{"id":"10625d46-5037-4891-9a28-b27bf4dfb84c","arxiv_id":"2505.23102","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"An RL agent iteratively adjusts a Bezier tone curve, rewarded by CLIP text-image similarity, for fast zero-reference low-light and multi-exposure image enhancement.","lead":"CURVE trains a reinforcement learning agent to enhance dark or malformed images by repeatedly adjusting a Bezier tone curve, with rewards from CLIP text prompts. It matches or beats several existing enhancement methods while running much faster on high-resolution images than most competitors.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Training on VOC plus N(0,1) tone curves is the load-bearing assumption: if the CLIP reward only tracks the synthetic augmentation family, the claimed zero-reference transfer to LOLv1/2 and SICE is unsupported.","rationale":"The strongest claim is an empirical one: best-or-second-best quality plus order-of-magnitude speed. The speed part is largely supported, though SCI is a near tie at UHD and faster at HD and FHD, so 'order of magnitude' should be read as relative to most, not all, baselines. The quality part rests on the distribution-shift assumption. The policy is trained only on VOC images with random Bezier tone curves (Sec. 3.1); the CLIP reward (Eq. 6) is never evaluated on real low-light inputs during training. The model could simply learn an inverse mapping for the N(0,1) curve family, which is a narrower claim than zero-reference enhancement of arbitrary low-light images. This is not contradicted by Table 1 alone, because those test sets may share histogram properties with the synthetic augmentation. The proposed check separates the two hypotheses by training on real unlabeled low-light images and comparing. I also flag that the SICE multi-exposure result is based on a single baseline and contains a contradictory sentence ('outperformed by the baseline methods' versus Table 4), which weakens the multi-exposure part of the claim. The train-by-loss ablation's 50x iteration gap is a real concern about the RL justification, but it is secondary to whether the learned policy transfers at all. The paper's own conclusion acknowledges global-processing limitations, which are consistent with the method's design rather than a fatal flaw.","tokens_in":904,"tokens_out":2036,"duration_ms":186169,"concrete_test":"Train the identical CURVE/SAC policy with the same CLIP reward on two sources: (a) VOC 2007 with the published N(0,1) augmentation, and (b) the LOLv1 training split used as unlabeled images (no ground truth, consistent with zero-reference training). Evaluate both policies on LOLv1 and LOLv2 test splits over three seeds. If the VOC-trained policy's PSNR and SSIM are within one standard deviation of the real-unlabeled-trained policy, the transfer assumption holds; if the real-trained policy is materially better, the reported enhancement is specific to the synthetic augmentation family rather than to the CLIP reward or RL.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim depends on a policy trained (Sec. 3.1) on VOC 2007 images with random Bezier tone curves sampled from N(0,1), with rewards computed only from CLIP text-image similarity (Eqs. 6-7). At test, the same policy is applied to real low-light and multi-exposure images. The paper never shows that maximizing this CLIP reward on synthetic curve-degraded VOC images tracks perceptual quality on real under/over-exposed photographs. The reported PSNR/SSIM numbers could be a distributional artifact: if the N(0,1) curves induce brightness and contrast shifts similar to those in the test sets, the policy may learn to invert that specific curve family rather than to enhance arbitrary illumination defects. There is no analysis of sensitivity to the augmentation distribution, no ablation trained on real unlabeled low-light images, and no held-out synthetic validation. The multi-exposure evidence is particularly thin: only Zero-DCE is compared on SICE Part 2, and the main text's statement that 'Our approach is outperformed by the baseline methods' directly contradicts Table 4, which shows CURVE ahead on SSIM, PSNR, and LPIPS. The transfer assumption, not any single metric, is the load-bearing element.