Directly Aligning the Full Diffusion Trajectory with Fine-Grained Human Preference
Xiangwei Shen
1,2*,
Zhimin Li1*,
Zhantao Yang1,
Shiyi Zhang3,
Yingfang Zhang
1,
Donghao Li
1,
Chunyu Wang1,
Qinglin Lu1,
Yansong Tang3,✝
1Hunyuan, Tencent
2School of Science and Engineering, The Chinese University of Hong Kong, Shenzhen
3Shenzhen International Graduate School, Tsinghua University
*Equal contribution
✝Corresponding author
Recent studies have demonstrated the effectiveness of directly aligning diffusion models with human preferences using differentiable reward. However, they exhibit two primary challenges: (1) they rely on multistep denoising with gradient computation for reward scoring, which is computationally expensive, thus restricting optimization to only a few diffusion steps; (2) they often need continuous offline adaptation of reward models in order to achieve desired aesthetic quality, such as photorealism or precise lighting effects. To address the limitation of multistep denoising, we propose Direct-Align, a method that predefines a noise prior to effectively recover original images from any time steps via interpolation, leveraging the equation that diffusion states are interpolations between noise and target images, which effectively avoids over-optimization in late timesteps. Furthermore, we introduce Semantic Relative Preference Optimization (SRPO), in which rewards are formulated as text-conditioned signals. This approach enables online adjustment of rewards in response to positive and negative prompt augmentation, thereby reducing the reliance on offline reward fine-tuning. By fine-tuning the FLUX.1.dev model with optimized denoising and online reward adjustment, we improve its human-evaluated realism and aesthetic quality by over 3x.
The
diffusion_pytorch_model.safetensors is online version of SRPO based on
FLUX.1 Dev, trained on HPD dataset with
HPSv21from diffusers import FluxPipeline
2prompt='The Death of Ophelia by John Everett Millais, Pre-Raphaelite painting, Ophelia floating in a river surrounded by flowers, detailed natural elements, melancholic and tragic atmosphere'
3pipe = FluxPipeline.from_pretrained('./data/flux',
4 torch_dtype=torch.bfloat16,
5 use_safetensors=True
6 ).to("cuda")
7state_dict = load_file("./srpo/diffusion_pytorch_model.safetensors")
8pipe.transformer.load_state_dict(state_dict)
9image = pipe(
10 prompt,
11 guidance_scale=3.5,
12 height=1024,
13 width=1024,
14 num_inference_steps=infer_step,
15 max_sequence_length=512,
16 generator=generator
17).images[0]
SRPO is licensed under the License Terms of SRPO. See ./License.txt for more details.
1@misc{shen2025directlyaligningdiffusiontrajectory,
2 title={Directly Aligning the Full Diffusion Trajectory with Fine-Grained Human Preference},
3 author={Xiangwei Shen and Zhimin Li and Zhantao Yang and Shiyi Zhang and Yingfang Zhang and Donghao Li and Chunyu Wang and Qinglin Lu and Yansong Tang},
4 year={2025},
5 eprint={2509.06942},
6 archivePrefix={arXiv},
7 primaryClass={cs.AI},\
8 url={https://arxiv.org/abs/2509.06942},
9}
