# Tencent-Hunyuan/SRPO

Directly Aligning the Full Diffusion Trajectory with Fine-Grained Human Preference

Repository: https://github.com/Tencent-Hunyuan/SRPO
Canonical: https://ross.abutalabs.com/products/srpo
Homepage: https://tencent.github.io/srpo-project-page/
Language: Python
License: NOASSERTION
License Family: other
Last push: 2026-05-11T03:51:29+00:00

## Health v2 (maintenance only)
Score: 54/100 (v2, computed 2026-09-02T17:46:02.011165+00:00)
- activity 81, release rhythm 35, longevity 25
- inputs: {"age_days": 358, "days_push": 114, "days_rel": null, "gap_med": null, "n_releases_24m": 0}
- flags: no_releases, no_license
- formula: round(0.45*activity + 0.35*rhythm + 0.20*longevity); archived -> min(score, 10)

## Adoption (not part of the score)
Stars 1278, forks 42 (observed 2026-08-28T04:04:13.444831+00:00)

## What it is
SRPO is Tencent Hunyuan's research code for fine-tuning diffusion image generation models (e.g., FLUX.1.dev) by aligning the full diffusion trajectory with fine-grained human preference via online reward adjustment. It introduces Direct-Align sampling and Semantic Relative Preference Optimization for faster, more stable training without reward hacking.

## Use cases
- fine-tune flux image generation for better realism
- align diffusion models with human aesthetic preference
- train text-to-image model with reward optimization
- improve photorealism of generated images
- avoid reward hacking in diffusion fine-tuning
- fast preference optimization for image models

## When to choose
- you want to fine-tune FLUX or similar diffusion models with human preference rewards
- you need faster, cheaper reward-based diffusion training than GRPO or ReFL
- you want to avoid reward hacking without KL regularization

## When to avoid
- you need a production image generation service rather than training code
- you work with non-diffusion generative models
- you lack GPU resources for diffusion model training

## Facets
- artifact type: library
- maturity: active
- function: machine-learning, llm-training, image-processing, stable-diffusion
- domain: deep-learning, image-processing, artificial-intelligence
- platform: python
- tags: diffusion-models, fine-tuning, human-preference-alignment, reward-optimization, flux, research-code, gpu, linux

## Member repositories
- Tencent-Hunyuan/SRPO (main) score 54

## Provenance
- Observed fields: from GitHub, fetched 2026-08-28T04:04:13.444831+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 2026-08-30T05:02:44.217745+00:00, confidence not recorded.
  - readme: https://github.com/Tencent-Hunyuan/SRPO (fetched 2026-08-28T04:04:13.444831+00:00, sha 49a6b4363026)
  - homepage: https://tencent.github.io/srpo-project-page/ (fetched 2026-08-29T12:13:29.029399+00:00, sha d48bd85b90d2)
- Data as of 2026-08-30T08:39:29.467469+00:00.