start/
│   └── generator_ema.pt    # Stage-1 AC-DMD LoRA (4 NFE base)
└── rtdmd/
    └── generator_ema.pt    # Stage-2 RTDMD LoRA (stacked on top of cold_start)
```

Each `generator_ema.pt` is a `state_dict` containing LoRA adapter keys (rank **32**, alpha **64**). The two adapters are designed to be **stacked**: the cold-start LoRA distills the model down to 4 NFE, and the RTDMD LoRA further fine-tunes it with reward-tilted RL.

---

## 🚀 Usage

### Option 1 — RTDMD inference CLI (recommended)

For exact reproduction of the paper numbers, please use the [official RTDMD repository](https://github.com/Harahan/RTDMD).

### Option 2 — Plain diffusers

You can use these LoRAs with the `diffusers` library as follows:

```python
import torch
from diffusers import StableDiffusion3Pipeline
from peft import LoraConfig
from huggingface_hub import hf_hub_download

base = "stabilityai/stable-diffusion-3.5-medium"
pipe = StableDiffusion3Pipeline.from_pretrained(base, torch_dtype=torch.bfloat16).to("cuda")

# Inject LoRA adapters with the rank/alpha used during training
TARGETS = [
    "to_q", "to_k", "to_v", "to_out.0",
    "add_q_proj", "add_k_proj", "add_v_proj", "to_add_out",
]
pipe.transformer.add_adapter(
    LoraConfig(r=32, lora_alpha=64, target_modules=TARGETS, init_lora_weights="gaussian")
)

# Sequentially load cold-start then RTDMD weights into the same adapter
for ckpt in ["cold_start/generator_ema.pt", "rtdmd/generator_ema.pt"]:
    path = hf_hub_download("Harahan/SD35M-RTDMD", ckpt)
    state = torch.load(path, map_location="cpu", weights_only=False)
    pipe.transformer.load_state_dict(state, strict=False)

# 4-step sampling
# Note: RTDMD is trained on the CPS scheduler with η = 0.9. 
# Default Flow-Matching Euler will still produce reasonable samples.
pipe(prompt="a cute cat sitting on a windowsill",
     num_inference_steps=4, guidance_scale=1.0).images[0].save("out.png")
```

---

## 📄 Citation

```bibtex
@misc{huang2026reinforcingfewstepgeneratorsrewardtilted,
      title={Reinforcing Few-step Generators via Reward-Tilted Distribution Matching}, 
      author={Yushi Huang and Xiangxin Zhou and Ruoyu Wang and Chi Zhang and Jun Zhang and Tianyu Pang},
      year={2026},
      eprint={2605.26108},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2605.26108}, 
}
```

---

## ⚖️ License

Apache 2.0. The base model [`stabilityai/stable-diffusion-3.5-medium`](https://huggingface.co/stabilityai/stable-diffusion-3.5-medium) is governed by its own license; please review and comply with it separately.