Remote-sensing super-resolution stops feeding the whole low-res image to a diffusion model: a few structural slots beat DiT4SR by 4.54 dB PSNR on AID

Remote-sensing super-resolution no longer hands the whole low-resolution image to a diffusion Transformer as its condition; it first compresses the observation into a small set of structural slots, so foreground objects and regular textures are not diluted by large background regions during denoising. Earlier DiT-based methods inject the low-resolution observation as a dense spatial condition, which entangles foreground objects with dominant backgrounds and weakens small targets and structural boundaries during denoising. Across UCMerced, AID, and RSSCN7 at two upscaling settings, PSNR is best on all three datasets; against DiT4SR the gain is 4.54 dB on AID, 2.03 dB on UCMerced, and 1.49 dB on RSSCN7.

Slots are extracted at the original and a resized scale, where the coarse scale carries stable layouts and the fine scale keeps boundaries, textures, and small objects, and the two are fused in slot space before reaching the frozen diffusion backbone. A single-scale dense condition retains too much background, and fixed averaging of scales gives only a small improvement, while learned cross-scale interaction opens a clear gap. On the UCMerced validation set, the single-scale baseline reaches 24.1520 dB PSNR, direct averaging 24.1620 dB, inter-scale fusion alone 24.2440 dB, and full multi-scale fusion 24.2510 dB with SSIM 0.6386 and LPIPS 0.2210.

Severe degradation can erase image-specific cues for some objects, so a prototype memory clustered offline from training slots adds recurring structures back as residual context rather than replacing the current image's information. Image-specific slots do not explicitly summarize structures shared across samples, while direct quantization or feature replacement can discard image-specific details. On the UCMerced validation set, adding prototype memory conditioning moves PSNR from 24.4340 dB to 24.4419 dB and LPIPS from 0.2223 to 0.2190; the memory is built offline, stays frozen, and is used at both adaptation and inference.

The report covers remote-sensing image super-resolution specifically: evaluation uses UCMerced, AID, and RSSCN7 at two upscaling settings, with PSNR and SSIM as full-reference metrics and the no-reference MANIQA and CLIP-IQA evaluated at only one of the settings. Adaptation targets a frozen DiT4SR backbone, updating only the prototype cross-attention and SlotFiLM adapter, with the prototype memory built offline and kept frozen. It suits researchers and engineers working on diffusion-model adaptation and remote-sensing image restoration.

The prototype memory is built offline from the training distribution, which the authors say may limit adaptation to unseen sensors and degradations, leaving dynamic and more transferable structural priors to future work. On perceptual quality, MOSAIC obtains the lowest LPIPS on RSSCN7 and AID, while MANIQA and CLIP-IQA vary across datasets and remain competitive. Ablations run on the UCMerced validation set, so their absolute scores need not match the test-set numbers, and scores across exploratory tables are not directly compared.