{"id":"5cd75281-78e3-4d47-9f19-1d457a590ab1","arxiv_id":"2508.13503","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"AdaptiveAE uses deep reinforcement learning to pick ISO and shutter speed for HDR bursts, and reports improved fused-image quality in dynamic scenes.","lead":"A team trained a reinforcement learning agent to choose the ISO and shutter speed for each shot in an HDR burst, balancing noise against motion blur in dynamic scenes. On simulator-generated test images, the agent's exposure schedules produce higher reconstruction quality than three existing exposure-setting rules, with only qualitative confirmation on a real camera.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Quantitative SOTA claim is computed on LDRs synthesized by the same uncalibrated pipeline used for training; real-camera validation is qualitative only.","rationale":"The reader's weakest assumption identifies the simulator-to-real transfer as the crux; my concern is a concrete, testable component of that assumption: the noise model's borrowed camera constants. If the calibrated-noise re-evaluation preserves the margin, the noise-model concern is mitigated, but the blur model still requires external validation; the CONDITIONAL verdict remains appropriate. If the margin collapses, the central claim is invalidated, also consistent with CONDITIONAL (pending external tests). No new evidence leads me to a more severe verdict than the reader's.","tokens_in":16388,"tokens_out":17138,"duration_ms":168704,"concrete_test":"Capture flat-field frames on the SONY Alpha 7C-II for every ISO and shutter speed in the action space (Sec. C), fit U, sigma_read, and sigma_ADC from Eq. 3, then regenerate the HDRV test LDRs for all methods with these calibrated parameters and rerun the Table 1 metrics. If AdaptiveAE's PSNR-mu margin over Hasinoff et al. drops from 2.1 dB to below 1 dB, the reported SOTA is an artifact of the uncalibrated noise simulation.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Table 1's PSNR-mu advantage over Hasinoff et al. (39.70 vs 37.59) is measured on LDRs generated by the same blur/noise synthesis pipeline (Fig. 2, Sec. 3.1) that trained the RL agent. The noise model in Eq. 3 uses camera constants U, sigma_read, sigma_ADC borrowed from Hasinoff et al. [6] without calibration to the SONY Alpha 7C-II used in the real-capture demo. Both training and test LDRs share this single unvalidated simulator, so the reported gain may reflect the agent's specialization to those specific noise statistics rather than a genuine real-capture advantage. The only real-camera evidence is qualitative (Fig. 6), so the central claim of state-of-the-art performance is currently unverified outside the simulation loop.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes AdaptiveAE, a reinforcement-learning method that sequentially selects ISO and shutter speed for multi-exposure HDR capture in dynamic scenes. A synthesis pipeline is introduced to generate training LDRs with motion blur (via RIFE interpolation of consecutive HDR frames) and sensor noise (following the model of Hasinoff et al.), and an A3C actor-critic agent is trained with a reward combining reconstruction error, saliency-weighted priority loss, motion-mask ghost loss, and a step penalty. The agent refines exposure parameters in three stages and can also produce more than three frames. Experiments are reported on the Real-HDRV and DeepHDRVideo datasets using DeepHDR for fusion, with PSNR-\\mu, SSIM-\\mu, PU-PSNR, PU-SSIM, and HDR-VDP-2; the paper reports state-of-the-art numbers (e.g., PSNR-\\mu 39.70 on HDRV versus 37.59 for Hasinoff et al.) and also shows a qualitative real-camera demo on a SONY Alpha 7C-II.","tokens_in":16552,"tokens_out":4523,"duration_ms":45664,"significance":"If the reported gains transfer to real cameras, the contribution is significant: it jointly optimizes shutter speed and ISO during exposure bracketing, explicitly models motion blur and noise during capture, and formulates exposure selection as an MDP solvable with RL. The reward design is thoughtful, the synthesis pipeline is a reasonable training device, and the real-camera demonstration, though qualitative, indicates that the learned policy can produce plausible exposure decisions. The paper also ships a substantial set of ablations and a cross-fusion-method study. However, the central quantitative claim of state-of-the-art performance currently