SD35M-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>

</div> <div align="center">

Yushi Huang^{1, 2,}*^†, Xiangxin Zhou^1,*, Ruoyu Wang^{1, 3,}*^†, Chi Zhang³, Jun Zhang², Tianyu Pang^1,‡

¹Tencent Hunyuan    ²The Hong Kong University of Science and Technology    ³Westlake University

* Equal contribution  ·  † Work done during internship at Tencent Hunyuan  ·  ‡ Corresponding author

</div>

---

📖 Abstract

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. With 4 NFE RTDMD reaches new SOTA on SD3-M / SD3.5-M / FLUX.2 4B; the distilled FLUX.2 4B even beats the full FLUX.2 9B teacher (50 NFE) on most rewards.

<table align="center"> <tr> <td align="center" width="50%"> <img src="https://cdn-uploads.huggingface.co/production/uploads/64b500fdf460afaefc5c64b3/MLr2YHfmAvKYQsfA50uOf.png" alt="RTDMD teaser" width="100%"> <br/> <em>4-step samples from RTDMD-distilled FLUX.2 4B (no classifier-free guidance).</em> </td> <td align="center" width="50%"> <img src="https://cdn-uploads.huggingface.co/production/uploads/64b500fdf460afaefc5c64b3/UJnH2QqNCw4aJDFgwtfjP.png" alt="RTDMD comparison" width="100%"> <br/> <em>Qualitative comparison for few-step diffusion models (4 NFE).</em> </td> </tr> </table>

---

🍭 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. <b>Det.</b> = deterministic final step, <b>Stoc.</b> = stochastic intermediate steps. Trajectories: teacher (blue), few-step generator (green), fake score (yellow).</em> </div>

For the generator $G_\theta$, the reward-tilted KL objective decomposes as

$$ \nabla_\theta D_{\text{KL}}(p_\theta | \tilde{p}\psi) = \underbrace{\nabla\theta D_{\text{KL}}(p_\theta | p_\psi)}{\text{distribution matching}} - \beta\underbrace{\nabla\theta \mathbb{E}_{\hat{\mathbf{x}}0 \sim p\theta}[r(\hat{\mathbf{x}}0)]}{\text{reward maximization}}. $$

The two terms map directly to the two trainers exposed by the CLI:

---

📦 Contents

This repository hosts the 4-NFE LoRA checkpoints distilled from Stable Diffusion 3.5 Medium with 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 torch.save-d state_dict containing only LoRA adapter keys (lora_A / lora_B, rank 32, alpha 64). The two adapters are designed to be stacked: the cold-start LoRA distills SD3.5-M down to 4 NFE, and the RTDMD LoRA further fine-tunes that distilled model with reward-tilted RL.

---

🚀 Usage

Option 1 — RTDMD inference CLI (recommended)

The simplest path is to clone the RTDMD repo and let it stack both LoRAs and run the CPS sampler for you:

git clone https://github.com/Harahan/RTDMD.git && cd RTDMD
pip install -r requirements.txt && pip install -e .

# Download this repo
huggingface-cli download Harahan/SD35M-RTDMD --local-dir ./ckpts/sd35m

# Run 4-NFE inference (single GPU)
python inference.py configs/inference/sd35m.yaml \
    --override lora_paths='["./ckpts/sd35m/cold_start/generator_ema.pt","./ckpts/sd35m/rtdmd/generator_ema.pt"]' \
    --override eval_reward=false \
    --prompt "a cute cat sitting on a windowsill"

Option 2 — Plain diffusers

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 CPS sampling
pipe(prompt="a cute cat sitting on a windowsill",
     num_inference_steps=4, guidance_scale=1.0).images[0].save("out.png")

Note: RTDMD is trained on the CPS (Coefficients-Preserving Sampling) scheduler with η = 0.9. Using the default Flow-Matching Euler scheduler will still produce reasonable samples at 4 NFE, but the RTDMD inference CLI is the only entry point that reproduces the paper numbers exactly.

---

📄 Citation

@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 — same as the upstream RTDMD repo. The base model stabilityai/stable-diffusion-3.5-medium is governed by its own license; please review and comply with it separately.