{"id":"3bfa0a8e-a781-43e0-95ed-cbdf5cdc2cee","arxiv_id":"2504.15176","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"DSPO trains real-world super-resolution models with per-object human-preference pairs and text feedback on hallucinated details, and reports higher human-preference win rates than its baselines on DRealSR and RealSR.","lead":"This paper applies human-preference training to image super-resolution, but at the level of individual objects within the image instead of the whole image. The authors report that their method, DSPO, wins human preference tests against their baselines and reduces false details in one-step and multi-step super-resolution models.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Instance-mask correspondence across SR outputs is assumed but never specified; without it, the preference pairs and reported win-rate claim are unverified.","rationale":"The reader identified exactly the load-bearing weak point, and I agree: all downstream results depend on the assumption that the same instance region across different SR outputs corresponds to the same physical content. The paper provides no registration or matching procedure, so the preference pairs used to train DSPO may compare different objects or regions. This is not a stylistic gap but a validity condition for the instance-level preference data and for Eq. 5. The human evaluation is otherwise a meaningful independent check, so the evidence does not warrant rejection, only a conditional verdict pending clarification or release of the matching procedure. Thus the reader's CONDITIONAL verdict should stand unchanged.","tokens_in":12231,"tokens_out":4627,"duration_ms":46088,"concrete_test":"Request the authors' mask-matching code or reconstruct the dataset: for each LQ image, segment the four SR outputs with SAM and compute pairwise IoU after aligning masks (e.g., Hungarian assignment). Report the fraction of 'same-instance' regions with IoU below 0.5; if it is substantial (>10%), the preference labels and Eq. 5 loss are computed on mismatched content, and the win rates must be recomputed with explicit correspondence.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that DSPO improves human preference alignment depends on comparing the 'same instance' across the four SR outputs of each LQ image (Sec. 4.2.1). The paper asserts that after SAM segmentation we can obtain the different instance-level image within the same instance sm region across different images, but it never defines how SAM masks from different outputs are registered or matched. SAM masks are not naturally ordered: the same physical object can receive different mask indices, segment counts can differ, and hallucinated content can introduce spurious regions. If masks are not matched by content, the preferred/dispreferred pair for a given sm can compare different physical regions, and the masked denoising-error loss in Eq. 5 is optimized against mislabeled pixels. Because every instance-level preference pair in the 2,500 annotated regions rests on this correspondence, the 73.6% and 75.6% human win rates are not yet supported. A matching procedure (or evidence that masks correspond) is a prerequisite, not a detail.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript introduces DSPO, a preference-alignment method for real-world image super-resolution that applies Direct Preference Optimization at the semantic-instance level rather than the whole-image level. Given a low-quality input, the pre-trained SR model generates several candidate outputs; each output is segmented with SAM into instance regions, and per-instance preferred/dispreferred pairs are selected either by human annotators or by an automatic sum of eight IQA metrics. The training loss is a masked diffusion-DPO objective (Eq. 5) with instance-size weights (Eq. 3), and a user-description feedback strategy is proposed to suppress hallucinations. Experiments compare DSPO against SFT, DDPO, and Diffusion-DPO on one-step OSEDiff and multi-step SD2-base frameworks; human win rates of 73.6% on DRealSR and 75.6% on RealSR are reported for the one-step framework.","tokens_in":12401,"tokens_out":6337,"duration_ms":53033,"significance":"The contribution is potentially valuable: it is a plausible first application of DPO-style preference alignment to real-world image SR, and the idea of per-instance preferred/dispreferred preferences over whole-image preferences addresses a genuine tension between pixel-level reconstruction and image-level reward. The human study is an independent, meaningful signal, and reporting 95% confidence intervals is a strength. The paper also provides ablations of the two proposed strategies. However, the current manuscript does not specify how SAM masks are matched across candidate images, and the automatic-IQA evaluation is circular with respect to the preference-label construction. These issues must be resolved before the central claims can be accepted.","major_comments":[{"comment":"The construction of instance-level preference pairs assumes that the same semantic region s_m can be compared across the N SR outputs, but no matching or registration procedure is specified. SAM applied independently to each SR output yields an