{"id":"f9abde1d-8d8d-452f-a0ee-b6545f393651","arxiv_id":"2508.10779","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"TriFlowSR injects low-resolution and reference-image features into a frozen diffusion model through extended attention, adds a Roma-based alignment step, and reports state-of-the-art perceptual RefSR scores including on its new Landmark-4K benchmark.","lead":"A super-resolution system that sharpens low-quality photos by explicitly matching and copying texture patterns from a second, high-quality reference photo of the same scene. It also introduces a small 4K landmark photo benchmark, reporting improved perceptual quality scores on that benchmark and two existing ones.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Full matching strategy is validated only on the authors' curated Landmark-4K; public wins likely reflect Patch-Ref Attention alone, leaving the UHD matching claim untested on uncurated references.","rationale":"The reader's weakest assumption is that Roma-based matching may fail under viewpoint/scale differences, and that the Landmark-4K curation removes exactly the cases needed to test it. I agree: the unique UHD component of the pipeline is the Reference Matching Strategy, yet the only experiments that can exercise it are on the authors' own benchmark, which was explicitly curated to remove significant viewpoint variation. The public CUFED5/WR-SR results, which are the strongest independent evidence, appear to use padding rather than the Roma warp/mask path, so they validate the attention module but not the matching strategy. The ablation on Landmark-4K further shows that the matching strategy's contribution over a simple reference branch is small, and that the matching map alone can hurt PSNR/SSIM. The implementation detail that the mask is actually the SISR output rather than a white image also muddies what the reference branch receives. These points do not overturn the conditional verdict, but they sharpen why the paper should remain conditional: the central claim is not yet established outside a curated, unreleased test set. The proposed WR-SR experiment is a single, public, decisive check of whether the matching strategy helps when references are genuinely misaligned.","tokens_in":16269,"tokens_out":12360,"duration_ms":154008,"concrete_test":"Run TriFlowSR on the public WR-SR benchmark with the Reference Matching Strategy (Roma warp + confidence mask) enabled, and compare against the currently reported padded/naive-alignment setting. If the matching-enabled variant does not improve PSNR/LPIPS/DISTS over the padded variant on these uncurated, misaligned references, the central claim that the matching strategy enables better reference utilization is not supported outside the curated Landmark-4K regime.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that the Roma-based Reference Matching Strategy (Eqs. 3-4) lets the model better exploit the reference in UHD landmark RefSR. However, the only experiments that can exercise this strategy are on Landmark-4K, a 185-image set the authors curated by deleting images with significant viewpoint variation and blur. The CUFED5/WR-SR evaluations—where references are genuinely misaligned—are described in the Appendix as 'padded to achieve alignment,' not run through the 560×560 Roma warp/confidence-mask path; they therefore support Patch-Ref Attention but not the UHD matching strategy. On Landmark-4K, the matching strategy adds only ~0.2 dB PSNR over a plain reference branch, and the matching map alone (without warping) lowers PSNR/SSIM, so the claimed benefit is marginal and confined to an easy-alignment regime. Since the strategy is specifically motivated by scale/perspective differences between LR and reference, the evaluation has removed the very condition it is designed for. The implementation also substitutes the SISR output for the white mask shown in Figure 3, further complicating interpretation of what the reference branch receives.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes TriFlowSR, a reference-based super-resolution (RefSR) method built on a frozen Stable Diffusion 3 backbone. A trainable Reference HR branch is integrated through Patch-Ref Attention, where the attention keys/values of the SR, LR, and reference branches are concatenated (Eq. 1), enabling patch-level matching between LR and reference features. For ultra-high-definition (UHD) inference, a Reference Matching Strategy (Eqs. 3-4) uses a pre-trained dense matching model (Roma) at 560x560 resolution to warp the reference HR image to the upscaled LR image and suppress low-confidence regions. The authors also introduce Landmark-4K, a self-curated dataset of 185 UHD landmark images in 49 categories. Experiments report improved