{"id":"db4bdaf4-d73a-4c20-9931-41f492d65795","arxiv_id":"2412.14705","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Event streams between extreme-exposure RGB frames enable explicit motion alignment, and a pre-trained diffusion prior fuses the brackets into ghost-reduced 12-stop HDR images; the authors also introduce the first 12-stop event-HDR dataset.","lead":"An event camera paired with an RGB camera is used to align five photos taken from -6 to +6 exposure stops, letting a diffusion model fuse them into a 12-stop HDR image of moving scenes. The work also contributes a new simulated dataset and a small real-world validation rig, and reports higher quality scores than prior HDR methods.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Real-world 12-stop claim is under-supported: Table 2 uses only non-reference IQA on 16 self-captured scenes, which can reward smoothing and does not verify radiometric range.","rationale":"The reader's weakest assumption identifies the same load-bearing gap: real-world performance is evaluated only through non-reference metrics on a small self-captured set, with no ground-truth HDR. My stress test confirms that this is the weakest point in the argument chain from 'strong simulated results' to 'state-of-the-art in real dynamic scenes.' I additionally note that the paper never directly measures the dynamic range of the output HDR, and that the Ours* configuration used in Table 1 includes a benchmark-specific VAE-decoder color-correction step, which slightly weakens the attribution of the margin to the proposed alignment and fusion modules. None of this constitutes evidence of error or bad faith; the simulated experiments and ablations are useful and internally consistent. It does mean the central real-world claim is not yet fully established, so conditional acceptance remains the right verdict rather than rejection or unconditional acceptance.","tokens_in":15134,"tokens_out":4942,"duration_ms":45279,"concrete_test":"Recruit a reference-HDR validation: capture 10 static high-contrast scenes on a tripod, acquire a 12-stop ground truth via multiple exposures and rank-minimization/median merge, then add a controlled moving object or hand-held shake and run all methods. Compute mu-PSNR, mu-SSIM, and HDR-VDP-2 on the moving/occlusion regions, and compare Ours versus Event-HDR* and HDR-Fusion. If Ours does not rank first on these reference metrics (or within noise), the real-world SOTA claim in Table 2 is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is empirical: event-assisted dual-camera alignment plus diffusion fusion produces ghost-free 12-stop HDR in real dynamic scenes. The simulated ESHDR results (Table 1) support the method on synthetic data, but the real-world evidence (Sec. 5.4, Table 2) consists of 16 self-captured scenes evaluated only with MUSIQ, MANIQA, and HyperIQA. These no-reference metrics are known to prefer perceptually clean, often smoother images; they do not measure radiometric fidelity, alignment accuracy, or whether the output actually spans the claimed 12 stops. The paper's own color-correction fine-tune (Sec. A.2) is benchmark-specific: one VAE decoder layer is trained on ESHDR color, and baselines are color-aligned to ground truth, so part of the Table 1 margin may come from this post-hoc correction rather than from the event-plus-diffusion pipeline. If the real-world scenes are unrepresentative or the no-reference scores are driven by the diffusion prior's smoothing, the statement that the method 'achieves state-of-the-art performance' and 'captures the full 12-stop range' in real scenes is not supported. This does not invalidate the synthetic contribution, but it makes the real-world generalization claim load-bearing and currently under-tested.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes a 12-stop dynamic HDR imaging pipeline built around a dual-camera RGB-event system. Five LDR images spanning -6EV to +6EV are first pre-processed by a UNet, then explicitly aligned by an event-assisted flow module that combines a RAFT initialization with event streams, and finally fused by a latent diffusion model with ControlNet conditioning. The authors introduce the ESHDR simulated dataset with synchronized events, a real-world dual-camera dataset of 16 scenes, and report state-of-the-art results on simulated and real-world benchmarks. The central claim is that the method produces ghost-free 12-stop HDR reconstructions in real dynamic scenes, going beyond the 6-stop range of prior event-RGB HDR methods.","tokens_in":15433,"tokens_out":7548,"duration_ms":77597,"significance":"If validated, this would be a meaningful advance: the paper introduces the first simulated dataset for 