unordered, image-dependent set of masks: mask indices can differ, segment counts can differ, and hallucinated content can introduce spurious regions. If the 'same s_m' across outputs is selected by mask index only, the preferred/dispreferred pairs in Eq. (4) and the masked losses in Eq. (5) can compare non-corresponding physical regions. Because all 2,500 annotated instance regions and the reported human win rates depend on this correspondence, the paper must define the matching rule (e.g., IoU-based mask matching, a shared mask from the LQ/HQ image, or post-hoc association), and must state that the same rule was used in both the human and automatic annotation pipelines.","section":"Sec. 4.2.1 (mask correspondence)"},{"comment":"The automatic preference labels are created by summing the same eight normalized IQA metrics (PSNR, SSIM, LPIPS, DISTS, NIQE, MUSIQ, MANIQA, CLIPIQA) that are then used as the evaluation metrics in Sec. 5.2.2. Training on labels derived from these metrics can only increase the same metrics, so the reported 'improvements across all metrics' are partly self-fulfilling and do not constitute independent evidence of quality. The human win rates in Sec. 5.2.1 are independent and should remain the primary evidence; the automatic-IQA comparison should either use held-out metrics not involved in label construction or be explicitly framed as a training-signal consistency check.","section":"Sec. 5.1.1 vs. Sec. 5.2.2"},{"comment":"The user-description feedback strategy is not defined at the level of the optimization objective. Eq. (5) contains no term involving the negative text prompt, while Eq. (8) only writes L(x_w, x_l | x_LQ, s_m, p_negative) without specifying how p_negative conditions the diffusion model or enters the preference loss. If the feedback prompts are used only at inference time, this should be stated and the inference-time mechanism described; if they are used during training, the exact loss term must be given. Without this, the reported hallucination suppression from the M2 ablation is not attributable to a concrete mechanism.","section":"Sec. 4.3 and Eq. (8)"},{"comment":"The model probability p_theta(x | x_LQ, s_m) is not defined, and it is unclear whether the semantic mask is a conditioning input to the diffusion model. The architecture description and inference procedure mention only the LQ image and cfg scale; Eq. (5) applies the mask as a spatial weight on the denoising error, not as a conditioning signal. Please state explicitly whether p_theta is conditioned on s_m, or reformulate Eq. (4) as a masked weighted DPO loss and clarify how the DPO likelihood ratio is computed in that case.","section":"Sec. 4.2.1, Eq. (4)"}],"minor_comments":[{"comment":"'may leading to' should be 'may lead to'.","section":"Abstract and Sec. 1"},{"comment":"The sentence 'Here, condition text is compactness.' appears garbled and should be removed or rewritten; the condition c is not defined in the diffusion-DPO context.","section":"Eq. (2)"},{"comment":"'DDPO [9]' is cited to Reference [9], which is the DDPM paper (Ho et al.), not the DDPO paper; the reference list needs correction.","section":"Sec. 2.3"},{"comment":"The notation sum_{m=1}^M s_m = 1 implies the SAM masks form a partition, but SAM masks typically overlap and may omit background; please specify any post-processing used to make the masks disjoint.","section":"Sec. 4.2.1"},{"comment":"The statement 'resulting in a total of 2500 instance regions for annotation, as each LQ has four corresponding SR results' is ambiguous; please clarify whether 2,500 refers to instances (500 images x 5 instances) or to instance-image pairs including the four candidates.","section":"Sec. 5.1.1"},{"comment":"'different different optimization objectives' contains a duplicated word.","section":"Table 1 caption"}],"recommendation":"major_revision","confidential_remarks":"Editor: the automatic-IQA circularity is the main reason I did not recommend minor revision; it is fixable, but the paper's current framing overstates the automatic evaluation. The mask-correspondence issue is the substantive technical gap. I would also like the authors to state clearly whether the human win rates cover both one-step and multi-step frameworks, since Sec. 5.2.1 reports them only for the one-step framework."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First things first: this is a legitimate new application, not a rehash. It adapts Diffusion-DPO to real-world SR and does something sensible about the image-level vs pixel-level mismatch by doing preference optimization at the instance level. SAM masks carve the SR outputs, Best/Worst-of-N picks winners and losers per instance, and Eq. 5 is a mask-weighted DPO loss. Plug-and-play on both one-step OSEDiff and multi-step SD2-base, with ablations showing both the instance alignment and the BLIP text-feedback strategies earn their keep. The human win rates — 73.6% on DRealSR and 75.6% on RealSR with three annotation rounds and confidence intervals — are the real result, and they are independent of the training labels.