- Rating
- 1
- Confidence
- 2
- Informativeness
- 3
- Presentation
- 3
- Soundness
- 3
Read review
Paper summary
This paper proposes the Representation Feature Adapter (RFA), a 2.9M-parameter network for unpaired image translation inside a frozen DINOv2 feature space. A frozen DINOv2 encoder maps the input to a feature map, the RFA maps it to a target-domain feature map, and a frozen one-step decoder (a DINO ControlNet on Sana-Sprint) renders the result. The adapter is trained with CycleGAN-style cycle and identity losses on features and with adversarial losses on the rendered image, whose gradients pass through the frozen decoder. The authors argue that the domain is a small residue of the feature map (13-14% of its norm) and show that the same adapter collapses to the identity on VAE latents. On adverse-weather removal, the RFA is ahead of CycleGAN-Turbo on fog and on night KID and level on snow, rain, and haze. On sim-to-real, it improves KID and FID over REGEN and HyPER-GAN, at a substantial cost in scene structure.
Critical strengths
S1. Clear diagnostic experiment. The comparison of the identical adapter on DINOv2 features and on two VAE latents (Table 1) is a clean and informative experiment, and it supports the claim that the choice of conditioning space determines what the adapter can change.
S2. Honest reporting of costs. The paper reports results that work against it, including halved detection mAP (App. G), lower structure retention, jackknife standard errors (App. H), and a decoder-only control (App. E).
S3. Efficiency and portability. A new domain pair requires training only a small adapter, which reduces per-condition training time to under a fifth of CycleGAN-Turbo's. Adapted features can also be rendered by decoders that never saw the adapter (PiD, RAE).
Critical weaknesses
W1. Limited technical contribution. Domain adaptation in feature space has been studied before (e.g., CyCADA), and decoders conditioned on frozen DINO features already exist (Control-DINO, RAE). The method mainly combines these with standard CycleGAN objectives and vision-aided discriminators. The "residue" is also measured as the adapter's own displacement, ||G(f) - f|| / ||f||, so it describes what the trained adapter does rather than an intrinsic property of the domain.
W2. Missing comparisons with modern editing and restoration models. The baselines are unpaired translators (CycleGAN, CUT, CycleGAN-Turbo) and Cosmos-Transfer. Instruction-based image editing models (e.g., FLUX.1 Kontext, Qwen-Image-Edit) can remove fog, rain, or snow and turn night into day zero-shot, and are known for strong structure preservation. Adverse-weather restoration models are also excluded on the grounds that their usual protocol is paired, although they could be run on the same crops. Without these comparisons, it is unclear whether the proposed approach has a practical advantage over currently available tools. I would be willing to raise my rating if the authors include recent image editing models as baselines and show that the RFA offers a clear advantage over them on the same crops, in particular on structure preservation and downstream detection.
W3. The structure loss undermines the target applications. On BDD100K snow and rain, a fixed detector's mAP on RFA outputs drops to about half of that on the untouched frames (0.148 vs. 0.276 and 0.146 vs. 0.254), whereas CycleGAN-Turbo keeps it (0.275 and 0.224; Table 12). On sim-to-real, only about a sixth of the detectable objects survive, compared with two thirds to three quarters for REGEN and HyPER-GAN. For driving applications, where restored frames are typically consumed by perception models, an output that is worse than doing nothing is a fundamental problem. The method appears to regenerate the scene rather than remove the weather from it, and the improved KID may partly reflect this.
W4. Evaluation protocol. KID and FID are computed on 50 crops per condition (25 for haze). Only fog and sim-to-real are separated by more than two standard errors (App. H), and FID on 50 images is rank-deficient, as the authors note. The "is the weather gone" classifier uses the same DINOv2 features that the adapter moves (App. F), so it is not an independent measure. In addition, the decoder is trained on Mapillary, which is also the sim-to-real target domain. Table 3 does not include a decoder-only row, so part of the sim-to-real KID gain may come from the decoder rather than the adapter.
Questions for authors
Q1. How does the RFA compare with instruction-based editing models (e.g., FLUX.1 Kontext, Qwen-Image-Edit) and adverse-weather restoration models on the same crops, in terms of KID, structure distance, and detection mAP?
Q2. What are the KID, FID, and structure distance of the decoder-only control on sim-to-real (Table 3)?
Q3. Can the authors evaluate weather removal with a classifier or metric that does not read DINOv2 features?
Q4. In which practical setting is the trade-off of the RFA preferable to that of CycleGAN-Turbo, given that detection mAP falls below that of the untouched input?