{"id":"6555bb31-e5cd-4766-9470-0988e4a3d7d9","arxiv_id":"2608.09133","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"By injecting pre-VAE pixel features into both the latent denoising trajectory and the frozen VAE decoder, PGSR improves fidelity of latent diffusion transformer super-resolution while maintaining perceptual quality.","lead":"This paper proposes PGSR, a super-resolution method that keeps the original low-resolution pixel information separate from the compressed latent representation and injects it at two points of a pretrained diffusion transformer. The result is a claimed better trade-off between perceptual quality and faithfulness to the input image, which matters because current latent generative SR models tend to hallucinate details.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The decoder-side ablation does not isolate the DFM architecture from the low-noise image-space losses, so the dual-stage claim is not yet established.","rationale":"The idea is plausible and the condition-side pathway is reasonably supported: fusing pre-VAE pixel evidence into the ControlNet condition is a natural way to anchor latent trajectories, and the Table 2 condition-side ablation is consistent with that. The concern I identify is narrower but load-bearing for the second half of the dual-stage claim: the paper does not currently separate the architectural effect of decoder-side DFM from the effect of adding low-noise image-space losses. Since both are introduced together in the full model, the reported gains could be driven by supervision, not by the multi-scale pixel taps. This is an experimental-identification issue rather than an internal inconsistency in the equations, though the Table 2/Table 5 PSNR discrepancy makes the quantitative support harder to trust. I do not see grounds to reject the paper outright; the method may well work as described, but the decoder-side claim needs the controlled ablation above before it can be accepted. The reader's conditional verdict is therefore unchanged.","tokens_in":15562,"tokens_out":9123,"duration_ms":89253,"concrete_test":"Run a four-variant ablation on DIV2K-Val with identical training budget and fixed condition-side guidance: (1) frozen decoder, L_fm only; (2) frozen decoder, L_fm + L_perc + L_pg (losses computed on the standard VAE decoder output); (3) DFM enabled, L_fm only; (4) DFM enabled, L_fm + L_perc + L_pg (full PGSR). Comparing (2) vs (4) isolates the DFM architecture while holding the loss terms constant; comparing (3) vs (4) isolates the image-space losses. If variant (2) matches (4) within noise, then the decoder-side injection contributes negligibly and the central dual-stage mechanism is unsupported. Report mean and standard deviation over at least three training seeds.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim requires showing that decoder-side pixel grounding itself improves fidelity. In Section 3.3, the full objective is L = L_fm + λ_perc·L_perc + λ_pg·L_pg, with image-space terms computed on the DFM-decoded output at low noise. Table 2 ablates 'w/o Decoder-Side Grounding' by removing the DFM branch, but the paper does not state whether the same L_perc/L_pg terms are still applied to the standard frozen-decoder output. Table 5, which is meant to separate supervision from architecture, adds the L_pg loss and the DFM branch simultaneously in its final row; there is no row with L_fm + L_perc + L_pg but no DFM. Thus the 1.11 dB PSNR and 0.0084 LPIPS gain attributed to decoder-side grounding in Table 2 could be partly or wholly due to the additional low-noise pixel-reconstruction and perceptual losses, not to the multi-scale pixel injection. An internal inconsistency compounds this: Table 2's 'w/o Pixel-Grounded Guidance' (PSNR 23.68, LPIPS 0.2431) and Table 5's 'FM control only' (PSNR 24.67, LPIPS 0.2431) appear to describe the same no-pixel-guidance configuration but disagree by ~1 dB PSNR, undermining confidence that the ablation tables are mutually calibrated.