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes CURVE, a zero-reference low-light image enhancement method that combines Soft Actor-Critic (SAC) reinforcement learning with a global Bezier tone-curve adjustment module. The reward is computed from CLIP text-image similarity, using object-class-conditioned positive prompts and a fixed negative prompt. At test time, the policy is run on a small downsampled image, and a composed look-up table (LUT) is applied to the full-resolution image. The paper reports SSIM and PSNR results on LOLv1, LOLv2Real, LOLv2Syn, and SICE Part 2, claiming best or second-best performance among compared zero-reference methods while being substantially faster on HD/FHD/UHD resolutions. The main claim is that a simple iterative global tone-curve policy trained with a CLIP-based reward can match or outperform conventional CNN-based enhancers without paired data.","tokens_in":14633,"tokens_out":4514,"duration_ms":44161,"significance":"If the reported results hold, the paper makes a practical contribution: it demonstrates that a compact global tone-curve policy can compete with much larger CNN-based zero-reference enhancers on standard low-light benchmarks, and the LUT-based implementation gives a genuine speed advantage on high-resolution images. The use of CLIP as a reward signal with object-conditioned prompts is a sensible and reasonably novel design, and the SAC formulation is clearly described. The paper does not provide code or trained models, but the method is simple enough to re-implement. The main risk is that the policy is trained exclusively on VOC 2007 images augmented with Bezier curves sampled from N(0,1), and the paper does not establish that the CLIP reward on these synthetic augmentations tracks perceptual quality on real under- or over-exposed photographs. This transfer assumption is load-bearing for the claimed zero-reference generalization.","major_comments":[{"comment":"The policy is trained only on VOC 2007 images augmented with Bezier tone curves sampled from N(0,1), yet it is evaluated on real low-light (LOLv1/v2) and multi-exposure (SICE) images without any domain adaptation or analysis. The paper does not show that maximizing the CLIP reward on this synthetic augmentation family tracks perceptual quality on real illumination defects. I ask for (i) an ablation trained on real unlabeled low-light images, (ii) a sensitivity analysis with different augmentation distributions, and (iii) a held-out synthetic validation demonstrating that the learned reward correlates with PSNR/SSIM. Without this, the reported zero-reference gains could be a distributional artifact of the specific N(0,1) curve family rather than a general enhancement capability.","section":"Sec. 3.1 and Sec. 3.2"},{"comment":"The sentence 'Our approach is outperformed by the baseline methods' directly contradicts Table 4, in which CURVE is better than Zero-DCE on SSIM, PSNR, and LPIPS. Please correct the text and state which comparison is intended. As written, the paper undercuts its own SICE result and leaves the reader uncertain about which numbers are trustworthy.","section":"Sec. 3.2, Table 4"},{"comment":"The comparison with 'train-by-loss' is not equal in training effort: CURVE is trained for 7.5e5 SAC iterations, while train-by-loss uses only 15k iterations. Moreover, no standard deviations or multiple-seed results are reported for any method. Since SAC is stochastic and the reported margins are small (e.g., SSIM 0.7164 versus 0.7035 on LOLv1), a single run does not establish superiority. Please report mean and standard deviation over at least three seeds and align the iteration count for the ablation.","section":"Sec. 3.1"},{"comment":"The reward uses object-class-conditioned positive prompts, but the paper does not specify how object classes are detected at test time. Is a pretrained detector used, or are VOC ground-truth labels assumed? For LOL and SICE images, no such labels exist, so this detail is essential for reproducing the method. The paper also does not state what happens when N=0 (no detected classes), in which case Eq. (6) is undefined.","section":"Sec. 2.2.2"}],"minor_comments":[{"comment":"The negative prompt text is written as 'a bad, saturated, blacked out photo of nothing' in the main text but 'a bad, saturated and blacked-out photo of nothing' in Fig. 7(b); please make the wording consistent.","section":"Sec. 2.2.2 and Fig. 7"},{"comment":"The policy samples actions from a Gaussian and then applies tanh and rescaling, but the rescaling factor that maps tanh outputs to the claimed action range [-2, 2] is not stated. Please clarify.","section":"Sec. 2.2.3 and Table 2"},{"comment":"The operation B(l_t, a_t) on a LUT vector is not defined; the Bezier module B was introduced for images, so please explain how it is applied to a 1-D lookup table.","section":"Sec. 2.3, Algorithm 1"},{"comment":"The runtime for ReLLIE is measured on CPU while all other methods are measured on GPU, making the speed comparison for ReLLIE not directly comparable. Please add an explicit caveat in the table or text.","section":"Table 1 and Sec. 3.1"}],"recommendation":"major_revision","confidential_remarks":"The paper is a compact application-oriented contribution that fits the venue. The central issue is the training-distribution transfer assumption; the authors should be asked to add the suggested experiments. The text/table contradiction in Sec. 3.2 is likely a copy-editing error but is serious because it reverses the reported result. Also, the absence of error bars is particularly problematic for an RL method where small performance differences are reported."