{"id":"c38e09ce-85f3-40a3-b8c3-3697ed5e1e65","arxiv_id":"2507.06146","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A prompt-free conditional diffusion framework with a detector-based counting loss augments multi-object images with matched per-category object counts and diverse layouts, achieving the best downstream COCO detection mAP among compared baselines (39.04 vs 38.65 without augmentation).","lead":"Prompt-free diffusion generates multi-object augmentation images: Stable Diffusion XL is conditioned on a full image plus cropped object patches, with a counting loss that uses an open-vocabulary detector to keep per-category object counts from falling. On COCO, the augmented mix raises downstream Mask R-CNN bbox mAP from 38.65 to 39.04, while generated images still fall well below the real-data instance count (IQS 29.17 vs 45.02).","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The count-preservation claim is contradicted by the paper's own IQS metric, and the counting-loss mechanism is evaluated only through that same self-defined, unpublished metric.","rationale":"The paper's central contribution is the reward-model counting loss that is supposed to keep per-category object counts from decreasing while increasing diversity. That claim is directly testable with the paper's own IQS metric, which they define and compute. The reported IQS is 29.17 for generated images versus 45.02 for originals; if IQS meaningfully reflects per-category counts, this is a substantial shortfall that the text never reconciles with the 'count of it o_l >= o_j' formulation. The reader identified the same weak spot: the counting loss is computed on one-step denoised estimates, and the IQS result is evidence the proxy does not perfectly transfer to final images. I agree with that analysis. My stress-test adds that the IQS algorithm is referenced to an Appendix that is not present in the supplement, so the very metric used to validate the core claim is not fully specified; a reader cannot independently verify the central number. This is an addressable issue, not a fatal one: the paper is otherwise coherent, the ablations show the counting loss moves IQS by about 1-3 points, and the downstream mAP gains are positive albeit small and without error bars. I keep the reader's CONDITIONAL verdict rather than moving to REJECT because the concern is about evidence completeness and metric-proxy alignment, not internal inconsistency or a demonstrated false mechanism; the authors could plausibly close the gap by providing the IQS algorithm, reporting per-category count statistics on final images, and comparing against an independent annotator. The recommendation is CONDITIONAL: accept only after the count-preservation claim is reconciled with the IQS gap using a fully specified metric and final-image evaluation.","tokens_in":14626,"tokens_out":1599,"duration_ms":14646,"concrete_test":"Run the released code on a fixed 100-image subset of COCO val2017: (1) regenerate final 50-step images, (2) implement IQS exactly as described (or, if the algorithm is missing, use YOLOv8m with the stated multi-confidence protocol) and compute per-category count ratios, (3) compare generated per-category counts to seed-image counts; if the fraction of categories with o_l < o_j exceeds, say, 10% or mean IQS stays below 35, the Sec. 3.1 count-nondecrease claim fails on final images.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim (Sec. 3.1: for each category c_j in the input, the generated image must contain a matching category with count o_l >= o_j) is contradicted by the paper's own aggregate metric. Table 2 reports IQS 29.17 for generated images vs 45.02 for real COCO val2017, a 35% shortfall, and the IQS algorithm is only 'shown in Appendix' but never included in the submitted supplement. Because IQS is the only quantitative support for the counting loss and count-preservation claims (Tables 2-5), a missing algorithm definition makes the flagship claim unfalsifiable from the manuscript. Additionally, the training-time loss (Eq. 5-7) uses Grounding DINO top-k confidence on a one-step denoising estimate x*_i at random timestep t, while the final images are 50-step Euler samples; the reader identifies this proxy misalignment, and I agree it is load-bearing, but the immediate blocker is that the claimed outcome (counts do not decrease) is not consistently demonstrated on final images with a sufficiently specified metric.