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes PGSR, a pixel-grounded super-resolution framework for latent diffusion transformers. The central idea is that the VAE compression bottleneck weakens fidelity-critical pixel evidence, so PGSR extracts pre-VAE pixel features from the upsampled LR image and reuses them at two stages: condition-side trajectory guidance, which fuses pixel evidence into the ControlNet condition with a gated residual term, and decoder-side pixel grounding, which injects multi-scale pixel features into the frozen VAE decoder through zero-initialized adapters. The FLUX.1-dev backbone and VAE remain frozen, with only lightweight modules trained alongside low-noise image-space losses. Experiments on DIV2K-Val, RealSR, and DRealSR report improved distortion and perceptual metrics over several generative SR baselines, and ablations attribute gains to the two pixel-grounded pathways.","tokens_in":15905,"tokens_out":4482,"duration_ms":41363,"significance":"If the central claim is supported, the paper makes a useful conceptual point: fidelity loss in latent generative SR is not only a conditioning-strength problem but also a consequence of the VAE information bottleneck, and pre-VAE pixel evidence can be reintroduced without retraining the backbone. The architecture is parameter-efficient and clearly specified, and the appendix contributes consistency arguments (gated fusion, multi-scale expressiveness, zero-initialized DFM identity) rather than overstated optimality guarantees. The main results are benchmark-based and reproducible in principle, with implementation details and ablation tables supplied. However, the significance is currently conditional because the decoder-side ablation does not isolate the architecture from the low-noise image-space losses, and because checkpoint selection on the RealSR LPIPS validation metric invites selection bias for the RealSR LPIPS headline.","major_comments":[{"comment":"The ablation tables contain an apparent inconsistency. Table 2's 'w/o Pixel-Grounded Guidance' row reports DIV2K-Val PSNR 23.68 and LPIPS 0.2431, while Table 5's 'FM control only' row reports PSNR 24.67 and LPIPS 0.2431 for what appears to be the same no-pixel-guidance configuration. The identical LPIPS value with a roughly 1 dB PSNR gap suggests the two tables are not mutually calibrated or describe different checkpoints/training stages. Please clarify the exact configuration of each row and, if they are meant to be the same, reconcile the discrepancy.","section":"Section 4.2, Tables 2 and 5"},{"comment":"The decoder-side grounding claim is not isolated from the image-space losses. The full objective in Section 3.3 is L = L_fm + λ_perc L_perc + λ_pg L_pg, and Table 5's progression adds L_perc and L_pg while the DFM branch is enabled in the PGSR default row. There is no row with L_fm + L_perc + L_pg applied to the standard frozen-decoder output without the DFM branch. Consequently, the gains attributed to decoder-side grounding in Table 2 (1.11 dB PSNR and 0.0084 LPIPS on DIV2K-Val) could be partly or wholly due to the additional low-noise reconstruction and perceptual losses rather than to multi-scale pixel injection. Please report an ablation that holds the loss terms fixed and toggles only the DFM adapters, or explicitly state whether the 'w/o Decoder-Side Grounding' row already includes L_perc and L_pg on the standard decoder output.","section":"Section 3.3, Table 5, and Section 4.2"},{"comment":"The main model is selected by RealSR validation LPIPS, and Table 1 then reports PGSR as best on RealSR LPIPS. Selecting the checkpoint on the same metric and benchmark used for the headline comparison introduces selection bias. Please either select the final model on a held-out split, report results for all candidate checkpoints, or add a caveat that the RealSR LPIPS number is a selected rather than independent estimate.","section":"Section 4, checkpoint selection and Table 1"},{"comment":"All quantitative results are reported as single runs without error bars, significance tests, or seeds beyond the fixed seed 42. Several headline differences are small, such as DIV2K-Val PSNR 25.89 vs. SinSR 25.78 and LPIPS 0.2104 vs. 0.2107. Given the central claim is that PGSR improves the realism-fidelity trade-off, please add variance across seeds or a statistical test for the main comparisons, or clearly state the single-seed limitation.","section":"Tables 1 and 4"}],"minor_comments":[{"comment":"The VAE reconstruction stress test is based on a single image, but the text states that 'both models exhibit substantial information loss' and later refers to 'unavoidable degradation.' Please aggregate this analysis over a validation set with standard deviations before making general claims about the VAE bottleneck.","section":"Figure 3"},{"comment":"The DIV2K-Val PSNR values in Table 3 (26.36 for standard PGSR and 25.81 for sparse attention) differ from the Table 1 value of 25.89 for the same dataset. Please clarify the evaluation resolution, tiling, or checkpoint differences between these tables, since the reader