12-stop event-HDR imaging, a practical dual-camera capture system, an explicit event-assisted alignment design, and a diffusion-prior fusion module. The simulated evaluation shows a large margin over the best event baseline (PSNR 38.25 vs 33.85), and the ablations in Fig. 3 are informative about the contribution of the alignment and fusion modules. However, the real-world generalization claim is currently supported only by non-reference metrics on 16 self-captured scenes, and the simulated comparison is partly asymmetric. The reported Ours* numbers also rely on a benchmark-specific color-correction fine-tune that is not part of the core method. The core idea is sound and the dataset contribution is useful, but the headline claim of full 12-stop real-world performance needs stronger evidence before publication.","major_comments":[{"comment":"The real-world evaluation that supports the headline claim ('capturing the full 12-stop range' in dynamic scenes) consists of 16 self-captured scenes measured only with the no-reference metrics MUSIQ, MANIQA, and HyperIQA. These metrics can reward perceptually clean, often smoothed, outputs and do not verify radiometric fidelity, alignment accuracy, or whether the reconstructed HDR actually spans 12 stops. With no ground-truth HDR or calibrated reference on the real set, the reported superiority in Table 2 does not establish the real-world generalization claimed in the abstract and conclusion. Please add a reference-based real-world validation, for example a static HDR reference captured with a calibrated target or a denser exposure bracket, or explicitly restrict the real-world claim to qualitative demonstration.","section":"Sec. 5.4, Table 2"},{"comment":"The main simulated comparison is asymmetric: HDRFlow and HDR-Transformer are not retrained on the ESHDR training set, while HDRV* and Event-HDR* are retrained on it. The RGB baselines are also evaluated with exposure ranges (3 or 6 stops) different from the 12-stop input, and all results are color-aligned to ground truth with a 3D bilateral grid. This makes the large margin over RGB baselines difficult to attribute to the proposed alignment and fusion modules rather than to training-domain mismatch or input range. Please retrain the RGB baselines on the same ESHDR split (or report their native-domain performance alongside), and state explicitly which color-alignment procedure is applied to each method.","section":"Sec. 5.2, Table 1"},{"comment":"Table 1 reports results for 'Ours*', but this variant includes a VAE-decoder color-correction layer that is fine-tuned specifically to improve scores on the ESHDR evaluation set, and the paper states this step is not part of the core method described in Section 3. Because only the color-correction module is updated during fine-tuning and all other parameters are frozen, part of the reported gain may come from a benchmark-specific post-hoc correction rather than from the event-assisted alignment and diffusion fusion. Please report the unmodified 'Ours' in Table 1 as well, describe the exact training data used for the color-correction fine-tuning, and analyze the contribution of this module separately.","section":"Supplementary A.2; Table 1"},{"comment":"The real-world fine-tuning uses frame-interpolation datasets [8,35] and randomly converts each frame to an LDR image with an exposure time between -6EV and 6EV, then supervises alignment toward the converted second frame. This provides real event statistics, but the supervision is still a single-exposure LDR frame, not a 12-stop HDR ground truth. It therefore does not directly validate whether the alignment module preserves information across the full exposure range in real scenes. Please provide an additional real-world alignment evaluation, such as flow error on a known motion pattern or a small set of real bracketed HDR references, to close this gap.","section":"Sec. 3.2, Real-world fine-tuning"}],"minor_comments":[{"comment":"The column header 'MUSID' should be 'MUSIQ'.","section":"Table 2"},{"comment":"The row label 'HDR-Event*' is inconsistent with 'Event-HDR*' used elsewhere in the paper; please unify the notation.","section":"Table 2"},{"comment":"The phrase 'previous event-based HDR [2,47]' cites event camera hardware papers; the relevant event-based HDR methods are Refs. [38,48].","section":"Sec. 3.2"},{"comment":"The conditioning representation is written as 'z_a = VAE-Decoder({I^a_n})'; in a latent diffusion pipeline the conditioning images should be encoded into latent space, so this is likely a typo for 'VAE-Encoder' and should be corrected or clarified.","section":"Sec. 3.3, Eq. (5)"},{"comment":"The section title 'Color correction for HSHDR dataset' appears