Skip to main content
Back to timeline
arXivSource publication:

One low-res image, any amount of detail: DetailScale-SR turns texture strength into a single dial

Synopsis

Real-world image super-resolution usually returns one fixed level of detail, yet the same degraded image calls for different amounts of detail in different scenes. DetailScale-SR separates general restoration from a user-specified detail preference: it first trains a shared restoration pathway, then freezes it and learns only a module that maps a scalar detail value to attention conditions, so texture expression can be adjusted continuously under the same low-resolution input and semantic constraint. On RealSR and DRealSR, a single checkpoint moves from restrained restoration to detail synthesis as the detail value rises, with PSNR and SSIM falling monotonically and MUSIQ and MANIQA rising, and the most conservative setting reaches the highest PSNR and SSIM.

Source-provided article image: DetailScale-SR: Decoupling Restoration and Continuous Detail Control for Real-World Super-Resolution
arXiv

Interpretation

Real-world super-resolution can now offer a continuous choice of detail on the same low-resolution input, from restrained restoration to richly textured synthesis, while scene content, spatial layout and readable text stay unchanged. Most prior generative super-resolution methods output a single fixed restoration, implicitly assuming the same detail preference across images and users; in existing controllable methods the control variable often affects degradation removal, texture strength, structural fidelity and semantic response to prompts at once, so it is hard to read as a detail preference alone. On the RealSR and DRealSR benchmarks, one checkpoint varies only the detail value and guidance strength; on RealSR, detail value 0 gives PSNR 26.52 and SSIM 0.7725, while value 1 gives PSNR 25.48 and SSIM 0.7300, and averaging per-metric ranks over PSNR, SSIM, LPIPS, DISTS and FID places the method first on both benchmarks.

Detail control comes from a lightweight conditioning module that turns the scalar detail value into extra key-value tokens in attention layers, without touching the already trained restoration model. Training runs in two stages: the first trains a rank-64 LoRA adapter on the Qwen-Image 2.1 image-editing backbone to build shared restoration capability; the second freezes the backbone and adapter and trains only the Scale Template, which maps the scalar through a sinusoidal embedding into block-specific key and value tokens appended to the original attention keys and values. Training targets are built by relative total variation filtering for a low-detail endpoint and interpolating toward the original high-resolution image by detail value, while the low-resolution condition is always generated from the original image independently of that value; in the ablation, joint one-stage training instead of the frozen two-stage scheme drops PSNR from 25.48 to 23.50, SSIM from 0.7300 to 0.6605 and raises FID from 88.40 to 102.77.

An optional scale-aware detail guidance extrapolates along the learned detail direction, producing textures richer than the paired high-resolution reference and extending the controllable range beyond conventional restoration. The guidance follows the classifier-free guidance form, but its two branches differ only in the detail condition while the low-resolution observation, text instruction and latent state stay identical, so the difference isolates the detail scale rather than an extra semantic direction. On RealSR and DRealSR, sweeping the detail value from 0 to 1 lowers PSNR and SSIM monotonically while MUSIQ and MANIQA rise, and FID reaches its minimum at a large detail value; LPIPS is best at an intermediate scale, 0.2215 on RealSR, and the minimum across the DRealSR rows of Table 2 is 0.2715.

Perspective

The demonstrated scope is the RealSR and DRealSR real-world super-resolution benchmarks plus qualitative comparisons on natural texture, thin structures and Chinese text-rich scenes; the model is built on Qwen-Image 2.1, with the first stage trained on 40,000 LSDIR images and 2,000 text-rich samples. Text restoration relies on an optional text prompt, and RGBA layered super-resolution builds on the backbone's four-channel generation capability.

As the detail value grows, PSNR and SSIM fall monotonically, so fidelity and detail expression trade off against each other; no-reference metrics are excluded from the ranking because they measure perceived quality rather than agreement with the reference. Under heavy degradation, text restoration needs a prompt supplying the text content, since some characters cannot be recovered precisely from the low-resolution observation alone.