cannot tell whether the sparse-attention comparison is on the same protocol.","section":"Table 3"},{"comment":"The pixel-conditioning strength appears as λ in the gated residual formula but as s_pc in Table 4. Please unify the notation or define the relationship explicitly to avoid ambiguity.","section":"Section 3.2 and Table 4"},{"comment":"The low-noise mask m_i = 1[σ_i ≤ τ] is defined, but the threshold τ is not given in the implementation details. Please report the value used for both L_perc and L_pg.","section":"Section 3.3"},{"comment":"The propositions in the supplementary analytical discussion are labeled A.1, A.2, and A.3, but they appear in Appendix C, not Appendix A. Please renumber the proposition labels to match the appendix.","section":"Appendix C"}],"recommendation":"major_revision","confidential_remarks":"The paper addresses a relevant and timely problem, and the proposed architecture is plausible, but the current evidence for the decoder-side contribution is not yet clean because the loss-component and architecture ablations are confounded. The RealSR checkpoint-selection issue also needs to be addressed before the empirical claims are fully trustworthy. I would be willing to review a revised version."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things you should know. First, the core idea is new: rather than only feeding pixel information into the latent diffusion condition, PGSR also injects multi-scale pixel features into the frozen VAE decoder, and this two-stage reuse of pre-VAE evidence is not present in PASD, PiSA-SR, GuideSR, or the other cited pixel-guided SR methods. Second, the headline claim about decoder-side grounding is not actually proven by the ablations as presented. The design is plausible and the benchmark numbers are strong, but there is a real confound.\n\nWhat is good: the paper is clearly motivated, the architecture is well specified, and the experiments are extensive. The results on DIV2K-Val, RealSR, and DRealSR are competitive on distortion and perceptual metrics, and the sparse-attention variant is a useful efficiency study. The appendix propositions are honestly framed as consistency/inductive-bias arguments rather than optimality guarantees. The writing is direct and the method is reproducible in principle, though no code link appears in the text.\n\nThe soft spots are in the ablations. The stress-test note lands. Table 2's \"w/o Decoder-Side Grounding\" row reports exactly the same numbers (24.78 PSNR, 0.2188 LPIPS) as Table 5's \"+ perceptual\" row, which uses only L_fm + L_perc with no pixel-grounded reconstruction loss L_pg. That means when the DFM branch is removed, L_pg is also dropped. Table 5 then adds L_pg and the DFM branch simultaneously in its next row. There is no row with L_fm + L_perc + L_pg but no DFM. So the 1.11 dB PSNR and 0.0084 LPIPS gain attributed to decoder-side grounding in Table 2 could be mostly due to the extra loss, not the architecture. This is a genuine flaw in the central claim.\n\nThere is also an internal inconsistency: Table 2's \"w/o Pixel-Grounded Guidance\" (PSNR 23.68, LPIPS 0.2431) and Table 5's \"FM control only\" (PSNR 24.67, LPIPS 0.2431) share the exact same LPIPS but differ by about 1 dB PSNR. These are meant to describe similar no-pixel-guidance configurations, and identical LPIPS with mismatched PSNR suggests the ablation tables were not cross-checked.\n\nOther concerns are minor but real: no error bars or significance tests, the VAE reconstruction stress test in Fig. 3 is a single image, and FaithDiff, GenDR, and GuideSR are cited but not compared despite being positioned as closely related. Checkpoint selection on validation metrics is standard practice and not a problem by itself.\n\nWho should read this: anyone working on generative SR or latent diffusion restoration, because the dual-stage pixel grounding is a sensible direction that others will build on. The paper deserves a serious referee, but the revision must include a clean ablation that separates the DFM architecture from the pixel-grounded loss terms, and should reconcile the two ablation tables. Send it to review with that condition.","headline":"A genuinely new dual-stage pixel-grounding idea for latent DiT SR, but the decoder-side ablation is confounded with the pixel-reconstruction loss, so the central claim needs a cleaner experiment.","tokens_in":16400,"tokens_out":3121,"would_cite":false,"duration_ms":28203,"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":"PGSR claims that preserving LR pixel evidence