to contain a typo and should read 'ESHDR'.","section":"Supplementary A.2"},{"comment":"The paper does not report the RGB camera's bit depth or the exact exposure times used for the -6EV to +6EV captures; these details are needed to judge whether the real-world captures actually cover the claimed 12 stops.","section":"Sec. 5.4"}],"recommendation":"major_revision","confidential_remarks":"I believe the core technical contribution and the ESHDR dataset are valuable, and the simulated experiments are largely convincing. The main risk is overclaiming real-world 12-stop performance from 16 scenes with no-reference metrics only. If the authors can add a credible reference-based real-world validation or carefully temper the real-world claim, and make the baseline comparison fair, I would be willing to support acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The genuinely new thing here is the task itself: 12-stop HDR fusion in dynamic scenes with synchronized events, plus the ESHDR dataset to train and evaluate on it. Prior event-based HDR methods top out around 4-6 stops, and 12-stop fusion was static-only. That is a meaningful extension, and the dataset, even if not yet released, fills a real gap. The method is also sensible: explicit event-assisted alignment with a real-world fine-tuning stage, then a diffusion-based fusion module with ControlNet conditioning. The simulated results in Table 1 are impressive, and the ablation study is useful, especially showing that the diffusion prior beats a CNN fusion baseline in compressed bright regions. The paper is honest about the sim-to-real gap and about the hardware limitations of the beam-splitter setup.\n\nNow the soft spots. The real-world generalization claim rests on Table 2: 16 self-captured scenes, no ground-truth HDR, and only no-reference metrics (MUSIQ, MANIQA, HyperIQA). Those metrics reward perceptually clean images and can be fooled by smoothing; they do not verify that the output actually spans 12 stops or that the alignment is geometrically correct. The stress-test note is on target here. Additionally, the color-correction fine-tune in Supplementary A.2 is explicitly benchmark-specific: it improves mu-PSNR on ESHDR but slightly degrades textures. Reporting that as Ours* in Table 1 mixes the contribution of the event-plus-diffusion pipeline with a separate post-hoc correction. That is not fatal, but it should be separated in the tables. Also, HDRFlow and HDR-Transformer were not retrained on 12-stop inputs, so the comparison is partially asymmetric, though the event baselines were retrained, which is fair.\n\nThis paper deserves a serious referee. The core task and the simulated evaluation are valuable enough to merit peer review, but the referee should insist on either a real-world validation with reference HDR (even if captured with a different method) or a human perceptual study, and on better separation of the color-correction step. I would not cite it yet without code/data access, but I would read it carefully if I work in HDR or event-based imaging.","headline":"The ESHDR dataset and 12-stop dynamic HDR task are a real step forward, but the real-world evidence is too thin to support the headline claim.","tokens_in":705,"tokens_out":1006,"would_cite":false,"duration_ms":23305,"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":"An event camera next to an RGB camera allows 12-stop HDR reconstruction of moving scenes, with no ghosting from large motion.","keywords":["high dynamic range imaging","event camera","dynamic scenes","ghosting artifacts","multi-exposure fusion","diffusion-based fusion","RGB-event dual camera","12-stop HDR"],"falsifier":"Capture a dynamic 12-stop scene with the same dual-camera rig while also recording a static ground-truth HDR of the same scene, then compare per-pixel radiance in the tone-mapped domain; if the method's output deviates from the reference in bright or dark regions beyond alignment tolerance, the ghost-free 12-stop claim is refuted.","tokens_in":14941,"feed_emoji":"📸","tokens_out":8722,"duration_ms":67269,"temperature":0.7,"pith_summary":"This paper tries to establish that dynamic-scene HDR photography can span the full 12-stop range that high-contrast real scenes require, not just the 6-stop range of earlier fusion methods. Its proposal is to put an event camera beside an RGB camera: the event stream supplies dense motion information that survives extreme exposure differences, so five LDR frames from -6 EV to +6 EV can be aligned without ghosting. A diffusion-based fusion module then reconstructs the HDR image, using pretrained image priors to suppress artifacts in compressed bright and dark regions and to tolerate alignment mistakes. If