rests entirely on a simulator that is also the training distribution, so the significance outside the simulation loop is not yet established.","major_comments":[{"comment":"The headline quantitative results are obtained by synthesizing test LDRs with the same function S (Eq. A4) that generated the training data. Motion blur uses the same RIFE-based interpolation (Eq. 2) and noise uses the model of Eq. (3) with constants U, sigma_read, and sigma_ADC borrowed from Hasinoff et al. [6] without calibration to the SONY Alpha 7C-II used in the real-capture demo. Consequently, the reported PSNR-mu advantage (39.70 vs. 37.59 on HDRV) measures performance on the training distribution, not on independently captured real data. The only real-camera evidence is qualitative (Fig. 6). To support the state-of-the-art claim, the authors should provide quantitative evaluation on real captured bracketed sequences whose HDR ground truth is measured or independently generated, or alternatively calibrate the noise model to the target camera and show that the ranking of methods is stable across a range of simulator parameters (e.g., varying U and noise variances). Without such evidence, the central claim is not yet verified outside the simulation loop.","section":"§4.1, Table 1 and §3.1/Eq. (A4)"},{"comment":"The evaluation on DeepHDRVideo uses synthesized HDR ground truth for frames other than the middle frame, generated by DeepHDR [33]. Since DeepHDR is also the fusion method used both in the reward during training and in the main evaluation, this creates a circular dependency: the test metric may favor exposure settings that are well matched to DeepHDR's particular error patterns, and the 'ground truth' HDR for non-middle frames is itself a DeepHDR reconstruction. Please clarify exactly how the test LDRs and reference HDRs are generated for DeepHDRVideo, and either evaluate only on frames with native ground-truth HDR or synthesize the missing reference frames with a different method. This is required to interpret the cross-dataset generalization claim.","section":"§4.1, Cross datasets test"}],"minor_comments":[{"comment":"In Table A2, the Ours row reports SSIM-mu = 0.9208, while Table 1 reports SSIM-mu = 0.9408 for the same PSNR-mu = 39.70 and identical PU-PSNR/PU-SSIM values. This is likely a typographical error, but it should be corrected for consistency.","section":"Supplementary Table A2"},{"comment":"The captions refer to 'DeepHDR [7]', but reference [7] is RIFE; DeepHDR is reference [33]. The same incorrect citation appears in Figure 4 and in the Section 3.4 text mentioning the 'adopted fusion method [7]'. Please update these citations to [33].","section":"Table 1 and Figure 4 captions"},{"comment":"The paragraph beginning 'Current datasets [3, 8, 9, 13, 26] are inadequate...' appears twice, verbatim, in the introduction. One copy should be removed.","section":"Introduction, repeated paragraph"},{"comment":"The sentence 'With average exposure time n (≤ 30ms) and prediction time m (≤ 10ms), the total execution time is 6n + 3m (≤ 250ms)' uses symbols n and m that are not defined in the main text; the formula also seems inconsistent with the preceding description of 'six LDR captures'. Please clarify the counting and the symbols.","section":"Section 4.1, Inference time"},{"comment":"The definition of m_j as a floor of a ratio is clear, but the formula for the blurred HDR b^L_j would benefit from a brief explanation of the averaging over the interpolated frames, especially the case m_j = 0. A short sentence would improve readability.","section":"Section 3.1, Eq. (2)"}],"recommendation":"major_revision","confidential_remarks":"The central quantitative claim (state-of-the-art PSNR/SSIM on HDRV and DeepHDRVideo) is currently based entirely on the authors' own synthesis pipeline, which is also the training distribution. The real-camera demo is only qualitative. In my view, this is a fixable issue: the paper could be strengthened by adding a real-capture quantitative benchmark or by calibrating the simulator and demonstrating robustness to its parameters. The repeated incorrect citation of DeepHDR as [7] should also be corrected. The idea is promising and the paper is generally well organized, but the evaluation independence needs to be established before the SOTA claim can be accepted."