--- license: apache-2.0 base_model: stabilityai/stable-diffusion-3.5-medium library_name: diffusers pipeline_tag: text-to-image tags: - text-to-image - diffusion - flow-matching - meanflow - reinforcement-learning - lora ---

Bringing Forward-Process RL to Average-Velocity Generators

[Yushi Huang](https://harahan.github.io/)^1,2\* · [Xiangxin Zhou](https://zhouxiangxin1998.github.io/)¹\*^✉ · [Jun Zhang](https://eejzhang.people.ust.hk/)² · [Liefeng Bo](https://research.cs.washington.edu/istc/lfb/)¹ · [Tianyu Pang](https://p2333.github.io/)^1,✉ ¹Tencent Hunyuan ²The Hong Kong University of Science and Technology \* Equal contribution ^✉Corresponding authors [![arXiv](https://img.shields.io/badge/arXiv-2607.15273-A42C25?style=for-the-badge&logo=arxiv&logoColor=white)](https://arxiv.org/abs/2607.15273) [![Project Page](https://img.shields.io/badge/Project-Page-0F9D8A?style=for-the-badge&logo=googlechrome&logoColor=white)](https://harahan.github.io/meanflownft-project-page/) [![Code](https://img.shields.io/badge/GitHub-MeanFlowNFT-181717?style=for-the-badge&logo=github)](https://github.com/Harahan/MeanFlowNFT) [![Demo](https://img.shields.io/badge/🤗-Interactive_Demo-FF4B4B?style=for-the-badge)](https://huggingface.co/spaces/Harahan/meanflownft-fewstep-generation)

MeanFlowNFT optimizes reward in induced instantaneous-velocity space while retaining the average-velocity parameterization and native any-step MeanFlow sampler. Training uses only forward-noised clean samples, without reverse-trajectory likelihood estimation.

## 🍭 Overview

MeanFlowNFT method overview

_{Optimize in instantaneous-velocity space while preserving MeanFlow sampling.}

MeanFlowNFT update algorithm

_{One practical MeanFlowNFT update.}

## 📦 Contents This repository contains the three ordered adapter components required to reconstruct the final policy: ```text . ├── anyflow_pretrain/ │ └── generator_ema.pt ├── anyflow_onpolicy/ │ └── generator_ema.pt └── meanflownft/ └── generator_ema.pt ``` The loading contract is strict: 1. Merge `anyflow_pretrain/generator_ema.pt` into the base transformer. 2. Merge `anyflow_onpolicy/generator_ema.pt` on top of that result. 3. Keep `meanflownft/generator_ema.pt` as the active final adapter. 4. Load the complete `delta_embedder` state from the final checkpoint. The adapter checkpoints use LoRA rank **32**, alpha **64**, and attention projections matching the released code. They also contain the trainable flow-map `delta_embedder`, so they are not conventional LoRA-only files. `SHA256SUMS` records the checksum of every checkpoint. ## 📊 Results

MeanFlowNFT is best on **6 of 8** image metrics among the evaluated few-step models, while 4-step Wan2.1 reaches **84.33 VBench**, surpassing 50-step LongCat-Video RL. See the [project page](https://harahan.github.io/meanflownft-project-page/) for full comparisons, scaling curves, and qualitative results. ## 🚀 Usage Use the [MeanFlowNFT codebase](https://github.com/Harahan/MeanFlowNFT), whose inference loader preserves the ordered adapter composition and full `delta_embedder` state. ```bash git clone https://github.com/Harahan/MeanFlowNFT.git cd MeanFlowNFT pip install -r requirements.txt pip install -e . export HF_REPO_ID=Harahan/MeanFlowNFT hf download "${HF_REPO_ID}" --local-dir ./checkpoints/meanflownft hf auth login hf download stabilityai/stable-diffusion-3.5-medium \ --local-dir ./models/stable-diffusion-3.5-medium ``` Run the final 4-step policy: ```bash python inference.py configs/inference/sd35m_meanflow_nft.yaml \ --override \ pretrained_path=./models/stable-diffusion-3.5-medium \ num_stages=3 \ stage1_lora_path=./checkpoints/meanflownft/anyflow_pretrain/generator_ema.pt \ stage2_lora_path=./checkpoints/meanflownft/anyflow_onpolicy/generator_ema.pt \ stage3_lora_path=./checkpoints/meanflownft/meanflownft/generator_ema.pt \ num_steps=4 \ eval_reward=false \ --prompt "a cinematic photo of a red panda" ``` The first two adapters are merged in order. The final MeanFlowNFT adapter stays active and supplies `delta_embedder`, matching training-time evaluation. The released inference config generates 1024×1024 images with CFG-free flow-map sampling. > [!IMPORTANT] > Do not load only the final checkpoint and do not use a generic LoRA loader > that discards non-LoRA tensors. Either mistake drops part of the trained > policy. > [!WARNING] > These `.pt` files use PyTorch serialization. Load checkpoints only from a > trusted source. ## 📄 Citation ```bibtex @misc{huang2026meanflownftbringingforwardprocessrl, title = {MeanFlowNFT: Bringing Forward-Process RL to Average-Velocity Generators}, author = {Huang, Yushi and Zhou, Xiangxin and Zhang, Jun and Bo, Liefeng and Pang, Tianyu}, year={2026}, eprint={2607.15273}, archivePrefix={arXiv}, primaryClass={cs.CV}, url={https://arxiv.org/abs/2607.15273}, } ``` ## 🙌 Acknowledgements This work builds on [MeanFlow](https://github.com/Gsunshine/meanflow), [AnyFlow](https://github.com/NVlabs/AnyFlow), [DiffusionNFT](https://github.com/NVlabs/DiffusionNFT), [Diffusers](https://github.com/huggingface/diffusers), [Transformers](https://github.com/huggingface/transformers), and [PEFT](https://github.com/huggingface/peft). ## ⚖️ License The repository metadata and accompanying code are released under Apache 2.0. The base model and all derivative weights remain subject to the base model's license and usage terms; review them before use or redistribution.