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a prompt-free conditional diffusion framework for multi-object image augmentation. It replaces text conditioning with a local-global semantic fusion of the input image and adds a Grounding DINO-based counting loss that, per Eq. (6)-(7), penalizes missing objects relative to per-category counts of the conditioning image. The method is evaluated on COCO with FID, DS, and a proposed IQS metric, plus Mask R-CNN downstream mAP. The paper reports best FID/DS among image-variation baselines and best bbox mAP (39.04) among all compared augmentation methods. The central claim is that generated images preserve or increase per-category object counts while diversifying layout.","tokens_in":14635,"tokens_out":4671,"duration_ms":46960,"significance":"If the claimed count preservation holds, the method is useful for multi-object augmentation because it avoids text-category bias and layout rigidity. The paper does provide reproducible code, detailed hyperparameters (supplement Table 1), ablations for each component (Tables 3-5), and out-of-domain qualitative results. However, the central claim is currently only supported by an unpublished IQS metric, and the reported IQS values contradict the 'counts do not decrease' claim (29.17 vs 45.02 for real COCO).","major_comments":[{"comment":"The central requirement ol >= oj in Sec. 3.1 is not met by the paper's own metric: Table 2 reports IQS 29.17 for generated images vs 45.02 for real COCO val2017 (35% lower). The text in §4.4 only notes that Add SD is lower than the original, but does not reconcile the shortfall of the proposed method. Because IQS is the only quantitative evidence for count preservation, either the claim should be relaxed or per-image/per-category count-difference statistics on final 50-step samples must be reported.","section":"§3.1, §4.4, Table 2"},{"comment":"The IQS algorithm is said to be 'shown in Appendix', but the supplementary material contains no IQS algorithm; Supplement §1.2 defines only FID and DS. Since Tables 2-5 rely on IQS, the central claim is currently not independently checkable. The authors should include the exact IQS computation (detector confidence settings, per-category aggregation, comparison protocol).","section":"§4.4 and Supplement §1.2"},{"comment":"The counting loss is computed on a one-step denoised estimate x*_i at a random timestep t, while inference uses 50-step Euler sampling. The paper does not demonstrate that gradients through this one-step proxy control object counts in the final images. A direct check--counting objects in final 50-step samples with the same Grounding DINO model used in Eq. (6)--should be added; if the proxy is unreliable, the count-preservation mechanism is unsupported.","section":"§3.4, Eq. (5), Algorithm 1"},{"comment":"The augmentation pipeline is optimized for Grounding DINO confidences (Eq. 6), and downstream annotations for the proposed method are also generated by Grounding DINO + SAM (Sec. 4.3). This circularity could inflate downstream mAP relative to methods whose annotations are not matched to their training signal. Please report downstream results with an alternative annotator (e.g., crowd-sourced COCO labels for generated data or a different open-vocabulary detector) to rule out this confound.","section":"§4.3, Eq. (6)"}],"minor_comments":[{"comment":"There are repeated grammar and typographical issues, e.g., 'Diffusion models has underpinned' and 'augementated'; the paper would benefit from a careful copyedit.","section":"Abstract and throughout"},{"comment":"The symbol ϵt is used for the noise prediction network, which conflicts with the noise ϵt at timestep t; use ϵθ consistently.","section":"Eq. (5)"},{"comment":"The caption contains an unnumbered bullet and the phrase 'Dataset Diffusion decrease in object amount' is grammatically incomplete; please revise.","section":"Fig. 1 caption"},{"comment":"The metric is called 'LSIPS' but the referenced metric is LPIPS; this typo should be fixed.","section":"Supplement §1.2"},{"comment":"The phrase 'Code is available at here' does not resolve to a URL in the PDF; please provide the full link.","section":"Abstract"},{"comment":"For the val2017 row, FID and DS are listed as '-'; the paper should explain explicitly why these are not computed for real validation images.","section":"Table 2"}],"recommendation":"major_revision","confidential_remarks":"The paper is within scope for a computer vision venue and the downstream result is promising. The main concern is that the headline contribution (count preservation) is not supported by a specified metric; if the authors supply the IQS algorithm and reconcile the shortfall, the paper may be acceptable. I do not see a need to reject on the basis of disagreement with consensus; the issues are fixable with additional measurement and analysis."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a coherent systems paper that deserves reading, but the headline claim is only partially supported. The method is a prompt-free diffusion augmentation pipeline: CLIP encodes the whole image plus cropped instances, LoRA adapts SDXL, and a reward-model counting loss (Grounding DINO top-k confidence on one-step denoised estimates) keeps per-category object counts from dropping. That combination is new relative to Prompt-Free Diffusion, Make It Count, and Add-SD, and the engineering story holds together—ablations are consistent and the downstream mAP improves from 38.65 to 39.04.\n\nWhat's good: local-global semantic fusion is simple and effective; the counting loss is a sensible reward term that raises IQS without sacrificing FID or DS; the qualitative OOD results support generalization beyond COCO; code is promised. This is a practical contribution for low-label detection and segmentation.\n\nSoft spots, in rough order.