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"AdaptiveAE is a real step beyond shutter-only RL exposure bracketing. It couples ISO and shutter selection, adds a blur/noise synthesis pipeline to training, and designs rewards that target salient and moving regions. The ablations support each design choice, the cross-fusion tests are a nice addition, and the real-camera demo suggests the idea transfers, at least visually. This is a worthwhile contribution to computational photography.\n\nWhere it gets wobbly is the quantitative claim. The test PSNRs in Table 1 are computed on LDRs generated by the authors' own synthesis function S, which is the same distribution used to train the agent. The noise model in Eq. 3 inherits camera constants U, sigma_read, sigma_ADC from Hasinoff et al. without calibration to the Sony Alpha used in the real capture. So the reported margins over Hasinoff and Wang are in-distribution with the simulator, not independent evidence of real-world superiority. The cross-dataset test on DeepHDRVideo uses the same synthesis pipeline, so it is a generalization test across video content, not across camera physics. That is not a fatal flaw, but it does mean the headline numbers should be read as 'performance on the authors' simulator' until externally validated. The only real-camera evidence is a single qualitative figure with manually set exposure, which is suggestive but not a quantitative test.\n\nThe internal consistency is good. The gap-to-best-achievable analysis is honest, and the ablation showing growth from Base to Base+Ppriority+Pghost is coherent. The step penalty and the option to output more than three frames are sensible engineering choices. I also appreciate that they show the fixed-ISO baselines with an oracle ISO (W-optimal) and still beat them, which strengthens the case on the simulator. The citation pattern looks fine; they are building directly on Wang and Hasinoff and say so.\n\nMy biggest ask for a revision would be an external quantitative evaluation: capture real LDR brackets with predicted settings on a static scene with controlled lighting (or use a real raw burst from a phone) and measure fused HDR quality, or at least calibrate the noise model to the actual camera and show the test set is not just the training simulator. Releasing code and the synthesis pipeline would also raise confidence.\n\nBottom line: this is a serious paper that deserves a proper peer review, and I would engage with it. It is not ready to be accepted as-is because the central SOTA claim is only verified inside the simulation loop. But the method is new, the formulation is sound, and the qualitative real-capture result points in the right direction.","headline":"A genuinely new joint ISO/shutter RL agent with a well-built simulator, but the SOTA numbers are only verified inside that simulator, not on real camera data.","tokens_in":17079,"tokens_out":1786,"would_cite":false,"duration_ms":17952,"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":"AdaptiveAE claims that a reinforcement-learning agent selecting ISO and shutter speed jointly captures higher-quality HDR images in dynamic scenes than fixed or shutter-only exposure rules.","keywords":["adaptive exposure","HDR imaging","reinforcement learning","motion blur synthesis","camera noise model","ISO and shutter speed","dynamic scenes","exposure bracketing"],"falsifier":"A concrete disconfirmation would be a real-camera experiment on dynamic scenes in which AdaptiveAE's chosen ISO and shutter-speed brackets, fused by the same network, fail to beat a fixed ±2 EV bracket with ISO set to a noise-optimal value, even though the simulated training pipeline predicts a clear advantage; that mismatch would show the synthetic blur-and-noise pipeline is not faithful enough to support the transfer claim.","tokens_in":16201,"feed_emoji":"📸","tokens_out":5671,"duration_ms":58248,"temperature":0.7,"pith_summary":"AdaptiveAE tries to establish that exposure bracketing for HDR should be treated as a sequential decision problem solved by reinforcement learning, rather than as a fixed or histogram-driven rule. The agent chooses each LDR's shutter speed and ISO together, balancing the noise a high ISO adds against the motion blur a long shutter speed causes, and it does so before the LDRs are fused. The paper argues that this capture-time optimization produces higher fused-HDR quality than prior exposure-setting methods, which either fix ISO, ignore motion, or tune only shutter speed. The reported evidence is a PSNR-µ of 39.70 on the