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's the take: the paper's actual novelty is modest but real—SAC learning to iteratively adjust a Bezier tone curve, rewarded by CLIP text-image similarity, with the whole sequence composed into one LUT at test time. The LUT composition is the clever part; it explains the 0.017 s/frame on UHD, and that speed claim is credible given the design.\n\nWhat the paper does well: the comparisons on LOLv1, LOLv2Real, and LOLv2Syn show CURVE at or near the top among zero-reference methods, and the speed advantage over CLIP-LIT and ReLLIE is substantial. The authors also borrow the Bezier module from their own prior work and cite it, which is fair.\n\nThe soft spots: first, the train-by-loss ablation in Table 1 runs 15k iterations against 750k for the full RL method. That 50x gap makes the ablation meaningless as evidence that RL is needed. Second, the SICE section says \"Our approach is outperformed by the baseline methods\" while Table 4 shows CURVE ahead on SSIM, PSNR, and LPIPS. That's a direct contradiction and should be fixed. Third, there are no error bars or multiple-seed results, so the small SSIM/PSNR improvements could be noise.\n\nThe stress-test worry about training on VOC 2007 with N(0,1) tone curves is real but not fatal. The paper doesn't analyze how sensitive the policy is to the augmentation distribution, and it doesn't show that maximizing the CLIP reward on synthetic curves tracks perceptual quality on real low-light images. But the same policy does transfer to three real datasets, which is weak evidence that the transfer assumption holds. A sensitivity analysis would make the claim much stronger; without it, the generalization is plausible but not fully supported.\n\nOverall, this is a pragmatic engineering paper with one genuinely useful trick (the LUT composition) and a solid empirical comparison. It deserves peer review, but the reviewers should ask for a fair RL ablation, error bars, and a fix for the SICE contradiction. I'd bring it to a reading group if you're interested in fast zero-reference enhancement.","headline":"A pragmatic RL+CLIP tone-curve method with a clever LUT speedup; the evaluation is mostly solid but the RL ablation is unfair and the SICE text contradicts its own table.","tokens_in":15193,"tokens_out":2702,"would_cite":false,"duration_ms":27413,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"CURVE claims that zero-reference image enhancement can be reduced to a sequence of global tone-curve adjustments, chosen by a reinforcement-learning policy and rewarded by CLIP text similarity, and that this simple recipe matches or beats…","keywords":["low-light image enhancement","zero-reference enhancement","reinforcement learning","CLIP","Bezier curve","tone mapping","look-up table","Soft Actor-Critic"],"falsifier":"Train CURVE exactly as described but replace the CLIP reward with a reward that only maximizes global brightness or contrast; if the resulting images score the same on LOLv1 SSIM and PSNR, then the CLIP text supervision is not essential. Conversely, run the trained policy on real low-light images with heavy sensor noise: since the action is a global tone curve, any large drop in perceived quality relative to clean synthetic test images would show that the synthetic Bezier-augmentation training distribution, not the CLIP reward, is the brittle link.","tokens_in":14164,"feed_emoji":"🌙","tokens_out":6339,"duration_ms":57626,"temperature":0.7,"pith_summary":"CURVE claims that zero-reference low-light and multi-exposure image enhancement can be reduced to a sequence of global tone-curve adjustments chosen by a reinforcement-learning policy. The reward is not a hand-designed image statistic but the agreement, in CLIP embedding space, between the enhanced image and text prompts such as \"a good photo of {object class}\", contrasted with a fixed negative prompt. Because the policy operates on a small 56×56 state and the final tone mapping is applied through a composed lookup table, the method reports roughly constant processing time across HD, FHD, and UHD resolutions, about 0.017 seconds per frame. On LOLv1, LOLv2Real, and LOLv2Syn the method reaches the best or second-best SSIM and PSNR among the zero-reference baselines compared, and on the multi-exposure SICE dataset it keeps brightness consistent across exposure levels. A sympathetic reading is that