{"id":"495bd4b0-9790-4fb7-8925-49d8c879db32","arxiv_id":"2505.04347","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"CountDiffusion improves object-count accuracy in text-to-image diffusion by detecting objects in a one-step predicted image and applying attention-map guidance to add or remove instances.","lead":"CountDiffusion is a training-free method that uses an off-the-shelf object counter and attention-map edits to push text-to-image models toward generating the requested number of objects. Its two-stage pipeline predicts an intermediate image, counts the objects, then modifies cross-attention to add or remove instances.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported gains are circular: Grounded SAM both drives the correction and defines Acc/MAE, so an independent counting protocol is required before the central claim can be credited.","rationale":"The reader's weakest assumption, that the one-step predicted x0,mid faithfully captures object quantity and positions, identifies a real internal risk; Section IV-A explicitly acknowledges the Markov violation and offers only a qualitative figure as support. However, the most load-bearing concern for the central claim is the closed evaluation loop: the same Grounded SAM model supplies both the correction signal and the accuracy metric, so the quantitative results cannot distinguish successful counting from successful optimization of the detector. If an independent evaluation confirms the gains, the one-step proxy may be acceptably accurate in practice; if not, no amount of internal consistency preserves the claim. I therefore place the concern on external validity rather than on the proxy assumption. The method itself is plausible, and its training-free universal-guidance design is a reasonable contribution, but the evidence as presented does not support the headline accuracy improvements. The reader's REJECT verdict remains appropriate.","tokens_in":10225,"tokens_out":5555,"duration_ms":57335,"concrete_test":"Re-run the evaluation on the same generated images and prompts with an independent counting protocol that never appears in the guidance loop: for example, humans annotate object counts, or a different detector (DINO, Florence-2, or MDETR) with a fixed threshold is used to compute Acc/MAE on CoCoCount and GPTSingleCount/GPTMultiCount. If CountDiffusion's gain over SDXL and Pixart-Sigma persists under this independent protocol, the concern is settled and the central claim is supported; if the gain shrinks or reverses, the reported numbers are artifacts of optimizing Grounded SAM's detections.","verdict_should_be":"UNCHANGED","load_bearing_attack":"CountDiffusion's only quantitative support for 'correct object quantity' is Acc/MAE computed in Section V-B, where 'We employ Grounded SAM to obtain the object quantities required for computing Acc. and MAE.' The same Grounded SAM is the counting model used in the detection stage (Section IV-A, Eq. 10), and the masks from its detections drive the correction loss in Section IV-B. The guidance loop is therefore optimized against the exact detector that produces the headline numbers, so the reported improvements (SDXL 34 to 59 on CoCoCount; Pixart-Sigma 40 to 60) can arise even if the method only changes images in ways that make Grounded SAM's counts match the prompt, rather than improving counts as perceived by humans or measured by an independent detector. In the limiting case, the method could concentrate attention into fewer high-confidence blobs per class and 'succeed' by Grounded SAM's standards while failing a human count. The one-step denoising proxy (Section IV-A) and test-set hyperparameter selection (Tables II and III) are additional risks, but the circular evaluation is the load-bearing concern because it makes the central claim unverifiable from the paper's own evidence.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes CountDiffusion, a training-free framework for improving object-count accuracy in text-to-image diffusion models. It operates in two stages: a detection stage that one-step denoises an intermediate latent to predict the final image, uses Grounded SAM to count objects and obtain masks, and a correction stage that modifies cross-attention maps with a Universal Guidance style loss to add or remove objects until the detected count matches the prompt. A multi-class strategy runs separate single-class corrections and averages the resulting latents. Experiments on CoCoCount and two new GPT-constructed datasets compare SDXL and Pixart-Sigma baselines and report large gains in accuracy and MAE, together with CLIP-score and ImageReward, plus ablations of loss strategies, guidance steps, Gaussian smoothing, and the multi-class strategy.","tokens_in":10530,"tokens_out":4634,"duration_ms":46941,"significance":"If validated, CountDiffusion would be a practically useful contribution: it is training-free, applicable to different diffusion backbones, and