From Sketch to Fresco: Efficient Diffusion Transformer with Progressive Resolution

TL;DR

Fresco addresses the computational bottleneck of diffusion transformers by introducing a training-free progressive-resolution framework that unifies cross-stage noise and uses variance-guided upsampling to selectively refine stable tokens. A token-encoded unified noise field preserves consistent stochastic evolution across resolutions, enabling coherent coarse-to-fine generation without re-noising resets. The approach yields substantial acceleration on both image and video tasks (up to 10× on FLUX and 5× on HunyuanVideo, up to 22× with distillation) while maintaining high fidelity and compatibility with distillation, quantization, and feature caching. Overall, Fresco provides a simple, scalable paradigm for efficient high-resolution diffusion generation with broad practical impact.

Abstract

Diffusion Transformers achieve impressive generative quality but remain computationally expensive due to iterative sampling. Recently, dynamic resolution sampling has emerged as a promising acceleration technique by reducing the resolution of early sampling steps. However, existing methods rely on heuristic re-noising at every resolution transition, injecting noise that breaks cross-stage consistency and forces the model to relearn global structure. In addition, these methods indiscriminately upsample the entire latent space at once without checking which regions have actually converged, causing accumulated errors, and visible artifacts. Therefore, we propose \textbf{Fresco}, a dynamic resolution framework that unifies re-noise and global structure across stages with progressive upsampling, preserving both the efficiency of low-resolution drafting and the fidelity of high-resolution refinement, with all stages aligned toward the same final target. Fresco achieves near-lossless acceleration across diverse domains and models, including 10$\times$ speedup on FLUX, and 5$\times$ on HunyuanVideo, while remaining orthogonal to distillation, quantization and feature caching, reaching 22$\times$ speedup when combined with distilled models. Our code is in supplementary material and will be released on Github.

Figures (6)

• Figure 1: Overview of re-noising strategies. (a) Standard Sampling: one initial noise with no re-noise during process. (b) Traditional Dynamic Resolution: inject stage-specific noise independently at every resolution change, disrupting semantic and reset denoising trajectory, causing aliasing and artifacts. (c) Unified re-noise (ours): all stages query the same noise field, ensuring stable refinement and clean results.
• Figure 2: Overview of the Fresco framework. Fresco starts sampling at a reduced resolution while assigning each token a fixed coordinate-bound noise vector from a unified noise field. During generation, Fresco tracks each token’s temporal variance: tokens with small variance, indicating stable semantics, are upsampled for high-resolution refinement, whereas unstable tokens remain at low resolution for further denoising. This unified-noise, variance-guided process enables smooth and efficient coarse-to-fine generation.
• Figure 3: Visualization of the image generated by different methods on FLUX.1-dev. Fresco delivers the most realistic and semantically faithful results while achieving the fastest speed (4.72$\times$), outperforming all dynamic resolution and feature caching baselines.
• Figure 4: Visualization of different acceleration methods on HunyuanVideo. Fresco achieves the best visual quality, semantic accuracy, and highest speedup ratio 4.92$\times$ among all methods, demonstrating strong generalization ability on video generation models.
• Figure 5: Compatibility with other acceleration methods. Fresco achieves 22$\times$ speedup when combined with step distillation and 9$\times$ training-free acceleration when paired with feature caching, while preserving high visual fidelity.