perceptual language supervision plus iterative simple processing is a viable alternative to heavier learned networks.","feed_headline":"CLIP-guided RL fixes dark photos in 0.017 seconds","feed_subtitle":"A four-parameter Bezier curve, adjusted step by step, matches or beats bigger networks on LOLv1 and LOLv2.","key_machinery":"The load-bearing object is the Bezier-curve tone adjustment module: a cubic Bezier curve with endpoints fixed at $(0,0)$ and $(1,1)$, whose two interior control points are moved by four action parameters, and which is evaluated as a piecewise-linear mapping so a full-resolution image can be transformed by a single lookup-table application. Around this module, the paper builds a Soft Actor-Critic (SAC) loop, an off-policy maximum-entropy reinforcement learning algorithm, in which the state is the concatenation of the current and previous small images ($x_t$ and $x_t - x_{t-1}$), the action is the four-parameter curve adjustment, and the reward is the decrease in a CLIP-based softmax cross-entropy loss that pushes the image embedding toward \"a good photo of {class}\" and away from \"a bad, saturated, blacked out photo of nothing\". The test-time trick is to apply the learned sequence of actions to a lookup table over all $2^{bit}$ pixel values rather than to the image itself, so the whole episode composes into one LUT that maps the original high-resolution image directly to the final result.","core_discovery":"On its own terms, the paper's claim is that CURVE establishes a new point in the design space of zero-reference low-light image enhancement: use a frozen CLIP model to define what \"good\" means, train a Soft Actor-Critic policy to take repeated small actions, and let each action modify the control points of a global cubic Bezier tone curve. The action vector $a_t = [\\theta_1, \\theta_2, r_1, r_2]$ is parameterized so that zero actions leave the image unchanged, and the reward $r_t = \\beta(L_t - L_{t+1})$ rewards any step that reduces a softmax cross-entropy loss between the CLIP image embedding and positive and negative text embeddings. The policy is trained on VOC 2007 images with randomly sampled Bezier augmentations and tested directly on low-light and multi-exposure datasets. The reported result is either the best or second-best SSIM and PSNR among the compared zero-reference methods on all three LOL datasets, with a processing speed an order of magnitude faster than most baselines at high resolutions because the state is tiny and the full-resolution image is never processed until the final lookup-table application.","pith_inferences":["If CLIP-text similarity is the right perceptual proxy, the same reward could steer other global image operators such as white balance, saturation, or gamma by swapping only the differentiable processing module, since the RL loop is operator-agnostic.","The reported UHD speed of 0.017 s/frame implies roughly 60 frames per second, so a frame-wise video extension is plausible; the paper does not address temporal flicker, which would be the next obstacle.","The positive prompt is built from object classes detected in the training image, coupling enhancement to recognition; this may make CURVE especially suited to preprocessing for detection, at the possible cost of human-perceptual quality on scenes with no clear object.","The global tone curve cannot repair local contrast or remove noise; the paper itself notes amplification of noise and tonal differences on over-exposed images, so a spatially varying extension would need to sacrifice the LUT speed trick."],"forward_implications":["Because the policy and Q-networks process only a 56×56 state, test-time cost is nearly independent of resolution; the paper reports 0.017 s/frame for HD, FHD, and UHD on an RTX 3080, versus 0.205 s for Zero-DCE and 20.68 s for CLIP-LIT at UHD.","The lookup-table composition means arbitrarily many iterative adjustments can be applied to a high-resolution image in one pass, so the speed advantage should persist for longer episodes.","Training on VOC 2007 with random Bezier-curve augmentations transfers to LOLv1, LOLv2Real, LOLv2Syn, and SICE, suggesting that synthetic tone-curve degradations can stand in for real low-light training data.","The RL formulation keeps improving with more steps and stays stable, while the train-by-loss ablation peaks and then degrades (Fig. 4), so the reward-as-improvement design is doing real work.","The same framework handles under-exposed and over-exposed images on SICE, because the policy learns to move brightness in the direction that lowers the CLIP loss."],"supporting_citations":[{"why":"Zero-DCE is the tone-curve enhancement baseline CURVE compares against and supplies the SICE evaluation protocol.","marker":"[3]"},{"why":"CLIP