LPIPS/FID/DISTS on CUFED5 and WR-SR (Table 2) and improved PSNR/SSIM/LPIPS/DISTS on Landmark-4K (Table 3), leading to the claim that TriFlowSR is the first diffusion-based RefSR pipeline for UHD landmark scenarios under real-world degradation.","tokens_in":16456,"tokens_out":4835,"duration_ms":60298,"significance":"If fully supported, the paper would make a useful contribution: it demonstrates competitive perceptual performance on public RefSR benchmarks, introduces a new UHD RefSR dataset, and proposes a relatively simple, decoupled way to inject reference information into a frozen diffusion model. The public-benchmark perceptual wins (especially LPIPS, FID, DISTS on CUFED5 and WR-SR) are concrete evidence that Patch-Ref Attention helps. The promised code and model would also aid reproducibility. However, the paper's most distinctive component, the Roma-based Reference Matching Strategy for UHD inference, is only evaluated on Landmark-4K, a small, self-curated dataset from which the hard cases (significant viewpoint variation) have been explicitly removed. The current evidence therefore supports Patch-Ref Attention more strongly than it supports the full UHD matching strategy.","major_comments":[{"comment":"The central claim that the Reference Matching Strategy (Eqs. 3-4) effectively handles scale/perspective differences between LR and reference is only exercised on Landmark-4K, a 185-image benchmark the authors curated by removing images with 'significant viewpoint variations'. On CUFED5 and WR-SR, where references are naturally misaligned, the Appendix states that the images are 'padded to achieve alignment' rather than run through the 560x560 Roma warp/confidence-mask path. The matching strategy is therefore untested in exactly the regime it is designed for. I request either an evaluation on uncurated or deliberately misaligned references, or a restriction of the claim to the curated UHD setting.","section":"Reference Matching Strategy / Landmark-4K Dataset"},{"comment":"The ablation does not cleanly support the claim that 'Introducing the Reference Matching Strategy can solve the problem of incorrect tile retrieval'. Adding the Matching map without Warping (RB+M) decreases PSNR/SSIM relative to RB alone (25.37/0.750 vs 25.67/0.756) and gives only marginal LPIPS/DISTS improvement (0.239/0.110 vs 0.241/0.113). The best result appears only after Warping is also added. Thus the individual contribution of M is negative on two of four metrics, and the benefit is attributable to M+W jointly. Because the table reports no error bars or statistical tests, the 0.2 dB difference between the best configuration and RB alone may not be significant on 49-136 test images.","section":"Table 4 / Ablation Study"},{"comment":"There is a direct inconsistency between the main text and the appendix about the confidence mask. Figure 3 and the surrounding text say the mask used to suppress low-confidence regions is 'a pure white image', but the Appendix states: 'In practical implementation, since the mask can be any available image, we will use the result of the SISR model as the actual mask used.' This is not a cosmetic detail: it changes the input to the reference branch and affects the interpretation of Eq. 4 and the ablations. The authors must state which variant was used in Tables 3-5 and justify the discrepancy.","section":"Figure 3 / Appendix 'Detail Experimental Setting'"},{"comment":"No error bars, confidence intervals, or paired statistical tests are provided for any table. This is especially important on Landmark-4K (Table 3), where the test set is small and many comparisons are within a few tenths of a dB. It is also relevant to the FID numbers on CUFED5/WR-SR, which are known to be unstable with small sample sizes. Please report per-image scores with variance or at least a paired comparison for the main claims.","section":"All quantitative tables"}],"minor_comments":[{"comment":"The notation 'P atchRef Attn' has a formatting artifact; also Eq. (1) omits layer indexing, making it unclear whether Patch-Ref Attention is applied at every decoder layer or only some layers.","section":"Eq. (1)"},{"comment":"The caption refers to 'TriFlow' while the method is named 'TriFlowSR'. Please make the nomenclature consistent.","section":"Figure 1 caption"},{"comment":"The Roma matching model is mentioned only by citation. Please specify the exact checkpoint, the resolution at which matching is computed, and any confidence thresholding applied to C before the mask is formed. The choice of 560x560 also appears arbitrary and should be justified.","section":"Implementation Details / Reference Matching Strategy"},{"comment":"The test protocol is under-specified. The text says one image per category is used for self-reference and the remaining images for cross-reference, but it is not stated how Table 3 is computed: how many images are in the test set, whether self- and cross-reference results are pooled or reported separately, and how the reference for each test image is selected.","section":"Landmark-4K Dataset"},{"comment":"The resource-efficiency table is missing TFLOPs for CoSeR, and no details are given about the input resolution, batch size, or number of diffusion steps used for each method. These affect both runtime and FLOPs.","section":"Table 11"},{"comment":"Several appendix tables duplicate main-text tables (e.g., Tables 4/7 and 5/8). This is not a technical issue but adds length without new information.