{"id":"545014b1-4896-48c7-8620-f4a6c32edf3e","arxiv_id":"2608.09373","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"FSP-Diff reduces content drift in one-step real-world image super-resolution by injecting structured details and modulating semantic guidance.","lead":"A new one-step diffusion model for real-world image super-resolution, called FSP-Diff, uses two parallel pathways to preserve fine details and correct text-based semantic errors. It reports improved fidelity and perceptual scores over prior one-step methods on standard benchmarks, which matters for practical image restoration at scale.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Main results and component ablations use different feature extractors, so the component ablation does not validate the claimed model's dual-pathway gains.","rationale":"The reader's conditional verdict already flags table inconsistencies and the absence of code, and I agree that conditional acceptance is the right level. My specific concern is narrower and more load-bearing than the reader's stated weakest assumption about ViT feature quality: the paper's own tables show that the component-wise ablation in Table 4 was run with the VAE-based extractor, while the headline model in Table 1 uses the ViT-based extractor. These are not interchangeable; Table 2 reports different FID, DISTS, and MANIQA for the two settings, and Table 4's 'Full' row numerically matches the VAE-extractor row rather than the headline model. As a result, the incremental gains attributed to SDCA, DATC, and SDM were not demonstrated for the architecture that is claimed to surpass prior one-step methods. This does not refute the method, but it means the central causal claim is currently supported only if one accepts an unverified transfer of ablations across extractors. The concrete test above would settle it directly: rerun the ablations with the final ViT extractor and check that the full configuration reproduces the Table 1 numbers. If it does, the concern is resolved and conditional acceptance is appropriate; if it does not, the headline claim should be downgraded or the configuration clarified.","tokens_in":16157,"tokens_out":8756,"duration_ms":82737,"concrete_test":"Run the Table 4 component ablations (Baseline, +SDCA, +DATC, +SDM, Full) on DRealSR using the exact Table 1 configuration: frozen ViT first-block extractor, L_q=400, s=0.9, and the same OSEDiff baseline. Require that the 'Full' row reproduces Table 1's FSP-Diff row (e.g., PSNR 28.85, FID 127.19) and that each added component shows the claimed incremental improvement. If 'Full' instead matches the VAE-extractor row of Table 2/Table 4 (FID 125.34, PSNR 28.93), the main comparison was not the configuration that was ablated, and the central dual-pathway claim remains unverified without release of code and checkpoints.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 4.3's component ablation does not test the configuration whose results are claimed. The text says the ablation uses the 'VAE+' extractor even though the final model in Tables 1-3 uses the 'ViT+' first-block extractor, asserting that this 'does not affect the conclusions.' That assertion is not supported by the reported numbers. Table 2's VAE+DATC_D4 row (FID 125.34, DISTS 0.2048, MANIQA 0.6053) matches Table 4's 'Full' row, while the headline FSP-Diff row in Table 1 (FID 127.19, DISTS 0.2014, MANIQA 0.6092) matches the ViT+DATC_D4 row. The two configurations are measurably different, yet Table 4 attributes the full-method gains to SDCA, DATC, SDM, and the shift using the VAE-based extractor. Because the ViT first-block extractor is the load-bearing source of 'structured details' and was itself selected via test-set comparisons, the causal claim that these components preserve fine structures and mitigate content drift in the final advertised model is not established by the experiments as reported.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes FSP-Diff, a one-step diffusion model for real-world image super-resolution (Real-ISR). It augments an OSEDiff-style Stable Diffusion 2.1 backbone with two pathways: a Detail-Conditioned Pathway that extracts structured details from frozen early ViT features, compresses them with a Detail-Aware Token Compression module, and injects them via Structured Detail Conditioned Attention; and a Detail-Modulated Semantic Pathway that refines text embeddings with a Semantic-Detail Modulator. Training uses L1+LPIPS reconstruction together with variational score distillation and a shifted timestep-sampling distribution. The model is evaluated on DIV2K-val, RealSR, and DRealSR against one-step