\n\nThe soft spots are real but mostly fixable. The load-bearing one is what the stress-test flags: Sec. 4.2.1 claims you can obtain the same instance region s_m across different SR outputs, but never says how the SAM masks are matched or registered. SAM mask indices are arbitrary, segment counts differ, and hallucinated content adds spurious regions. If the masks don't correspond to the same physical object, every instance-level preference pair is potentially comparing different content, and the loss in Eq. 5 is trained on mislabeled pixels. A matching procedure — or at least evidence that masks line up — is required. The central claim isn't dead, but it's unverified until that's addressed.\n\nSecond: the automatic-IQA evaluation in Sec. 5.2.2 uses the same eight metrics that were summed in Sec. 5.1.1 to create the training preferences. Those gains are partly self-fulfilling; the human win rates carry the claim, and the IQA section should be read as a sanity check, not independent evidence. Minor stuff: no code or artifacts, baselines don't include current Real-ISR methods like DiffBIR or SeeSR, and the 10 annotators are described but no inter-annotator agreement is reported. Also there's a stray 'compactness' typo in Sec. 3.2 that should be cleaned up.\n\nBottom line: worth sending to peer review. The idea is transferable, the engineering is honest, and the human evaluation is the right kind of evidence. The authors need to specify mask matching and recalibrate the IQA claims, but I'd rather see this in the review process than desk-rejected. I'd bring it to the reading group.","headline":"A legitimate instance-level DPO adaptation for real-world SR with strong human win rates, but the unspecified mask-correspondence step and circular automatic-IQA evaluation leave the central claim only partially verified.","tokens_in":12955,"tokens_out":2490,"would_cite":true,"duration_ms":21604,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that applying DPO at the semantic-instance level, with negative text feedback, aligns real-world super-resolution models with human preferences, raising average win rates to 73.6% and 75.6% on DRealSR and RealSR.","keywords":["real-world image super-resolution","direct preference optimization","semantic instance alignment","human preference alignment","hallucination suppression","diffusion models","prompt injection","instance-level DPO"],"falsifier":"Run SAM on the four SR outputs for a set of LQ images, align the outputs, and compute the overlap between the mask used for the preferred image and the mask used for the dispreferred image; if those masks do not correspond to the same scene instance, the instance-level preference pairs are comparing different content and the reported win-rate gains should be re-examined.","tokens_in":12000,"feed_emoji":"🖼️","tokens_out":5548,"duration_ms":48477,"temperature":0.7,"pith_summary":"This paper claims that human preference alignment can be brought to real-world image super-resolution by applying Direct Preference Optimization not to whole images but to semantic instances within them. It identifies a core conflict: image-level preference judgments reward outputs that are sharp overall yet hallucinate or artifact in local regions, which fights the pixel-level reconstruction objective. To resolve it, the proposed Direct Semantic Preference Optimization (DSPO) segments each upscaled output into instances, builds preferred and dispreferred pairs instance-by-instance, weights the DPO loss by instance area, and injects user-selected text descriptions of hallucinated instances as negative prompts. The paper reports that this raises human preference win rates over the pre-trained model to 73.6% and 75.6% on DRealSR and RealSR with a one-step SR model, and improves perception and fidelity metrics on both one-step and multi-step frameworks. A reader would care because it offers a way to align generative restoration models with human taste without training a separate reward model.","feed_headline":"Super-resolution tuned on human taste wins 75% of comparisons","feed_subtitle":"Instance-level direct preference optimization beats pretrained, SFT, and DPO baselines on real-world SR benchmarks.","key_machinery":"The load-bearing object is the mask-weighted, instance-level DPO loss (Eq. 5): for each semantic region $s_m$, the denoising error difference between the current model and the reference model is computed only inside the mask, weighted by $w_m = |s_m| / \\sum_m |s_m|$, and passed through a log-sigmoid margin. SAM supplies the masks that define 'the same instance' across outputs; the reference model's per-region denoising error acts as a distribution-drift anchor; the user-description feedback supplies negative prompts constructed from BLIP captions of hallucinated regions. The claim is that this machinery makes the preference signal fine-grained enough to coexist with reconstruction objectives.","core_discovery":"On its own terms, the paper's central claim is that the mismatch between holistic human preferences and pixel-level super-resolution objectives can be dissolved by making preferences instance-level and semantic. DSPO generates several SR outputs per low-quality input, segments each with SAM, selects best and worst outputs within each instance region under human or automatic evaluation, and optimizes the diffusion denoising objective with a DPO-style margin weighted by the relative