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The paper is best viewed as an empirical systems contribution rather than a theoretical one. The public-benchmark perceptual wins give it credible value. My main concern, as reflected in the major comments, is that the signature Reference Matching Strategy is validated only on a small, self-curated, easy-alignment benchmark, while the public benchmarks that would test its robustness are processed without this strategy. The mask inconsistency between Figure 3 and the Appendix also needs to be resolved before the paper is publishable. I see no reason to suspect misconduct, but the evaluation section should be substantially strengthened or its claims narrowed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a solid, incremental RefSR paper that deserves a serious referee rather than a desk reject. The genuinely new pieces are Patch-Ref Attention, which concatenates LR and reference keys/values into a frozen SD3 cross-attention, and the Landmark-4K dataset. The public-benchmark wins on LPIPS/FID/DISTS on CUFED5 and WR-SR are real evidence that the attention design helps perceptually, and the reported FLOPs/time are a plus. If code, weights, and the dataset ship as promised, this becomes a reproducible contribution.\n\nWhat the paper does well: the recipe is simple and easy to build on. Freeze SD3, train an LR branch first, then a reference branch, and inject both through concatenated cross-attention. That plausibly explains the perceptual improvements. The public evaluation is standard, and the visual results in the figures are believable. The Landmark-4K dataset is small (185 images, 49 categories) but it is a first step toward evaluating RefSR at true UHD resolution.\n\nThe soft spots line up with the stress-test note. The Reference Matching Strategy is only exercised on Landmark-4K. On CUFED5 and WR-SR the authors pad LR and reference to the same size rather than running the Roma warp/confidence-mask path, so those wins are about Patch-Ref Attention, not the matching strategy. On Landmark-4K, warping adds only ~0.2 dB over the plain reference branch, and the matching map without warping actually lowers PSNR/SSIM. And because the dataset explicitly removes images with significant viewpoint variation, the exact regime the strategy is designed for is absent. The claim that the full pipeline 'better utilizes' the reference is therefore narrower than the abstract suggests.\n\nTwo issues deserve explicit referee attention. First, the appendix says the white mask shown in Figure 3 is implemented as the SISR output. That changes the method: the confidence gate is actually a weighted blend between the warped reference and SISR content, not a mask-out to white. This needs to be disclosed and ablated. Second, all tables are single-run without error bars, and the kscale sweep jumps from 24.79 at 0.8 to 25.89 at 1.0, which is enough of a discontinuity to make me want a significance test. The default settings (kscale=1, 23 layers) are chosen via ablations on the same Landmark-4K set used for the headline comparison, which is common but still inflates the headline number.\n\nNone of this is a load-bearing flaw. The core mechanism is plausible and the public evidence supports it. But the strong form of the claim, about UHD misalignment, is untested. Recommendation: send to peer review. Ask for code/data, error bars, a test set with real viewpoint differences, and an honest description of the mask. I'd bring it to a reading group if super-resolution is on the agenda, but I wouldn't cite it as a benchmark until the dataset and code are actually released.","headline":"Incremental but useful RefSR paper with a credible diffusion recipe; the matching strategy is only tested on a small curated set and the mask substitution needs disclosure.","tokens_in":17098,"tokens_out":4147,"would_cite":false,"duration_ms":42559,"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":"TriFlowSR claims explicit patch-scale matching between low-resolution and reference images lets