# NVlabs/DiffusionNFT

[ICLR 2026 Oral] DiffusionNFT: Online Diffusion Reinforcement with Forward Process

Repository: https://github.com/NVlabs/DiffusionNFT
Canonical: https://ross.abutalabs.com/products/diffusionnft
Language: Python
License: Apache-2.0
License Family: permissive
Last push: 2026-02-10T11:18:15+00:00

## Health v2 (maintenance only)
Score: 47/100 (v2, computed 2026-09-02T17:46:02.011165+00:00)
- activity 66, release rhythm 35, longevity 24
- inputs: {"age_days": 348, "days_push": 204, "days_rel": null, "gap_med": null, "n_releases_24m": 0}
- flags: no_releases
- formula: round(0.45*activity + 0.35*rhythm + 0.20*longevity); archived -> min(score, 10)

## Adoption (not part of the score)
Stars 1034, forks 45 (observed 2026-08-28T04:03:18.668890+00:00)

## What it is
DiffusionNFT is a research library implementing an online reinforcement learning paradigm for diffusion models that optimizes policy directly on the forward diffusion process. Built on the Flow-GRPO codebase, it supports multiple reward models (GenEval, OCR, PickScore, CLIP, HPSv2, Aesthetic) for training diffusion models like SD3.5M.

## Use cases
- fine-tune diffusion models with reinforcement learning
- train text-to-image models with reward-based optimization
- apply online RL to diffusion policy without sampling trajectories
- optimize image generation with reward models like PickScore or HPSv2
- research forward-process policy optimization for diffusion models
- integrate RL training into existing flow-matching codebases

## When to choose
- you need solver-agnostic RL fine-tuning of diffusion models compatible with any black-box sampler
- you want memory-efficient training that only requires clean images rather than full sampling trajectories
- you want to reward-optimize diffusion models using standard flow-matching objectives
- you are reproducing ICLR 2026 research on diffusion reinforcement learning

## When to avoid
- you need general-purpose diffusion model inference or image generation without RL training
- you are training non-diffusion models such as LLMs or standard supervised classifiers
- you need a production-ready, stable training framework with extensive documentation and support
- you lack GPU resources, as diffusion RL training is compute-intensive

## Facets
- artifact type: library
- maturity: active
- function: machine-learning, llm-training, image-processing
- domain: machine-learning, deep-learning, image-processing, artificial-intelligence
- platform: python
- tags: diffusion-models, reinforcement-learning, flow-matching, policy-optimization, reward-models, research-code, iclr-2026, gpu, linux

## Member repositories
- NVlabs/DiffusionNFT (main) score 47

## Provenance
- Observed fields: from GitHub, fetched 2026-08-28T04:03:18.668890+00:00.
- Health v2: computed from the inputs above; adoption is never an input.
- Inferred fields (summary, facets, guidance): AI-extracted, prompt v1, taxonomy v1, on 2026-08-30T07:06:30.826835+00:00, confidence not recorded.
  - readme: https://github.com/NVlabs/DiffusionNFT (fetched 2026-08-28T04:03:18.668890+00:00, sha ff8360d4c8d1)
- Data as of 2026-08-30T08:39:29.467469+00:00.