Preserve
An identity embedding explicitly tells the model when no restoration is needed.
In brief
A restoration model that listens precisely.
Composite degradations are intertwined: removing one corruption can alter another, and processing order can change the result. CURE learns disentangled and adjustable representations so a single model can preserve the input, remove selected degradations, and continuously control restoration strength.
Read the full abstract
The presence of composite degradations poses a significant challenge, since the underlying corruption factors exhibit complex and interdependent interactions. Even when the degradation types are known, accurately restoring the image remains difficult due to the intertwined nature of their effects and the need for selective control during the recovery process. To address this, we introduce CURE, a unified framework that enables controllable restoration in complex degradation settings by learning disentangled and adjustable representations. CURE is driven by four complementary objectives. First, an identity embedding is incorporated, along with a reconstruction constraint, to ensure that the model can reproduce the input image when restoration is unnecessary. Second, the ratio control mechanism blends the identity embedding with degradation-specific embeddings using user-regulated mixing ratios, allowing continuous control over restoration intensity. Third, an intermediate loss supervises stepwise outputs, each encouraged to tackle only a single degradation factor within a composite mixture. Finally, a permutation-invariant loss produces consistent restoration quality regardless of the order in which multiple degradations are addressed.
Control
A ratio-controlled embedding continuously adjusts how strongly a degradation is removed.
Disentangle
Intermediate supervision isolates individual corruptions inside complex mixtures.
Reorder
Permutation-invariant learning keeps outputs consistent across restoration sequences.
01 Intensity control
Dial restoration up or down.
Move the slider to inspect the exact ratio-controlled outputs shown in the paper.
Snow example
Restoration ratio
w = 0.5
Live paper result
Partial restoration
The model removes part of the target degradation while retaining a controlled amount.
The interactive range uses the eight outputs visualized in Figure 3(a). In the formulation, w = 0 is identity and w = 1 is full restoration.
02 Selective restoration
Remove one. Preserve the rest.
CURE isolates a requested corruption instead of blindly clearing the entire image.
Low + Snow
Remove Low-light
Preserve Snow
Selective low-light enhancement should retain the snow component.
Composite input
OneRestore
CURE
Target
03 Order invariance
Same goal. Either order.
Sequential restoration should not depend on which degradation is removed first.
Input · Low + Snow
Delow → Desnow
vs.
Desnow → Delow
Ground truth
CURE produces closely aligned outputs across both processing orders.
04 Method
Controllability learned through four objectives.
CURE changes the training strategy, not the underlying restoration architecture.
Identity embedding
Constrains the model to reproduce the input when restoration is unnecessary.
Ratio-control loss
Interpolates identity and degradation embeddings for continuous intensity control.
Intermediate loss
Supervises partial outputs so each step removes only the requested degradation.
Permutation-invariant loss
Aligns outputs from reversed restoration sequences for consistent behavior.
05 Results
Qualitative comparisons on CCDD-11.
Figure 2 compares CURE with Restormer, NAFNet, PromptIR, and OneRestore across four complex degradation settings.
Key quantitative results
CURE with the OneRestore backbone on CCDD-11.
Average PSNR on CCDD-11
Selective restoration accuracy
Parameters
Average gap between restoration orders
06 Supplementary evidence
Every control, under a closer lens.
Explore twelve supplementary figures by experiment family. Open any figure for a full-resolution inspection.
When snow is meant to remain, the identity embedding preserves the degraded input instead of over-restoring it.
Figures 5–16 from the paper supplementary material, grouped and loaded by experiment family.
07 Dataset
CCDD-11
CDD-11 was originally introduced with OneRestore. CCDD-11 follows its composite-degradation synthesis pipeline, then adds supervision pairs for removing each degradation factor individually. For example, a Low + Haze input is paired with a Haze-only target when removing Low and a Low-only target when removing Haze. It also provides a half-restored target at approximately 50% restoration strength for every degradation—enabling both selective restoration and continuous intensity control rather than only full recovery.
Download CCDD-11 on Hugging Face
Original composite pipeline
CDD-11 by OneRestore
Add factor-level and half-restoration targets
Composite input
Low-light
+
Haze
CCDD-11 · supervision for selective and strength-controlled restoration
Citation
If CURE helps your research, please cite:
@inproceedings{kim2026cure,
title = {CURE: Controllable Unified Image Restoration for Complex Degradations},
author = {Kim, Boseong and Cho, Donghyeon},
booktitle = {International Conference on Pattern Recognition (ICPR)},
year = {2026}
}