diffusion baselines, and the paper reports ablations of components and hyperparameters as well as inference complexity.","tokens_in":16475,"tokens_out":7871,"duration_ms":68871,"significance":"If the claimed gains are reproducible, the core idea—using early-layer ViT features as a structured-detail complement to the lossy VAE latent and conditioning both detail injection and semantic modulation on those features—is a plausible and useful direction for one-step Real-ISR. The paper has strengths: it uses standard benchmarks and baselines, reports parameter/FLOP/latency comparisons, and makes an effort to match OSEDiff's training setup. However, the experimental validation as reported does not fully support the central claims: the headline model and the component ablations use different feature extractors, several hyperparameters are selected directly on the test benchmarks, and the abstract overstates the quantitative comparison.","major_comments":[{"comment":"The component ablation is run with the VAE+ extractor, while the headline FSP-Diff model in Tables 1–3 uses the ViT+ first-block extractor. The paper acknowledges this but asserts it 'does not affect the conclusions' without support. Table 2 shows the two extractors produce measurably different results: ViT+DATC_D4 gives DISTS 0.2014, MANIQA 0.6092, FID 127.19, and CLIPIQA 0.6868, whereas VAE+DATC_D4 gives DISTS 0.2048, MANIQA 0.6053, FID 125.34, and CLIPIQA 0.6725. Table 4's 'Full' row (PSNR 28.93, DISTS 0.2048, NIQE 6.21, MANIQA 0.6053, FID 125.34) matches the VAE-based configuration, not the ViT-based model whose results are advertised. The incremental gains attributed to SDCA, DATC, and SDM in Table 4 are of the same order as the differences caused by the extractor choice, so the component-wise attribution for the final model is not established. The ablations should be rerun with the ViT+ extractor, or the authors should provide direct evidence that the extractor choice does not change the conclusions.","section":"Section 4.3, Tables 2 and 4"},{"comment":"Several key hyperparameters are selected using the same benchmark test sets that are later reported as the main results. Table 3 selects ViT depth on DRealSR, Table 5 selects L_q (presumably on DRealSR, though the table is not labeled), and Table 6 selects the shift parameter on DIV2K according to the caption but on DRealSR according to the main text. Since the final configuration is chosen by comparing directly on these benchmarks, the numbers in Table 1 are not held-out evaluations, and the reported improvements are inflated by test-set optimizer's curse. The authors should use a separate validation split or nested evaluation procedure for hyperparameter selection and then report unbiased test-set numbers.","section":"Section 4.3, Tables 3, 5, and 6"},{"comment":"The abstract claims that FSP-Diff 'surpasses existing one-step diffusion methods in both quantitative and qualitative metrics,' but Table 1 contradicts this on several no-reference metrics. TSD-SR outperforms FSP-Diff on CLIPIQA on all three benchmarks (0.7344 vs 0.6868 on DRealSR, 0.7160 vs 0.6687 on RealSR, 0.7416 vs 0.6583 on DIV2K-Val) and on MUSIQ on all three (66.62 vs 65.17, 71.19 vs 68.60, 71.69 vs 67.78). TSD-SR also has better NIQE on RealSR and DIV2K-Val, and PiSA-SR has better MANIQA on RealSR and DIV2K-Val. The text in Section 4.2 partially acknowledges this, but the abstract and contribution statements should be revised to claim competitiveness or superiority only on the specific metrics and benchmarks where it holds.","section":"Abstract and Table 1"},{"comment":"The table shows that 'Full' (with shift s=0.9) has worse PSNR than 'Add SDM' (s=1.0) by 0.13 dB and worse FID by 0.79, while improving only NIQE and MANIQA by small margins. The manuscript states that 'Full' is the best configuration, but it is not supported by the fidelity metrics reported in the same table. This is a specific internal inconsistency in the ablation story for the shift parameter. The authors should report the trade-off explicitly and justify why the perceptual gains outweigh the fidelity loss, or adjust the claimed best configuration.","section":"Table 4 and Section 4.3"}],"minor_comments":[{"comment":"The text says the DIV2K-val dataset contains 3,000 LR–HR image pairs, but the standard DIV2K validation set contains only 100 images; please clarify whether this is a different synthesized validation set of crops.","section":"Section 4.1"},{"comment":"The caption of Table 6 says 'DIV2K Benchmark,' while the main text says the shift hyperparameter is evaluated on the 'DRealSR test set.' These need to be