Paper 2603.24800
Calibri: Enhancing Diffusion Transformers via Parameter-Efficient Calibration
- Published
- Mar 2026
- Research lab
- Independent
- Citations
- 0
- GitHub
- 59 stars
01 In brief
Summary
The paper introduces Calibri, a parameter-efficient method to enhance Diffusion Transformers (DiTs) by calibrating block outputs with learned scaling parameters.
The authors show that selectively disabling or re-weighting DiT blocks can improve generation quality, leading to a black-box optimization problem solved via CMA-ES, modifying only ~10^2 parameters.
Calibri is evaluated on FLUX, SD-3.5M, and Qwen-Image, consistently improving metrics like HPSv3, ImageReward, and Q-Align while reducing inference steps (e.g., FLUX from 30 to 15 NFE).
Calibri Ensemble combines multiple calibrated models, further boosting performance and reducing optimal steps to 10-15.
A user study confirms perceptual gains.
Calibri also complements alignment methods like Flow-GRPO, achieving comparable or better results with far fewer parameters (216 vs 18.78M).
Calibration cost ranges from 32 to 356 H100 GPU-hours, a one-time offline expense.
The method preserves generation diversity, unlike Flow-GRPO, and CMA-ES is shown to be more efficient than gradient-based optimization.
02 From the paper
Abstract
In this paper, we uncover the hidden potential of Diffusion Transformers (DiTs) to significantly enhance generative tasks. Through an in-depth analysis of the denoising process, we demonstrate that introducing a single learned scaling parameter can significantly improve the performance of DiT blocks. Building on this insight, we propose Calibri, a parameter-efficient approach that optimally calibrates DiT components to elevate generative quality. Calibri frames DiT calibration as a black-box reward optimization problem, which is efficiently solved using an evolutionary algorithm and modifies just ~100 parameters. Experimental results reveal that despite its lightweight design, Calibri consistently improves performance across various text-to-image models. Notably, Calibri also reduces the inference steps required for image generation, all while maintaining high-quality outputs.