HDRV test set versus 37.59 for Hasinoff et al. and 36.46 for Wang et al., with the advantage holding on DeepHDRVideo and across three different fusion networks.","feed_headline":"RL exposure policy lifts HDR quality in dynamic scenes","feed_subtitle":"AdaptiveAE picks ISO and shutter speed jointly, lifting HDR PSNR to 39.70 via blur- and noise-aware training.","key_machinery":"The machinery is an actor-critic reinforcement-learning agent (A3C) that casts exposure bracketing as a Markov decision process: the state is the current set of LDR images, and the action is a discrete ISO and shutter-speed pair from a fixed camera-parameter grid. The load-bearing component that makes it trainable is the blur-aware synthesis pipeline: motion blur is generated by interpolating consecutive HDR frames with RIFE and integrating the chosen shutter speed (Equation 2), and noise is added with the three-source camera noise model of Hasinoff et al. (Equation 3). This pipeline converts any proposed ISO/shutter-speed pair into a realistic LDR, enabling the policy to learn from the reward computed on DeepHDR-fused outputs.","core_discovery":"On the paper's own terms, the discovery is that a policy trained with asynchronous advantage actor-critic (A3C) can learn shutter-speed-and-ISO choices that optimize the final fused HDR image instead of a proxy such as SNR or histogram coverage. The agent operates in three refinement stages: it first re-anchors the mid-exposed frame to a new EV zero, then adjusts the underexposed frame, then the overexposed one, each choice conditioned on the previously synthesized LDRs. The reward is a negative sum of a reconstruction loss, a saliency-masked loss, a motion-masked ghost loss, and a penalty for brackets longer than three frames. The paper further claims that its blur-aware synthesis—motion blur from RIFE interpolation between consecutive HDR frames plus photon/readout/ADC noise from the Hasinoff model—is what lets the policy transfer to real dynamic scenes. The reported result is state-of-the-art fused HDR quality on Real-HDRV and DeepHDRVideo, with increasing robustness over baselines as scene motion grows.","pith_inferences":["If simulation-to-real transfer holds, a testable extension is to train the policy on a camera's own preview buffer (zero shutter lag) so it adapts to that specific sensor's noise and readout characteristics, potentially improving real-capture quality beyond the paper's fixed-noise-model experiment.","The same MDP formulation could be extended to optimize aperture as a third parameter, which the paper lists as future work; such an extension would let the policy trade depth of field against noise and blur.","One implicit consequence is that the optimal exposure policy depends on the fusion network; if the fusion network is retrained, the exposure policy might need recalibration, and jointly training both could yield further gains beyond the paper's frozen-fusion setup.","The motion-mask loss hints that exposure decisions could be conditioned on semantic scene understanding rather than saliency alone—for example, prioritizing text or faces—which could improve perceived quality in specific applications like document scanning or portrait photography."],"forward_implications":["If the policy is correct, exposure bracketing becomes scene-adaptive: the agent will assign faster shutter speeds to the reference frame when it detects motion, and rely on higher ISO for dark scenes, rather than using a fixed EV bracket.","Because capture-time choices are optimized, downstream fusion networks yield better results without any post-hoc deblurring or deghosting; the gap over prior exposure methods widens when a stronger fusion network is used.","The same trained policy transfers across datasets without retraining, as demonstrated by testing on DeepHDRVideo after training on Real-HDRV.","The reward design, with saliency and motion masks, implies that exposure choices can be steered toward the regions viewers care about most, such as faces and moving subjects.","The framework naturally extends to more than three frames when a scene demands them, with the step penalty keeping the bracket within the user's time budget."],"supporting_citations":[{"why":"Supplies the camera noise model (photon, readout, ADC) used in Eq. 3 to synthesize noisy LDRs and defines the noise-optimal exposure baseline.","marker":"[6]"},{"why":"RIFE frame interpolation is the core of the