requires no human annotation at inference time. The proposed multi-loss universal guidance and the two new evaluation datasets are also potentially useful resources. The paper is clearly written and includes ablations for each design component. However, the quantitative claims are not credible as presented because the same Grounded SAM model that drives the correction loop is also used to compute the headline Acc. and MAE, making the evaluation circular. In addition, the one-step denoising proxy is asserted without quantitative verification, and key hyperparameters appear to be selected on the same CoCoCount test set that is later reported in the main comparison. As a result, the paper's central claim that CountDiffusion improves counting accuracy by a large margin is unverifiable from the evidence provided.","major_comments":[{"comment":"The evaluation is circular. Grounded SAM is used in the detection stage to count objects in the one-step predicted image and to supply the masks that drive the correction loss (Eqs. 12-14), and in Section V-B the same Grounded SAM is used to compute the Acc. and MAE that appear in Table I. The reported improvements (e.g., SDXL 34% to 59% on CoCoCount and Pixart-Sigma 40% to 60%) therefore partly measure the detector's self-consistency rather than counting accuracy by an independent standard. In the limiting case, the method could change images so that Grounded SAM's detections match the prompt while a human or a different detector would still disagree. An independent counting protocol—human counts, a different detector, or at least a report of Grounded SAM's per-class counting accuracy—is required before the central claim can be credited.","section":"§IV-A, §IV-B, §V-B, Eq. (10)"},{"comment":"The entire detection stage depends on the assertion that a one-step denoising prediction x_{0,mid} accurately preserves the object quantity and positions of the final synthesized image. The paper provides only qualitative examples in Fig. 3 and no quantitative comparison between Grounded SAM counts on x_{0,mid} and on the final image. If the proxy is unreliable, the correction masks and the add/remove decisions are computed from an image that does not reflect the final result. Please add a quantitative validation, such as count agreement and mask overlap between x_{0,mid} and the final x_0 over a sample of prompts.","section":"§IV-A, Eqs. (8)-(9), Fig. 3"},{"comment":"The hyperparameters of the method appear to be selected on the same CoCoCount dataset that is later used for the headline comparison. Table II selects P=50 for the top-k loss strategy and Table III selects the number of universal guidance steps and total steps using accuracy on CoCoCount, and Table I then reports CoCoCount accuracy for the final method. No held-out validation split or nested evaluation is described. This test-set fitting inflates the reported gains and should be fixed by reporting validation-selected hyperparameters or by using a separate validation set.","section":"Tables II and III vs. Table I"},{"comment":"The multi-class correction strategy averages per-class corrected latents z_{mid,i} at the intermediate step. The paper provides no analysis of whether this averaged latent stays on the denoising manifold or whether the average of class-specific corrections introduces inter-class conflicts in the subsequent DDIM steps. The ablation in Fig. 8 is qualitative only. A quantitative ablation, including per-class accuracy, failure-mode examples, and a comparison against a jointly optimized multi-class loss, is needed to support the claim that the averaging strategy avoids competition among class losses.","section":"§IV-C, Eq. (15)"}],"minor_comments":[{"comment":"The display equations for the denoising update and the one-step prediction are difficult to parse because of missing parentheses; please format them with unambiguous operator grouping.","section":"§III, Eqs. (2) and (3)"},{"comment":"The experiments state T=40 and t_mid=30, but Table III varies universal guidance steps up to 40 with total steps 40; please clarify how the guidance-step count relates to t_mid and the interval on which the loss is applied.","section":"§V-B and Table III"},{"comment":"The caption says the red smiling face indicates a correct count and the green crying face indicates an error; this color convention is the opposite of typical usage and should be checked and clarified.","section":"Fig. 5"},{"comment":"All reported Acc., MAE, CLIP-score, and ImageReward values lack error bars or significance tests; given the observed differences between some methods are small, standard deviations over multiple runs or seeds should be reported.","section":"§V-B"},{"comment":"Reference [37] contains a malformed URL ('https://https://github.com/...') and should be corrected.","section":"References"},{"comment":"The paper states that two new datasets are built but does not provide a release link or a detailed description of the data collection protocol beyond the prompt counts; please include availability information for reproducibility.","section":"§V-A"}],"recommendation":"reject","confidential_remarks":"The decisive issue is the circular evaluation: the counting model that drives the correction is the same model that produces the headline metrics. The one-step proxy and test-set hyperparameter selection are additional serious concerns. If the authors can supply an independent counting evaluation (e.g., human counts or a different detector) and hold out the hyperparameter selection, the work could potentially be resubmitted as a substantially revised manuscript."