provides the frozen image and text encoders used to define the reward.","marker":"[4]"},{"why":"CLIP-LIT is the CLIP-based enhancement baseline that CURVE compares against, including its UHD runtime.","marker":"[5]"},{"why":"CLIP-IQA supplies the idea of comparing image embeddings against good and bad text prompts, which the reward adapts.","marker":"[8]"},{"why":"Ogino et al. supplies the Bezier-curve tone adjustment module and the random-augmentation training recipe CURVE builds on.","marker":"[11]"},{"why":"ReLLIE is the reinforcement-learning enhancement baseline and the CPU-timing comparison point.","marker":"[15]"},{"why":"The Soft Actor-Critic algorithm provides the off-policy maximum-entropy RL method used for training.","marker":"[16]"},{"why":"The SAC applications paper supplies the soft Bellman residual and policy loss used to update the networks.","marker":"[17]"},{"why":"The VOC 2007 dataset is the training source whose images are randomly augmented with Bezier curves.","marker":"[22]"}],"fun_headline_variants":["CLIP-guides RL to adjust Bezier curve for low-light photos","Fast low-light enhancement with CLIP-trained RL and a simple curve","Bezier curve + CLIP reward = fast low-light fix","Zero-reference LLIE: CLIP-guided RL on a four-parameter curve"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a policy trained on VOC 2007 images randomly darkened or brightened with Bezier curves drawn from N(0,1), and rewarded by CLIP text similarity, will enhance real low-light and multi-exposure images without any real paired or unpaired low-light training data.","fun_headline_variants_meta":{"raw":{"variants":["CLIP-guides RL to adjust Bezier curve for low-light photos","Fast low-light enhancement with CLIP-trained RL and a simple curve","Bezier curve + CLIP reward = fast low-light fix","Zero-reference LLIE: CLIP-guided RL on a four-parameter curve"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001267,"raw_usage":{"total_tokens":5176,"prompt_tokens":923,"completion_tokens":4253,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":539,"completion_tokens_details":{"reasoning_tokens":4176}},"tokens_in":539,"tokens_out":4253,"duration_ms":32148,"temperature":1.0,"reasoning_tokens":4176,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T12:53:09.040533+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train CURVE exactly as described but replace the CLIP reward with a reward that only maximizes global brightness or contrast; if the resulting images score the same on LOLv1 SSIM and PSNR, then the CLIP text supervision is not essential. Conversely, run the trained policy on real low-light images with heavy sensor noise: since the action is a global tone curve, any large drop in perceived quality relative to clean synthetic test images would show that the synthetic Bezier-augmentation training distribution, not the CLIP reward, is the brittle link.","supporting_citations":[{"cited_title":"Experimental Settings Implementations and training details.We used the VOC 2007 trainval dataset [22] for training","cited_arxiv_id":null,"evidence_quote":"Zero-DCE is the tone-curve enhancement baseline CURVE compares against and supplies the SICE evaluation protocol."},{"cited_title":"Our approach addresses both low-light and multi-exposure image enhancement challenges while maintaining computational efficiency","cited_arxiv_id":null,"evidence_quote":"CLIP provides the frozen image and text encoders used to define the reward."},{"cited_title":"Iterative prompt learning for unsupervised backlit image enhancement,","cited_arxiv_id":null,"evidence_quote":"CLIP-LIT is the CLIP-based enhancement baseline that CURVE compares against, including its UHD runtime."},{"cited_title":"CURVE: CLIP-Utilized Reinforcement Learning for Visual Image Enhancement via Simple Image Processing","cited_arxiv_id":"2505.23102","evidence_quote":"CLIP-IQA supplies the idea of comparing image embeddings against good and bad text prompts, which the reward adapts."},{"cited_title":"Unsupervised image prior via prompt learning and CLIP semantic guidance for low-light image en- hancement,","cited_arxiv_id":null,"evidence_quote":"Ogino et al. supplies the Bezier-curve tone adjustment module and the random-augmentation training recipe CURVE builds on."},{"cited_title":"Discover- ing an image-adaptive coordinate system for photogra- phy processing,","cited_arxiv_id":null,"evidence_quote":"ReLLIE is the reinforcement-learning enhancement baseline and the CPU-timing comparison point."},{"cited_title":"Learning to control camera exposure via reinforcement learning,","cited_arxiv_id":null,"evidence_quote":"The SAC applications paper supplies the soft Bellman residual and policy loss used to update the networks."}],"review_version":1}