correct, a hand-held dual-camera rig could capture scenes with both a bright light source and deep shadows in one bracket, where static-scene methods previously required a locked-down camera.","feed_headline":"Event camera unlocks 12-stop HDR for dynamic scenes","feed_subtitle":"Five bracketed exposures plus event streams align frames where RGB-only flow fails, doubling practical HDR range.","key_machinery":"The load-bearing mechanism is the event-assisted explicit alignment module paired with a diffusion-based fusion module. Events are triggered when log-intensity change exceeds a contrast threshold, and because they encode change rather than absolute radiance, they stay informative in regions that are over- or underexposed in the RGB frames. The alignment module extracts features from the reference and moving LDR frames, initializes optical flow with RAFT, and then refines that flow using the event stream between the two timestamps before warping the moving frame into the reference geometry. The fusion stage feeds the aligned LDRs through a ControlNet, a trainable copy of the encoder that injects conditioning into a frozen latent diffusion model, which denoises a latent of the tone-mapped HDR candidate; a small color-correction layer is added so the output matches ground-truth color. Together the two stages convert a five-frame -6 EV to +6 EV bracket into a single HDR image intended to be free of ghosting artifacts.","core_discovery":"The paper's central claim is that 12-stop HDR reconstruction of a dynamic scene becomes tractable when the RGB bracket is accompanied by an event stream. Events record log-intensity changes at microsecond resolution across roughly 140 dB, so they retain the same motion information in a -6 EV frame and a +6 EV frame; the authors exploit this by estimating optical flow explicitly between LDR frames from event streams, rather than letting a network discover alignment implicitly. They then fuse the aligned frames with a latent diffusion model conditioned on the aligned LDRs, so pretrained natural-image priors fill in compressed details and absorb residual alignment errors. To support the claim they build the ESHDR simulated dataset, a synchronized dual-camera rig, and a real-world fine-tuning step for the alignment module, and report that the full pipeline outperforms previous RGB-based and event-based HDR methods on simulated fidelity metrics and on non-reference quality metrics for 16 real scenes.","pith_inferences":["Editorial inference: if the alignment module's real-world transfer is genuine, the same fine-tuning recipe should improve other event-assisted multi-frame tasks such as burst denoising and event-based frame interpolation; measuring transfer there would isolate the alignment gain from the diffusion fusion gain.","Editorial inference: because the real-world evaluation uses only non-reference metrics, the paper leaves open the possibility that the reported gains reflect perceptual smoothness rather than radiometric accuracy; a comparison against a calibrated high-dynamic-range reference would settle that.","Editorial inference: the explicit event-flow design implies that a simpler pipeline, event-based flow plus a strong conventional fusion network, might capture much of the alignment benefit; an ablation with a CNN fusion under reference-based metrics would reveal how much of the final gain comes from generative priors.","Editorial inference: the beam-splitter dual-camera hardware is the main practical obstacle, so the result points directly at hybrid event-RGB sensors; if those reach consumer cameras, 12-stop dynamic HDR could become a single-sensor feature."],"forward_implications":["If the central claim holds, the practical capture range for hand-held dynamic HDR rises from ±3 EV to ±6 EV with a five-frame bracket, so scenes containing both a bright light source and deep shadows can be recovered in one shot.","Event streams become a standard component of HDR pipelines: any bracket with exposure differences large enough to defeat RGB-only flow can be aligned by events, which record motion in the log-intensity domain rather than in absolute brightness.","Pretrained diffusion priors are shown to serve a fidelity task rather than only generation: the fusion module uses them to repair compressed-detail artifacts and residual alignment errors in the HDR output.","The ESHDR dataset provides a first 12-stop event-assisted HDR benchmark, making it possible to train and compare methods on this harder setting rather than the previous 6-stop standard.","Fine-tuning on easier real-world interpolation data narrows the simulated-to-real event gap, suggesting the alignment module generalizes beyond purely synthetic