diffusion-based super-resolution transfer reference texture faithfully, setting the best reported perceptual scores on UHD landmark images.","keywords":["reference-based super-resolution","diffusion models","patch-scale attention","image matching","ultra-high-definition","landmark images","real-world degradation","RefSR dataset"],"falsifier":"Take a landmark pair with a large viewpoint change (for instance, a facade shot from 30 degrees apart, which Landmark-4K explicitly filters out), run TriFlowSR's Reference Matching Strategy, and inspect whether the warped reference aligns with the LR. If perceptual metrics (LPIPS, DISTS) on such pairs fall back to the SISR-only baseline while the SISR branch alone holds, the matching assumption is the bottleneck; visual misalignment in the masked warped reference would confirm it.","tokens_in":16048,"feed_emoji":"🖼️","tokens_out":4495,"duration_ms":41958,"temperature":0.7,"pith_summary":"This paper tries to establish that reference-based image super-resolution can be made to work at ultra-high-definition under real-world degradation by explicitly matching the low-resolution and reference high-resolution images at the patch scale, instead of injecting reference features through ControlNet-style additions. To test this, the authors build TriFlowSR, a three-branch diffusion pipeline with a frozen Stable Diffusion 3 backbone that receives extra keys and values from a reference branch through concatenated cross-attention, and they introduce Landmark-4K, a 185-image UHD landmark benchmark. Reported metrics place TriFlowSR best on perceptual and distribution measures (LPIPS, FID, DISTS) on CUFED5 and WR-SR, and best on PSNR, SSIM, LPIPS, DISTS on Landmark-4K. If the claim holds, the practical payoff is that UHD photos of landmarks can be restored with textures drawn from a few reference shots of the same site rather than hallucinated by the generative model.","feed_headline":"Patch-matched diffusion tops 4K landmark super-resolution","feed_subtitle":"Explicit patch-scale matching plus a warped reference beat ControlNet-style RefSR on perceptual metrics.","key_machinery":"Patch-Ref Attention: cross-attention where the SR branch's query attends over a concatenation [K_SR, K_LR, K_Ref] and [V_SR, V_LR, V_Ref], so each output patch is explicitly matched to reference patches at the same feature scale. The Reference Matching Strategy is the companion mechanism for UHD: it runs the pretrained matcher at 560×560, upscales the mapping and confidence map, warps the reference HR into the LR frame, and gates low-confidence regions with a mask (in practice the SISR output). Together they feed spatially consistent reference texture to the frozen diffusion backbone.","core_discovery":"TriFlowSR's central claim is that explicit pattern matching between the low-resolution image and the reference HR image at the patch scale is what lets a diffusion-based RefSR pipeline transfer semantic and texture information faithfully. The mechanism is Patch-Ref Attention, a cross-attention operation in which the SR branch's query attends to the concatenation of keys (and values) from the SR, LR, and reference branches, so reference features are selected by their similarity to the target patches rather than by a coarse global alignment. For UHD inference the paper adds a Reference Matching Strategy: a pretrained matcher (Roma) establishes dense correspondences at 560×560 between the bicub","pith_inferences":["Because Landmark-4K curation explicitly removes images with 'significant viewpoint variations', the claimed advantage over ControlNet-style methods is demonstrated only for near-frontal landmark views; robustness to large viewpoint or scale changes remains untested.","The pipeline inherits the limitations of the pretrained matcher Roma; swapping in a stronger dense matcher should directly raise the ceiling of the Reference Matching Strategy, and the confidence map could be treated as an explicit uncertainty signal rather than a binary gate.","The kscale and layer-count controls suggest a practical product feature, per-region user control over how much reference texture is trusted, which the paper does not explore.","Using the SISR output as the mask means the pipeline already has a natural fallback; comparing a white mask against the SISR mask would isolate how much of the gain comes from the warping versus the masking."],"forward_implications":["Diffusion-based RefSR need not rely on ControlNet-style injection; explicit patch-scale matching gives better perceptual fidelity on the tested benchmarks.","Tile-based UHD inference is feasible when the reference is warped and masked before each tile is processed, avoiding incorrect tile retrieval due to scale or position mismatch.","The kscale knob