before VAE compression and reusing it during both latent trajectory formation and VAE decoding improves the realism-fidelity trade-off in diffusion transformer super-resolution, reducing…","keywords":["super-resolution","latent diffusion","diffusion transformer","VAE bottleneck","fidelity","flow matching","pixel grounding","ControlNet"],"falsifier":"A concrete test is to scramble the scale-matched taps during inference, feeding the H/2 tap to the H/8 decoder stage and so on, and measure PSNR and LPIPS; if the model is largely insensitive to tap-to-stage alignment, then the scale matching central to the mechanism is not load-bearing.","tokens_in":15383,"feed_emoji":"🖼️","tokens_out":7114,"duration_ms":61062,"temperature":0.7,"pith_summary":"The paper claims that in latent diffusion transformer super-resolution, the VAE compression bottleneck—not the conditioning strength—is a primary cause of hallucinated details, and that preserving LR pixel evidence before VAE encoding and reusing it at two stages yields a better realism-fidelity trade-off. To test this, the authors propose PGSR, which extracts pre-VAE pixel features from the bicubically upsampled LR image, fuses them into the latent LR condition to guide the flow-matching trajectory, and injects multi-scale pixel taps into the frozen VAE decoder to ground the final rendering. They keep the pretrained DiT backbone and VAE frozen, training only lightweight ControlNet-style, LoRA, gated fusion, and decoder-side adapters. On DIV2K-Val, DRealSR, and RealSR, PGSR reports the best or near-best scores across distortion, perceptual, distribution-level, and no-reference metrics, and ablations show that each grounding pathway contributes. A sympathetic reader would take this as evidence that latent generative SR can be made more faithful by compensating for the VAE bottleneck rather than by strengthening conditioning alone.","feed_headline":"Reusing pre-VAE pixels grounds diffusion super-resolution","feed_subtitle":"Two-stage pixel grounding beats compressed-latent-only SR on fidelity and realism, keeping the model frozen.","key_machinery":"The central object is pre-VAE pixel evidence: image-domain features extracted from the bicubically upsampled LR image before VAE compression, consisting of a latent-resolution feature and three multi-scale taps at half, quarter, and eighth resolution. The argument is carried by two mechanisms. Condition-Side Trajectory Guidance uses a gated residual fusion to combine the VAE-encoded LR latent with the pre-VAE pixel feature before the ControlNet branch, anchoring the flow-matching trajectory without replacing the compact condition. Decoder-Side Pixel Grounding injects scale-matched pixel taps into three upsampling blocks of the frozen VAE decoder via zero-initialized projections, so the decoder is identical to the pretrained VAE at initialization and learns grounded rendering only under image-space supervision. These mechanisms specifically target the two loss points the paper identifies: latent trajectory drift and detail loss during final VAE decoding.","core_discovery":"The central discovery is that the information bottleneck of VAE tokenization, rather than insufficient conditioning, is a key driver of fidelity loss in latent generative super-resolution. PGSR extracts pre-VAE pixel evidence from the upsampled LR image—a latent-resolution feature and three multi-scale decoder taps—and reuses it at two points: Condition-Side Trajectory Guidance applies a gated residual correction to anchor the ControlNet condition and the flow trajectory; Decoder-Side Pixel Grounding adds zero-initialized adapters to three matched upsampling stages of the frozen VAE decoder, so the decoder starts identical to the pretrained VAE and gradually learns pixel-grounded rendering. The paper reports that this dual design improves PSNR, LPIPS, DISTS, FID, and NIQE on DIV2K-Val and DRealSR over prior latent generative SR methods, and the ablations show removing either pathway degrades the headline metrics. The authors also give two analytical arguments: gated residual fusion is no worse than fixed additive fusion under a zero-mean noise model of the pixel correction, and multi-scale taps are at least as expressive as a flat latent-resolution extractor.","pith_inferences":["One implication the paper leaves implicit is that if the VAE bottleneck is the root cause, training or fine-tuning the VAE itself with a high-frequency-aware reconstruction loss could complement, or even reduce the need for, decoder-side grounding.","A testable