\n\n1. The central count-preservation claim is contradicted in aggregate by the paper's own IQS metric: 29.17 generated vs 45.02 real COCO, a 35% gap the text never mentions. The IQS algorithm is referenced to an Appendix that isn't in the supplement, so the metric can't be checked. If IQS is right, the \"same or more objects\" claim is false; if it isn't right, the paper needs a validated count metric. Either way this needs to be fixed, not just softened.\n\n2. Counting loss trains on one-step denoising at random timesteps, but final images are 50-step Euler samples. That proxy can mislead, and the IQS gap is consistent with it doing so. A final-image detection count with an independent detector would settle it.\n\n3. Grounding DINO is both the training-time reward and the annotation engine for generated training data. That coupling is a real concern; an annotator-agnostic experiment (or real labels) would rule it in or out.\n\n4. mAP differences of 0.1-0.4 between methods have no seeds or error bars. Probably real, but report them.\n\nBottom line: send it to serious review. The new recipe and mostly solid ablations warrant referee time, with expected moderate revision on the count claim, metric disclosure, and error bars. I agree with the conditional verdict, and I think the missing IQS algorithm is the one blocker.","headline":"A useful prompt-free augmentation recipe that deserves review, but the count-preservation claim is weaker than the abstract suggests and is undercut by the paper's own missing IQS metric.","tokens_in":15387,"tokens_out":5583,"would_cite":true,"duration_ms":58097,"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 prompt-free conditional diffusion method generates multi-object augmentations that preserve per-category object counts while varying layout.","keywords":["diffusion models","dataset augmentation","multi-object image generation","prompt-free conditioning","local-global semantic fusion","counting loss","reward model","object detection"],"falsifier":"Count objects in the final 50-step generated images with an independent detector at multiple confidence thresholds and compare with the counts returned on the one-step estimates used for Eq. 5; if the one-step counts do not track the final counts across timesteps and seeds, the counting loss cannot deliver its per-category count guarantee.","tokens_in":14204,"feed_emoji":"🖼️","tokens_out":7840,"duration_ms":72049,"temperature":0.7,"pith_summary":"The paper claims that a diffusion model can augment multi-object training images without any text prompt, using the seed image itself as the condition, and that a counting loss can enforce a per-category count guarantee: each generated image must contain at least as many objects of every category as its seed, while layout and composition are left free to vary. This matters because text-conditioned generation drifts from the target category distribution, and image-editing augmentation adds too little variation to help downstream models. The proposed framework reports better downstream detection and segmentation than existing baselines when generated images are mixed into COCO training, and better fidelity-diversity balance than prior variation methods. The core quantity that improves is the instance quantity score, which moves from 25.67 in the image-only baseline to 29.17 with the full method.","feed_headline":"Image-based diffusion keeps object counts while adding layout variety","feed_subtitle":"Augmented images keep the seed's per-category object counts with new layouts and lift Mask R-CNN mAP to 39.04.","key_machinery":"The load-bearing machinery is the local-global semantic fusion module and the reward-model counting loss. The fusion module uses the CLIP image encoder to encode the whole seed image together with up to nine cropped instances, producing a condition $C(p^{img}_i)$ in the same feature space as the text encoder, so the latent diffusion model can be adapted by LoRA without retraining the full network. The counting loss $L^C_i$ (Eq. 7) sums, over categories, the ReLU distance between a threshold $\\tau$ and the top-$k$ confidence scores returned by Grounding DINO on a one-step denoising estimate $x^*_i$ (Eq. 5), with $k$ equal to the seed object count for that category; it is added to the MSE denoising loss with weight $\\lambda$. This counts objects without constraining their positions, which is what allows layout diversity to coexist with count preservation.","core_discovery":"On the paper's own terms, the central claim is that replacing text prompts with a local-global image condition—global CLIP features of the whole seed image plus features of up to nine cropped instances—and adding a reward-model counting loss produces augmentations whose per-category object counts are at least those of the input image. The counting loss detects objects in a one-step denoised estimate of the generated image with an open-vocabulary detector, and penalizes the top-k confidence scores of each category when they fall below a threshold, without ever constraining bounding-box positions. This is the mechanism the paper says “bridges the quantity deviation” between generated and original data while improving diversity, and it is supported by the reported gains in instance quantity score and by the downstream Mask R-CNN mAP