blur synthesis pipeline (Eq. 2), generating intermediate frames from which motion blur is accumulated based on shutter speed.","marker":"[7]"},{"why":"Real-HDRV (HDRV) is the main training and evaluation dataset used for the quantitative comparisons and ablation studies.","marker":"[26]"},{"why":"Wang et al. is the reinforcement-learning baseline that predicts shutter speed only, and it fails to account for motion; the paper compares against it and extends it by adding ISO control.","marker":"[32]"},{"why":"DeepHDR is the exposure-fusion network used to compute the reward during training and to produce fused HDR outputs for evaluation.","marker":"[33]"},{"why":"RAFT optical flow is used to compute the motion mask for the ghost penalty, identifying pixels with large motion that are prone to blur and ghosting artifacts.","marker":"[28]"},{"why":"SalGAN saliency prediction generates the importance mask used in the priority reward term, steering the agent to focus quality on salient regions.","marker":"[20]"}],"fun_headline_variants":["AdaptiveAE: RL tunes ISO and shutter for motion-robust HDR","Blur-aware RL exposure policy improves HDR in dynamic scenes","Joint shutter-ISO optimization via RL beats fixed bracketing","AI learns exposure from simulated blur and noise for HDR"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the synthesized LDRs—motion blur from RIFE interpolation plus the Hasinoff noise model—match what a real sensor would capture closely enough that a policy trained only on simulated images still picks the right exposure settings on a real camera.","fun_headline_variants_meta":{"raw":{"variants":["AdaptiveAE: RL tunes ISO and shutter for motion-robust HDR","Blur-aware RL exposure policy improves HDR in dynamic scenes","Joint shutter-ISO optimization via RL beats fixed bracketing","AI learns exposure from simulated blur and noise for HDR"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001128,"raw_usage":{"total_tokens":4687,"prompt_tokens":938,"completion_tokens":3749,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":554,"completion_tokens_details":{"reasoning_tokens":3676}},"tokens_in":554,"tokens_out":3749,"duration_ms":32796,"temperature":1.0,"reasoning_tokens":3676,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T17:13:20.889182+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A concrete disconfirmation would be a real-camera experiment on dynamic scenes in which AdaptiveAE's chosen ISO and shutter-speed brackets, fused by the same network, fail to beat a fixed ±2 EV bracket with ISO set to a noise-optimal value, even though the simulated training pipeline predicts a clear advantage; that mismatch would show the synthetic blur-and-noise pipeline is not faithful enough to support the transfer claim.","supporting_citations":[{"cited_title":"Noise-optimal capture for high dynamic range photography","cited_arxiv_id":null,"evidence_quote":"Supplies the camera noise model (photon, readout, ADC) used in Eq. 3 to synthesize noisy LDRs and defines the noise-optimal exposure baseline."},{"cited_title":"Real-time intermediate flow estimation for video frame interpolation","cited_arxiv_id":null,"evidence_quote":"RIFE frame interpolation is the core of the blur synthesis pipeline (Eq. 2), generating intermediate frames from which motion blur is accumulated based on shutter speed."},{"cited_title":"Towards real-world HDR video reconstruc- tion: A large-scale benchmark dataset and a two-stage align- ment network","cited_arxiv_id":null,"evidence_quote":"Real-HDRV (HDRV) is the main training and evaluation dataset used for the quantitative comparisons and ablation studies."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Wang et al. is the reinforcement-learning baseline that predicts shutter speed only, and it fails to account for motion; the paper compares against it and extends it by adding ISO control."},{"cited_title":"Deep high dynamic range imaging with large foreground motions","cited_arxiv_id":null,"evidence_quote":"DeepHDR is the exposure-fusion network used to compute the reward during training and to produce fused HDR outputs for evaluation."},{"cited_title":"RAFT: Recurrent all-pairs field transforms for optical flow","cited_arxiv_id":null,"evidence_quote":"RAFT optical flow is used to compute the motion mask for the ghost penalty, identifying pixels with large motion that are prone to blur and ghosting artifacts."}],"review_version":1}