extension is to apply the same pre-VAE grounding to other latent diffusion tasks that suffer from text or structure corruption, such as text-to-image editing or inpainting, where decoder-side feature injection might reduce artifacts similarly.","The paper's comparisons are against methods that largely use pixel evidence only as an external condition; a fair next step would be a controlled comparison where a strong baseline gets the same compute budget for conditioning, to isolate the contribution of decoder-side injection rather than overall adaptation capacity.","The analytical propositions suggest that gated residual fusion is near-optimal under zero-mean noise; one could measure the learned gate values under severe degradation to check whether the model actually attenuates pixel corrections as degradation increases."],"forward_implications":["If correct, latent generative SR can be made more faithful by adding pixel-grounded conditioning and decoding around a frozen VAE and DiT, without retraining the base model.","The two-stage design gives a recipe: any latent-space generator that needs pixel-level fidelity should consider preserving pre-compression cues for both the trajectory and the decoder, not just appending them as a condition.","The sparse-attention results suggest that local-window attention can cut inference cost substantially while keeping most of the fidelity gain, making 2K-to-8K SR practical on the same pipeline.","The gated residual and zero-initialized DFM adapters are generic enough to transfer to other frozen latent diffusion models for restoration and editing.","The paper's own limitation about slightly smoother textures indicates a path forward: combine pixel-grounded gains with distillation to approach one-step SR without sacrificing fidelity."],"supporting_citations":[{"why":"Supplies the VAE tokenizer and latent diffusion formulation whose compression bottleneck the paper identifies as the cause of fidelity loss.","marker":"Rombach et al. [2022]"},{"why":"The FLUX.1-dev pretrained flow-matching transformer and VAE that PGSR adapts with frozen weights.","marker":"Black Forest Labs [2024]"},{"why":"ControlNet; the conditioning architecture that Condition-Side Trajectory Guidance builds on and initializes from.","marker":"Zhang et al. [2023]"},{"why":"Real-ESRGAN; the second-order degradation pipeline used to synthesize real-world LR-HR training pairs in Stage 2.","marker":"Wang et al. [2021]"},{"why":"Formalizes the perception-distortion trade-off that PGSR claims to improve, providing the evaluation framing.","marker":"Blau and Michaeli [2018]"},{"why":"Prior analysis of the reconstruction-generation trade-off in latent diffusion models that motivates the VAE bottleneck argument.","marker":"Yao et al. [2025]"}],"fun_headline_variants":["Pixel evidence before VAE compression grounds super-resolution","Pre-VAE pixels fix diffusion SR fidelity","Reuse pixels from before VAE to restore SR fidelity","Ground latent diffusion SR with pre-compression pixels"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The decoder-side grounding assumes that pixel features from the bicubic-upsampled LR at H/2, H/4, and H/8 line up with the frozen VAE decoder's internal activations at the three matched upsampling stages, and that zero-initialized adapters can learn a useful additive correction; if the scale matching is imperfect, the decoder pathway contributes little and the fidelity gains come mainly from condition-side guidance.","fun_headline_variants_meta":{"raw":{"variants":["Pixel evidence before VAE compression grounds super-resolution","Pre-VAE pixels fix diffusion SR fidelity","Reuse pixels from before VAE to restore SR fidelity","Ground latent diffusion SR with pre-compression pixels"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000351,"raw_usage":{"total_tokens":1966,"prompt_tokens":1047,"completion_tokens":919,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":663,"completion_tokens_details":{"reasoning_tokens":871}},"tokens_in":663,"tokens_out":919,"duration_ms":6772,"temperature":1.0,"reasoning_tokens":871,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T22:52:10.102676+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A concrete test is to scramble the scale-matched taps during inference, feeding the H/2 tap to the H/8 decoder stage and so on, and measure PSNR and LPIPS; if the model is largely insensitive to tap-to-stage alignment, then the scale matching central to the mechanism is not load-bearing.","supporting_citations":[],"review_version":1}