makes the model interpolate continuously between SISR and RefSR behavior, offering a single checkpoint that can serve both tasks.","Landmark-4K provides a UHD benchmark whose reference images contain fine-grained detail, making it possible to measure whether diffusion RefSR actually uses reference texture.","The ablations show each component matters: the reference branch, the matching map, and the warping each increase PSNR, SSIM, LPIPS, or DISTS on Landmark-4K."],"supporting_citations":[{"why":"Supplies Roma, the pretrained dense matcher that produces the coarse correspondence and confidence map used by the Reference Matching Strategy.","marker":"(Edstedt et al. 2024)"},{"why":"ReFIR, the primary diffusion-based RefSR baseline that injects reference information via ControlNet-style gating and is outperformed on perceptual metrics.","marker":"(Guo et al. 2024)"},{"why":"CoSeR, another diffusion RefSR baseline using separate ControlNets for LR and reference, compared against in the experiments.","marker":"(Sun et al. 2024)"},{"why":"StableSR, whose Real-ESRGAN degradation configuration is adopted to synthesize real-world degraded LR inputs, and a SISR baseline.","marker":"(Wang et al. 2024)"},{"why":"Stable Diffusion 3, the frozen rectified-flow diffusion backbone that serves as the SR branch of TriFlowSR.","marker":"(Esser et al. 2024)"},{"why":"C2-Matching, a classic CNN RefSR baseline and the source of the WR-SR dataset used for evaluation.","marker":"(Jiang et al. 2021)"},{"why":"CUFED5, the standard RefSR benchmark dataset used for evaluation.","marker":"(Zhang et al. 2019)"}],"fun_headline_variants":["Patch-level reference matching boosts diffusion SR to 4K","TriFlowSR: cross-attention to references for UHD landmarks","Explicit patch matching: key to diffusion-based 4K SR","First UHD landmark RefSR with diffusion and patch attention","Roma-matched references sharpen diffusion super-resolution"],"cache_read_input_tokens":2816,"weakest_assumption_plain":"Everything rests on the pretrained matcher Roma giving reliable dense correspondences between a bicubically upscaled, heavily degraded LR image and the reference HR at 560×560, and on the confidence map C correctly suppressing bad warps; the Landmark-4K curation removes large viewpoint differences, so that regime is not tested.","fun_headline_variants_meta":{"raw":{"variants":["Patch-level reference matching boosts diffusion SR to 4K","TriFlowSR: cross-attention to references for UHD landmarks","Explicit patch matching: key to diffusion-based 4K SR","First UHD landmark RefSR with diffusion and patch attention","Roma-matched references sharpen diffusion super-resolution"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00047,"raw_usage":{"total_tokens":2185,"prompt_tokens":760,"completion_tokens":1425,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":504,"completion_tokens_details":{"reasoning_tokens":1343}},"tokens_in":504,"tokens_out":1425,"duration_ms":12334,"temperature":1.0,"reasoning_tokens":1343,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T20:14:25.280435+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a landmark pair with a large viewpoint change (for instance, a facade shot from 30 degrees apart, which Landmark-4K explicitly filters out), run TriFlowSR's Reference Matching Strategy, and inspect whether the warped reference aligns with the LR. If perceptual metrics (LPIPS, DISTS) on such pairs fall back to the SISR-only baseline while the SISR branch alone holds, the matching assumption is the bottleneck; visual misalignment in the masked warped reference would confirm it.","supporting_citations":[{"cited_title":"o kman, G.; Wadenb \\","cited_arxiv_id":null,"evidence_quote":"Supplies Roma, the pretrained dense matcher that produces the coarse correspondence and confidence map used by the Reference Matching Strategy."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"ReFIR, the primary diffusion-based RefSR baseline that injects reference information via ControlNet-style gating and is outperformed on perceptual metrics."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"CoSeR, another diffusion RefSR baseline using separate ControlNets for LR and reference, compared against in the experiments."},{"cited_title":"C.; and Loy, C","cited_arxiv_id":null,"evidence_quote":"StableSR, whose Real-ESRGAN degradation configuration is adopted to synthesize real-world degraded LR inputs, and a SISR baseline."},{"cited_title":"C.; Wang, X.; Loy, C","cited_arxiv_id":null,"evidence_quote":"C2-Matching, a classic CNN RefSR baseline and the source of the WR-SR dataset used for evaluation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"CUFED5, the standard RefSR benchmark dataset used for evaluation."}],"review_version":1}