reconciled.","section":"Section 4.3 and Table 6"},{"comment":"The table does not state which benchmark is used for the L_q ablation; it should be labeled for reproducibility.","section":"Table 5"},{"comment":"The 'vit_b' feature extractor is not precisely specified; please state which pretrained ViT checkpoint (e.g., CLIP ViT-B/32, DINOv2, or another model) is used, how the input is preprocessed, and at which spatial resolution the features are taken.","section":"Section 3.3"},{"comment":"There are minor typos: 'we adopt adopt this value' and 'experimemts'; also Table 2's VAE row is missing spaces between numeric entries.","section":"Section 4.3"}],"recommendation":"major_revision","confidential_remarks":"The proposed architecture is plausible and the paper contains useful experimental infrastructure, but the current validation is not sufficient: the component ablations are run with a different feature extractor than the final model, hyperparameters are tuned on the test benchmarks, and the abstract overclaims on several no-reference metrics. These issues are fixable with additional experiments and a more careful evaluation protocol, but they are load-bearing for the central claims. I would not recommend acceptance without those revisions."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's the short version: this is a competent, useful increment in one-step real-ISR. The dual-pathway design - pulling fine structures from early ViT features and using those same structures to modulate the text semantics - makes sense, and the experimental coverage is thorough. The main problems are a gap between the ablation setup and the final model, and an abstract that oversells the numbers.\n\nWhat's actually new: the paper reframes the failure mode of one-step diffusion SR as 'content drift' with two distinct flavors, visual detail degradation and text-semantic shift. That's a helpful way to organize the problem. The specific modules (DATC, SDCA, SDM) are variations on attention-based feature injection, but the combination in a dual-pathway is not something I've seen in the cited one-step baselines. The comparisons are against the right methods: SinSR, OSEDiff, TSD-SR, PiSA-SR, TVT. And the gains on PSNR, SSIM, LPIPS, DISTS, and FID are consistent across all three benchmarks - that's real evidence, even if the no-reference metrics are mixed.\n\nThe soft spots, in order of importance:\n\n1. The ablation doesn't validate the final model. Section 4.3 states the component ablations use the VAE-based extractor, not the ViT first-block extractor that the final model actually uses. Table 2 shows those two extractor choices give measurably different results (DISTS 0.2014 vs 0.2048, MANIQA 0.6092 vs 0.6053, FID 127.19 vs 125.34). So when Table 4 attributes the improvements to SDCA, DATC, and SDM, it's doing so under the VAE extractor, while the headline results in Table 1 come from the ViT extractor. The authors wave this off with 'this choice does not affect the conclusions,' but they don't show that. That's a referee-fixable problem, but a real one.\n\n2. The abstract overclaims. Table 1 shows TSD-SR and PiSA-SR beat FSP-Diff on CLIPIQA, MUSIQ, NIQE, and MANIQA on several benchmarks. The body text is more careful, noting where competitors do better. The abstract should be toned down.\n\n3. Test-set hyperparameter selection. The shift s, query count Lq, and ViT depth are chosen by evaluating on the test sets. That inflates the reported numbers and there are no error bars or repeated runs. This is common in the field, but it's still a limitation.\n\n4. Minor inconsistencies: Table 6's caption says DIV2K but the text says DRealSR; Table 4's 'Full' has worse PSNR and FID than 'Add SDM'. Small, but messy.\n\nWho this is for: researchers working on one-step diffusion super-resolution will get value from this paper. It's not a breakthrough, but it's a concrete, well-scoped improvement with honest comparison against a strong baseline. The central idea holds up. I'd send it to peer review - with a request to rerun the ablations with the final extractor and to fix the abstract.","headline":"Solid incremental one-step SR paper with a good idea; the ablation doesn't match the final model configuration and the abstract overclaims, but it's worth refereeing.","tokens_in":16963,"tokens_out":4482,"would_cite":true,"duration_ms":37970,"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":"A dual-pathway one-step diffusion model preserves fine structures and corrects semantic drift in real-world super-resolution.","keywords":["real-world image super-resolution","one-step diffusion","content drift","structured details","semantic