training."],"supporting_citations":[{"why":"Prior event-based multi-bracket HDR method with implicit alignment; re-trained as the Event-HDR* baseline the paper must beat.","marker":"[38]"},{"why":"Event-to-image feature distillation approach for bracketed HDR; earlier implicit event/RGB fusion that the paper contrasts with explicit alignment.","marker":"[48]"},{"why":"HDRFlow, the RGB-only dynamic HDR baseline that fails when exposure differences exceed what RGB flow can align.","marker":"[58]"},{"why":"HDR-Transformer, a context-aware transformer baseline compared on both simulated and real data.","marker":"[31]"},{"why":"Debevec-Malik HDR radiance recovery, the 12-stop static-scene fusion baseline that ghosts under motion.","marker":"[10]"},{"why":"HDRV, the single-frame event-guided HDR reconstruction baseline re-trained as HDRV*.","marker":"[63]"},{"why":"Latent diffusion model whose pretrained image priors the fusion module uses to repair artifacts.","marker":"[44]"},{"why":"ControlNet conditioning that feeds aligned LDR features into the frozen diffusion UNet during fusion.","marker":"[64]"},{"why":"v2e event simulator used in the ESHDR pipeline to generate realistic event streams from HDR video.","marker":"[20]"},{"why":"Real-world RGB-event interpolation datasets used to fine-tune the alignment module and close the simulated-to-real event gap.","marker":"[8, 35]"}],"fun_headline_variants":["Event streams align brackets for 12-stop HDR","Doubling HDR range with event-guided alignment","12-stop HDR from events: no more ghosting","Event camera fixes motion in extreme HDR","Five brackets plus events reach 12 stops"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the 16 real-world scenes and the non-reference quality metrics used to validate the method genuinely reflect HDR fidelity; if those scenes are unrepresentative or the metrics reward smoothing over radiometric correctness, the claim that the method works in real dynamic scenes is not established.","fun_headline_variants_meta":{"raw":{"variants":["Event streams align brackets for 12-stop HDR","Doubling HDR range with event-guided alignment","12-stop HDR from events: no more ghosting","Event camera fixes motion in extreme HDR","Five brackets plus events reach 12 stops"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000143,"raw_usage":{"total_tokens":1174,"prompt_tokens":949,"completion_tokens":225,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":565,"completion_tokens_details":{"reasoning_tokens":153}},"tokens_in":565,"tokens_out":225,"duration_ms":2195,"temperature":1.0,"reasoning_tokens":153,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T11:59:01.475216+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Capture a dynamic 12-stop scene with the same dual-camera rig while also recording a static ground-truth HDR of the same scene, then compare per-pixel radiance in the tone-mapped domain; if the method's output deviates from the reference in bright or dark regions beyond alignment tolerance, the ghost-free 12-stop claim is refuted.","supporting_citations":[{"cited_title":"Multi-bracket high dynamic range imaging with event cameras","cited_arxiv_id":null,"evidence_quote":"Prior event-based multi-bracket HDR method with implicit alignment; re-trained as the Event-HDR* baseline the paper must beat."},{"cited_title":"HDR Reconstruction from Bracketed Exposures and Events","cited_arxiv_id":"2203.14825","evidence_quote":"Event-to-image feature distillation approach for bracketed HDR; earlier implicit event/RGB fusion that the paper contrasts with explicit alignment."},{"cited_title":"Hdrflow: Real-time hdr video reconstruction with large motions","cited_arxiv_id":null,"evidence_quote":"HDRFlow, the RGB-only dynamic HDR baseline that fails when exposure differences exceed what RGB flow can align."},{"cited_title":"Ghost-free high dynamic range imaging with context-aware transformer","cited_arxiv_id":null,"evidence_quote":"HDR-Transformer, a context-aware transformer baseline compared on both simulated and real data."},{"cited_title":"Recovering high dy- namic range radiance maps from photographs","cited_arxiv_id":null,"evidence_quote":"Debevec-Malik HDR radiance recovery, the 12-stop static-scene fusion baseline that ghosts under motion."},{"cited_title":"Learning event guided high dynamic range video re- construction","cited_arxiv_id":null,"evidence_quote":"HDRV, the single-frame event-guided HDR reconstruction baseline re-trained as HDRV*."},{"cited_title":"v2e: From video frames to realistic dvs events","cited_arxiv_id":null,"evidence_quote":"v2e event simulator used in the ESHDR pipeline to generate realistic event streams from HDR video."}],"review_version":1}