TL;DR
DetailScale-SR introduces a "detail slider" for real-world super-resolution, allowing users to continuously adjust the richness of reconstructed textures using a single scalar value. By decoupling general restoration from detail preference, the framework enables everything from conservative, artifact-free smoothing to "beyond-reference" enhancement that generates textures even more vivid than the original high-resolution source.
What Problem Does This Paper Address?
In real-world super-resolution (SR), there is rarely a "one-size-fits-all" solution. A portrait might require smooth, restrained restoration to avoid skin artifacts, while a landscape or a shot of animal fur benefits from aggressive, high-frequency detail synthesis.
Existing generative SR models typically produce a single "canonical" output. While some methods offer controllability, they often entangle detail levels with other variables like degradation strength or text prompts. This makes it difficult for users to precisely dial in the "richness" of a texture without accidentally changing the scene's identity or the semantic content. DetailScale-SR aims to provide a continuous, predictable control dimension specifically for detail.
Methods
The authors propose a two-stage controllable restoration framework built upon the Qwen-Image 2.1 image-editing backbone:
- Stage I: Shared Restoration Adaptation: The model learns to handle real-world degradations (blur, noise, compression) by training a LoRA (Low-Rank Adaptation) module. This establishes a baseline capability for recovering clean content.
- Stage II: Continuous Detail Conditioning: The restoration adapter is frozen, and a new Scale Template module is trained. This module maps a user-defined scalar $s \in [0, 1]$ into layer-wise attention tokens.
To train this "detail slider," the authors create a continuous supervision path. They use Relative Total Variation (RTV) filtering to create a "low-detail" version of the ground truth. The model is then trained to reconstruct targets that are linear interpolations between this smoothed version and the original high-resolution image, all while the input low-resolution image remains identical.
The framework also includes an optional Scale-Aware Detail Guidance (Detail-CFG). Similar to classifier-free guidance, this extrapolates the learned detail direction beyond the training range to achieve hyper-realistic enhancement.
Datasets
- Training: The model was trained on 40,000 images from the LSDIR dataset, supplemented by 2,000 text-rich samples to improve performance on signs and posters.
- Evaluation: Performance was benchmarked on the RealSR and DRealSR datasets, which are standard for evaluating real-world super-resolution.
Key Insights and Ideas
- Decoupling is Key: By freezing the restoration adapter in Stage II, the authors ensure that the detail slider only affects texture richness, not the model's fundamental ability to remove blur or noise.
- RTV as a Detail Anchor: Using Relative Total Variation to define the "zero-detail" endpoint is a clever way to provide a mathematically grounded reference for "smoothness" that still preserves the underlying structure of the image.
- Text as a First-Class Citizen: Unlike many SR models that treat text as just another texture (often leading to "gibberish" characters), DetailScale-SR leverages the strong text-rendering priors of its pretrained backbone. This makes it exceptionally good at restoring legible Chinese and English characters in complex scenes.
Key Results
- Quantitative Superiority: DetailScale-SR achieved the #1 average rank across multiple metrics (PSNR, SSIM, LPIPS, DISTS, and FID) on both the RealSR and DRealSR benchmarks.
- Fidelity-Realism Trade-off: The model allows users to navigate the Pareto front of SR. Lower $s$ values yield higher PSNR/SSIM (fidelity), while higher $s$ values (and Detail-CFG) yield better perceptual scores like FID and MUSIQ (realism).
- Text Restoration: Qualitative results show that the model recovers much more legible and structurally accurate text than competitors like ODTSR or SeeSR, especially in "text-heavy" environments like storefronts.
Limitations and Future Work
While the paper demonstrates impressive control, a few areas remain for exploration:
- Inference Overhead: As a diffusion/flow-based model, inference is likely slower than simple GAN-based or CNN-based SR methods, though the use of flow-matching helps mitigate this.
- Hallucination at Extremes: While Detail-CFG provides impressive enhancement, very high guidance scales can eventually lead to over-sharpening or the introduction of synthetic-looking textures.
- Future Directions: The authors suggest that this "decoupled control" philosophy could be extended to other image properties beyond just detail, or adapted for video super-resolution where temporal consistency of the detail slider would be a new challenge.
Summary updated