guidance","Vision Transformer features","detail-aware token compression","structured detail conditioned attention"],"falsifier":"A controlled experiment that replaces the ViT feature map with random noise of the same shape, while keeping DATC and SDCA intact, would settle the claim: if the noise-injected model still matches FSP-Diff's PSNR and LPIPS scores on any of the three benchmarks, then the structured details themselves are not the cause of the improvement.","tokens_in":1813,"feed_emoji":"🔍","tokens_out":1977,"duration_ms":74119,"temperature":0.7,"pith_summary":"Real-world image super-resolution must reconstruct sharp images from low-quality inputs, but diffusion-based methods suffer content drift: the compressed latent discards fine structures, and captions generated from unclear images can shift the semantics (a house rendered as a snow pile). This paper proposes FSP-Diff, a one-step diffusion model whose dual-pathway architecture counteracts both failure modes. One pathway extracts structured details from the first block of a frozen Vision Transformer and injects them into the diffusion network; the other uses the same details to modulate the text embeddings so the semantic guidance stays faithful to the input. Benchmarks on DIV2K-val, RealSR, and DRealSR show the model outperforming existing one-step diffusion methods on most fidelity and perceptual metrics. The paper claims that preserving fine-grained structure from the low-quality input, both as pixel-level guidance and as a check on the text prompt, is what resolves content drift.","feed_headline":"One-step diffusion model beats content drift by preserving fine details","feed_subtitle":"A frozen early ViT preserves structures the VAE loses; those details also refine the text guidance.","key_machinery":"The load-bearing mechanism is the dual-pathway architecture built on a pre-trained Stable Diffusion backbone: (1) the Detail-Conditioned Pathway extracts a structured-detail feature map $X_S$ from the first block of a frozen ViT encoder, compresses it to a small token set $X_H$ with a learnable Detail-Aware Token Compression (DATC) transformer, and injects it into the UNet through Structured Detail Conditioned Attention (SDCA), in which the diffusion latents query the detail tokens; (2) the Detail-Modulated Semantic Pathway passes the text embeddings and $X_S$ through a Semantic-Detail Modulator (SDM), whose scaled dot-product attention produces a detail-aware text conditioning $C_T$. DATC keeps the attention cost tractable, SDCA feeds fine structure directly into generation, and SDM lets the details correct semantic misalignment in the caption.","core_discovery":"The central discovery claimed is that content drift in one-step diffusion-based real-world super-resolution arises from the loss of structured details in the VAE-compressed latent and from the unreliability of text captions derived from low-quality images, and that both can be mitigated by explicitly extracting fine structures from the input and using them in two complementary ways. FSP-Diff's Detail-Conditioned Pathway uses a frozen first-block ViT feature map at one-eighth resolution as structured detail, compresses it through a Detail-Aware Token Compression module, and injects the resulting tokens into each UNet block through Structured Detail Conditioned Attention placed after the text cross-attention. The Detail-Modulated Semantic Pathway computes attention between text embeddings and the structured detail map, producing a refined text conditioning that replaces the original caption embedding. Trained with L1 and LPIPS reconstruction losses plus variational score distillation regularization and a shifted timestep distribution, the model attains the best or second-best results on most benchmark metrics and, in the paper's qualitative comparisons, avoids both the stochastic artifacts of the baseline and the polygonal artifacts of a fine-structure-preserving competitor.","pith_inferences":["The dual-pathway principle may transfer to other diffusion-based restoration tasks, such as deblurring or deraining, where the input carries high-frequency structure that the latent drops and where text prompts from degraded images are unreliable.","The success of the frozen first-block ViT features suggests a general recipe: bypass the VAE's lossy compression with an off-the-shelf high-resolution feature extractor, then compress adaptively before injection.","A testable extension is to generate the semantic conditioning directly from the structured-detail tokens, which could remove the dependency on the captioning module entirely.","The ViT-depth ablation