area of each semantic instance. The same pipeline adds a hallucination-suppression branch: a vision-language model describes each instance, and descriptions of instances judged inconsistent with the input are stored as negative text feedback and injected into the prompt branch during training. The paper's empirical assertion is that this combination, applied as a plug-in to a one-step diffusion SR model and a multi-step text-to-image backbone, outperforms the pre-trained models, SFT, DDPO, and Diffusion-DPO on human preference win rates and on the reported IQA metrics.","pith_inferences":["The paper never specifies how SAM masks are registered across different SR outputs; a testable extension would check mask correspondence, since the instance-level pairs are the whole basis of the loss.","A practical extension would replace human annotation with learned IQA scores entirely and measure whether the win-rate gains survive; the paper already uses an automatic IQA variant, so the human step may be compressible.","The same instance-level preference recipe could transfer to other dense generative restoration tasks where local hallucinations dominate, such as deblurring or deraining."],"forward_implications":["Instance-level semantic preference alignment can be added to existing one-step and multi-step diffusion SR models without changing inference-time cost, since the alignment is a training-phase objective.","Human annotators prefer DSPO outputs over the pre-trained model roughly three-quarters of the time on real-world benchmarks, and over SFT, DDPO, and Diffusion-DPO as well.","Automatic quality metrics improve alongside human preference, so the method does not merely chase annotator taste.","The negative text-feedback branch contributes separately to artifact suppression, since removing it degrades performance in the ablation."],"supporting_citations":[{"why":"Supplies the direct preference optimization objective that DSPO adapts to the instance level.","marker":"[22]"},{"why":"Supplies the diffusion-timestep DPO loss form and the Diffusion-DPO baseline that DSPO compares against.","marker":"[29]"},{"why":"Supplies SAM, the semantic extraction model whose masks define the instance regions used in the loss.","marker":"[13]"},{"why":"Supplies BLIP, which generates the text descriptions used to build hallucination feedback prompts.","marker":"[15]"},{"why":"Supplies the Real-ESRGAN degradation pipeline used to construct the LQ-HQ training pairs.","marker":"[32]"},{"why":"Supplies the OSEDiff one-step diffusion SR framework on which DSPO is plugged in.","marker":"[36]"},{"why":"Supplies the DRealSR test set used for the reported human preference win rates.","marker":"[34]"},{"why":"Supplies the RealSR test set used for the reported human preference win rates.","marker":"[4]"}],"fun_headline_variants":["DSPO: semantic preferences sharpen real-world super-resolution","Human-preference alignment boosts super-resolution realism","Semantic DPO makes super-resolution prefer what humans prefer","DSPO tops DPO on real-world super-resolution"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the SAM mask of an instance in one SR output and the SAM mask of the 'same' instance in another output refer to the same real-world object; if the masks are never registered or matched across outputs, the preferred and dispreferred pair compares different content and the mask-weighted DPO loss is computed on mislabeled regions.","fun_headline_variants_meta":{"raw":{"variants":["DSPO: semantic preferences sharpen real-world super-resolution","Human-preference alignment boosts super-resolution realism","Semantic DPO makes super-resolution prefer what humans prefer","DSPO tops DPO on real-world super-resolution"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001211,"raw_usage":{"total_tokens":5005,"prompt_tokens":985,"completion_tokens":4020,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":601,"completion_tokens_details":{"reasoning_tokens":3956}},"tokens_in":601,"tokens_out":4020,"duration_ms":22712,"temperature":1.0,"reasoning_tokens":3956,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T11:31:19.138226+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run SAM on the four SR outputs for a set of LQ images, align the outputs, and compute the overlap between the mask used for the preferred image and the mask used for the dispreferred image; if those masks do not correspond to the same scene instance, the instance-level preference pairs are comparing different content and the reported win-rate gains should be re-examined.","supporting_citations":[{"cited_title":"Diffusion model align- ment using direct preference optimization","cited_arxiv_id":null,"evidence_quote":"Supplies the diffusion-timestep DPO loss form and the Diffusion-DPO baseline that DSPO compares against."},{"cited_title":"Segment any- thing","cited_arxiv_id":null,"evidence_quote":"Supplies SAM, the semantic extraction model whose masks define the instance regions used in the loss."},{"cited_title":"One-step effective diffusion network for real-world image super-resolution","cited_arxiv_id":null,"evidence_quote":"Supplies the OSEDiff one-step diffusion SR framework on which DSPO is plugged in."}],"review_version":1}