Hunyuan Flux SRPO
Hunyuan Flux SRPO is FLUX.1 dev (12B) fine-tuned by Tencent Hunyuan with SRPO, a human-preference training method aimed at realism and aesthetic quality without oversaturated color.
Overview
Hunyuan Flux SRPO is FLUX.1 dev, the 12B-parameter flow transformer from Black Forest Labs, with its transformer weights fine-tuned by Tencent Hunyuan. Tencent published the checkpoint, paper and inference code on September 8, 2025, and the training code four days later, under the name SRPO (Semantic Relative Preference Optimization).
The method has two parts. Direct-Align uses a predefined noise prior so a clean image can be recovered from any denoising step, which lets training cover the whole diffusion trajectory and avoids over-optimizing the late steps. SRPO turns the reward into a text-conditioned signal that can be adjusted with positive and negative prompt words. This checkpoint was trained on the HPD dataset with HPSv2 as the reward. Tencent reports that the method improves the human-evaluated realism and aesthetic quality of FLUX.1 dev by over 3x, without reward hacking such as drifting toward oversaturated color.
Only the transformer changes. The text encoders and the VAE come from FLUX.1 dev, so it reads prompts the same way FLUX.1 dev does.
What you can build
- Side-by-side realism studies against FLUX.1 dev on the same prompt and seed, looking at skin, fabric and light. The samples include a close-up of a woman in traditional Chinese dress and a full-length shot of a man on a misty street.
- Evaluation sets for human-preference and aesthetic scoring research, across photographic and illustrated prompts such as the blue night scene of people at a bus stop.
- Studies of how preference tuning treats painterly styles. The Ophelia sample uses the Pre-Raphaelite prompt from Tencent's own quick-start example.
- Tests of title lettering and creature prompts, such as the Underealm cover and the furry, amber-eyed creature in the samples.
Inputs
- Prompt: a description of the image in plain sentences. The T5 text encoder reads up to 512 tokens, while the CLIP encoder keeps only the first 77, so put the subject and style at the start of a long prompt.
- Width and height: up to 2048 pixels each, 1024 x 1024 by default.
- Inference steps (default 20) and guidance scale (1 to 20, default 3.5).
- Samples: 1 to 8 images per run.
- Seed: fix it to reproduce a result or to compare prompt edits.
Outputs
PNG images, one file per sample, at the width and height you set rounded down to a multiple of 16 pixels.
Recommended settings
- Tencent's reference settings are guidance 3.5, 1024 x 1024 and 50 steps. The default of 20 steps is enough for trying out prompts; move toward 50 for final images where texture and small detail matter.
- Keep guidance close to 3.5.
- Wide frames work well. Two of the samples here were made at 1024 x 576.
Limitations
- The fine-tune targets human-preference scores for realism and aesthetics. Prompt understanding still comes from the unchanged FLUX.1 dev text encoders.
- Large title lettering comes out legibly, but small secondary text can be garbled, as in the line under the title on the Underealm sample.
Safety & compliance
Tencent's license terms for the SRPO weights allow academic use only and rule out commercial or production use. As a fine-tune of FLUX.1 [dev], the weights also fall under Black Forest Labs' FLUX.1 [dev] Non-Commercial License, which limits the model and its derivatives to non-commercial use, forbids using outputs to train, fine-tune or distill a model that competes with FLUX.1 [dev], and prohibits unlawful or infringing content.
Example prompts
Great starting points for Hunyuan-Flux-SRPO.
API quick start
Run Hunyuan-Flux-SRPO with a single API call.
{
"prompt": "A close up portrait of a Chinese princess…",
"inputImage": "https://your-cdn.com/input.png",
"steps": 20,
"scale": 3.5
}curl -X POST "https://api.wiro.ai/v1/Run/tencent/hunyuan-flux-srpo" \
-H "Content-Type: application/json" \
-H "x-api-key: YOUR_WIRO_API_KEY" \
--data-binary @- <<'JSON'
{
"prompt": "A close up portrait of a Chinese princess…",
"inputImage": "https://your-cdn.com/input.png",
"steps": 20,
"scale": 3.5
}
JSONOther models you might like
Explore similar image generation models.