implies that only the earliest layer of the encoder matters for this purpose, so a lightweight edge-and-boundary extractor might achieve similar gains at lower computational cost."],"forward_implications":["One-step real-world super-resolution can preserve fine structures without retraining a less-compressed VAE, avoiding the multi-stage pipeline used by methods that transfer the VAE to a lower compression ratio.","Adding the two pathways to an existing one-step diffusion baseline raises fidelity and perceptual metrics together, suggesting that detail injection and semantic modulation do not force the usual fidelity-perception trade-off.","Because the structured-detail extractor is frozen and the token set is small, the dual-pathway design adds only modest inference cost, about 0.15 seconds versus 0.12 seconds for the baseline on the reported hardware.","The semantic pathway uses the same detail features as the detail pathway, so semantic mis-reconstructions such as rendering a house as a snow pile are corrected without introducing an extra captioning model beyond the existing prompt extractor.","The design is compatible with the one-step latent diffusion formulation and can be trained in a single stage on the same training data as the baseline."],"supporting_citations":[{"why":"Provides the one-step diffusion baseline, VSD regularization objective, and LoRA fine-tuning setup that FSP-Diff extends with its two pathways.","marker":"[49]"},{"why":"Supplies the frozen ViT encoder whose first block yields the structured-detail feature map used throughout the architecture.","marker":"[12]"},{"why":"Provides the pre-trained Stable Diffusion backbone and VAE whose compressed latent is the source of the detail loss FSP-Diff addresses.","marker":"[34]"},{"why":"Supplies the DAPE prompt-extraction module that generates the initial text guidance for each low-quality image.","marker":"[50]"},{"why":"Is the fine-structure-preserving one-step competitor whose multi-stage VAE transfer and artifacts FSP-Diff is compared against.","marker":"[55]"},{"why":"Is the one-step diffusion competitor with target score distillation used as a benchmark for quantitative and qualitative comparison.","marker":"[11]"},{"why":"Provides the Real-ESRGAN degradation pipeline that generates the low-quality/high-quality training pairs.","marker":"[44]"},{"why":"Supplies the low-rank adaptation method used to fine-tune the diffusion network and VAE encoder.","marker":"[17]"}],"fun_headline_variants":["Frozen ViT details mitigate content drift in real-world super-resolution","Dual-path diffusion uses frozen ViT details to stop drift","FSP-Diff: Structured details mitigate content drift in image upscaling","Preserve fine details to stop content drift in super-resolution"],"cache_read_input_tokens":19072,"weakest_assumption_plain":"The whole scheme rests on the assumption that the first block of a frozen ViT, applied to the low-quality image, retains the fine structures that the VAE latent has discarded, and that the DATC and SDCA modules can feed those structures into the generator without introducing their own artifacts.","fun_headline_variants_meta":{"raw":{"variants":["Frozen ViT details mitigate content drift in real-world super-resolution","Dual-path diffusion uses frozen ViT details to stop drift","FSP-Diff: Structured details mitigate content drift in image upscaling","Preserve fine details to stop content drift in super-resolution"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000877,"raw_usage":{"total_tokens":3791,"prompt_tokens":941,"completion_tokens":2850,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":557,"completion_tokens_details":{"reasoning_tokens":2788}},"tokens_in":557,"tokens_out":2850,"duration_ms":19422,"temperature":1.0,"reasoning_tokens":2788,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T18:29:20.272325+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A controlled experiment that replaces the ViT feature map with random noise of the same shape, while keeping DATC and SDCA intact, would settle the claim: if the noise-injected model still matches FSP-Diff's PSNR and LPIPS scores on any of the three benchmarks, then the structured details themselves are not the cause of the improvement.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the one-step diffusion baseline, VSD regularization objective, and LoRA fine-tuning setup that FSP-Diff extends with its two pathways."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Is the one-step diffusion competitor with target score distillation used as a benchmark for quantitative and qualitative comparison."}],"review_version":1}