LucidFlux: Caption-Free Universal Image Restoration via a Large-Scale Diffusion Transformer

TL;DR

LucidFlux tackles universal image restoration under unknown real-world degradations without relying on image captions. It leverages a large-scale diffusion transformer (Flux.1) with a lightweight dual-branch conditioner and a timestep- and layer-adaptive modulation to align conditioning with the backbone, preserving global structure while recovering textures. A caption-free SigLIP semantic alignment and a scalable data curation pipeline enable large-scale, structure-rich supervision, achieving state-of-the-art perceptual and semantic fidelity across real and synthetic benchmarks with competitive runtime. The work demonstrates that targeted conditioning placement and data quality, rather than extra parameters or textual prompts, are key to robust caption-free restoration in the wild.

Abstract

Universal image restoration (UIR) aims to recover images degraded by unknown mixtures while preserving semantics -- conditions under which discriminative restorers and UNet-based diffusion priors often oversmooth, hallucinate, or drift. We present LucidFlux, a caption-free UIR framework that adapts a large diffusion transformer (Flux.1) without image captions. LucidFlux introduces a lightweight dual-branch conditioner that injects signals from the degraded input and a lightly restored proxy to respectively anchor geometry and suppress artifacts. Then, a timestep- and layer-adaptive modulation schedule is designed to route these cues across the backbone's hierarchy, in order to yield coarse-to-fine and context-aware updates that protect the global structure while recovering texture. After that, to avoid the latency and instability of text prompts or MLLM captions, we enforce caption-free semantic alignment via SigLIP features extracted from the proxy. A scalable curation pipeline further filters large-scale data for structure-rich supervision. Across synthetic and in-the-wild benchmarks, LucidFlux consistently outperforms strong open-source and commercial baselines, and ablation studies verify the necessity of each component. LucidFlux shows that, for large DiTs, when, where, and what to condition on -- rather than adding parameters or relying on text prompts -- is the governing lever for robust and caption-free universal image restoration in the wild.

Figures (13)

• Figure 1: We present LucidFlux, a universal image restoration framework built on a large-scale diffusion transformer that delivers photorealistic restorations of real-world low-quality (LQ) images, outperforming state-of-the-art (SOTA) diffusion-based models across diverse degradations.
• Figure 2: Overview of the proposed architecture for universal image restoration. Our method integrates dual condition streams (LQ and LRP) with timestep- and layer-adaptive modulation modules, and incorporates SigLIP semantic priors through a connector into a Flux-based DiT backbone to jointly enhance perceptual quality and semantic consistency.
• Figure 3: Comparison of dataset attributes. Our dataset exhibits higher CLIP-IQA scores, lower flatness, and more diverse resolutions than Flickr2K Flickr2K and DIV2K DIV2K.
• Figure 4: Qualitative comparisons on RealLQ250. Baseline methods either leave noticeable artifacts or yield over-smoothed textures, while our approach restores sharper details. See Figure \ref{['fig:RealLQ250_sup1']} to Figure \ref{['fig:LSDIR_sup']} in Appendix for more visual comparisons.
• Figure 5: Impact of captions with and without degradation-related descriptions on restoration results. The second to fourth columns illustrate that inconsistent captions generated by the same MLLM across different runs lead to variations in the restoration outcomes. The fifth and sixth columns show that captions containing explicit degradation descriptions misguide the restoration model and result in inferior quality compared with captions focusing purely on content and style.