increase from 38.65 to 39.04 on COCO.","pith_inferences":["A direct test of the method's load-bearing proxy would be to compute the counting loss on the final 50-step sample instead of the one-step estimate; if the IQS gap to real data (29.17 vs 45.02) narrows, the proxy mismatch is the cause.","The counting reward could be combined with a spatial-diversity term (for example, a penalty on layout similarity between paired generations) to make the trade-off between count fidelity and layout variation explicit rather than emergent.","The downstream gain of roughly 0.4 mAP is small compared with typical Mask R-CNN training noise; re-running the augmentation mixing with several seeds would indicate whether the gain is stable.","Because the framework conditions on images, it should extend to attribute-heavy or predicate-heavy datasets where text prompts are known to be imprecise, though the paper only demonstrates categories and scenes."],"forward_implications":["Mixing 10,000 generated images into COCO train2017 raises Mask R-CNN bbox mAP from 38.65 to 39.04 and mask mAP from 35.24 to 35.43.","The instance quantity score reaches 29.17 with the full method, up from 25.67 for the image-only baseline, while diversity score rises from 0.36 to 0.71.","Because the condition is an image rather than a category-name prompt, the method generalizes to categories absent from COCO and to cross-view remote sensing images, as shown in the out-of-domain experiments.","Augmenting an already-augmented image recurrently preserves object counts and increases layout spread, so the augmentation can be applied repeatedly to a single seed.","Counting constraints apply per category, not per pixel or box, so the model can vary background and layout while keeping quantity fidelity."],"supporting_citations":[{"why":"Supplies the latent diffusion backbone that the framework adapts from text conditioning to image conditioning.","marker":"[Rombach et al., 2022]"},{"why":"Provides the CLIP image encoder used in the local-global semantic fusion to embed global and cropped image conditions.","marker":"[Radford et al., 2021]"},{"why":"Grounding DINO is the open-vocabulary detector that acts as the reward model for the counting loss.","marker":"[Liu et al., 2024]"},{"why":"SDXL is the specific diffusion model and Euler scheduler setup on which the LoRA fine-tuning and 50-step inference run.","marker":"[Podell et al., 2023]"},{"why":"LoRA is the parameter-efficient adaptation mechanism that injects target-dataset knowledge into the UNet.","marker":"[Hu et al., 2021]"},{"why":"MS-COCO is the dataset used for training, generation-quality evaluation, and downstream detection and segmentation tests.","marker":"[Lin et al., 2014]"},{"why":"Mask R-CNN is the downstream model whose mAP on COCO measures the practical value of the augmented data.","marker":"[He et al., 2017]"},{"why":"Prompt-Free Diffusion is the closest prior image-variation approach and the baseline the paper extends and compares against.","marker":"[Xu et al., 2024]"},{"why":"ControlNet Reference-Only is a comparison method used to show that attention-spliced variation loses object semantics.","marker":"[Zhang et al., 2023]"}],"fun_headline_variants":["Prompt-free diffusion keeps object counts while adding layout variety","Image-based diffusion preserves counts, boosts diversity in augmentations","Counting-loss diffusion retains object counts, enhances image augmentation","No text prompts: diffusion keeps per-category counts and adds diversity"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The counting loss is trained on a noisy one-step denoising estimate, and the method assumes that the detector's top-k confidence on that estimate faithfully reflects the object counts in the final 50-step generated image.","fun_headline_variants_meta":{"raw":{"variants":["Prompt-free diffusion keeps object counts while adding layout variety","Image-based diffusion preserves counts, boosts diversity in augmentations","Counting-loss diffusion retains object counts, enhances image augmentation","No text prompts: diffusion keeps per-category counts and adds diversity"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000209,"raw_usage":{"total_tokens":1410,"prompt_tokens":950,"completion_tokens":460,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":566,"completion_tokens_details":{"reasoning_tokens":393}},"tokens_in":566,"tokens_out":460,"duration_ms":5440,"temperature":1.0,"reasoning_tokens":393,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T19:11:48.232806+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Count objects in the final 50-step generated images with an independent detector at multiple confidence thresholds and compare with the counts returned on the one-step estimates used for Eq. 5; if the one-step counts do not track the final counts across timesteps and seeds, the counting loss cannot deliver its per-category count guarantee.","supporting_citations":[{"cited_title":"Microsoft coco: Common objects in context","cited_arxiv_id":null,"evidence_quote":"MS-COCO is the dataset used for training, generation-quality evaluation, and downstream detection and segmentation tests."}],"review_version":1}