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Z. — quick take on CountDiffusion. The idea is genuinely useful: instead of training a model to count, they let the base diffusion model generate normally and then steer it using attention-map losses, with Grounded SAM providing both the count error and the masks. The multi-class per-class averaging trick is sensible, and building LLM-generated prompts is a nice practical addition. The paper is readable and the ablations are reasonable.\n\nBut the evidence for the central claim doesn't hold. The accuracy and MAE are computed by Grounded SAM (Section V-B), and the very same Grounded SAM is what drives the correction (Eq. 10 and the loss in Eq. 12). So the reported gains—SDXL 34 to 59 on CoCoCount, Pixart 40 to 60—may reflect the method making images that Grounded SAM counts better, not images that people actually perceive as having the right count. In the limit, the method could be 'successful' by collapsing attention into fewer high-confidence blobs per class. That's not a minor concern; it makes the headline claim unverifiable from the paper as written. An independent evaluation, either a human study or a different detector, is required.\n\nThere are two other soft spots, both smaller. The hyperparameters (P=50, guidance steps) are selected on CoCoCount, which is also the test set, so the final numbers carry selection bias. And the one-step denoising proxy that drives the whole detection stage is defended with a single qualitative figure, not a quantitative check that the proxy layout matches the final image. Missing code and error bars make it harder to judge, but the methodological issues are the real problem.\n\nIf the authors fix the evaluation—use a held-out split, an independent count oracle, and ideally release code—the method could well work; the idea is plausible and the qualitative examples look convincing. As it stands, I wouldn't cite it or rely on the numbers. But I'd send it to peer review, because the task matters and the approach is sound enough that a referee should see it with the flaws clearly flagged. In short: promising idea, not yet credible evidence.","headline":"Clever training-free counting-correction idea, but the headline numbers are circular—the same detector drives the correction and defines the metric—so the claimed gains are unverified.","tokens_in":10990,"tokens_out":3323,"would_cite":false,"duration_ms":30550,"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":"This paper claims that a training-free detection-and-correction loop can attach to any diffusion text-to-image model and sharply improve how often generated images contain the requested number of objects.","keywords":["text-to-image synthesis","diffusion models","object counting","training-free guidance","attention map control","universal guidance","object quantity","grounded segmentation"],"falsifier":"Take a fixed set of prompts with small, overlapping, or similarly shaped objects, run the detection stage, and compare the Grounded SAM counts and object positions in the one-step preview $x_{0,\\text{mid}}$ against those in the final decoded image. If the preview and final image disagree on counts in a large share of cases where the final image is correct, then the documented correction mechanism is not what is fixing the counts; if they agree reliably, the counting signal is as claimed.","tokens_in":10076,"feed_emoji":"🔢","tokens_out":7087,"duration_ms":65286,"temperature":0.7,"pith_summary":"CountDiffusion is a training-free way to make diffusion-based text-to-image models produce the exact number of objects a prompt asks for. It first generates an early preview of the image with a single denoising step, counts objects in that preview with a grounded segmentation model, and then uses a universal-guidance loss on the model's attention maps to add missing objects or remove extras. The paper claims this two-stage correction can be attached to any diffusion text-to-image model without retraining and that it improves counting accuracy substantially, for instance raising SDXL from 34% to 59% and Pixart-Sigma from 40% to 60% on the CoCoCount dataset. A curious reader would care because getting object counts right is a known weakness of text-to-image systems, and this fix is model-agnostic and requires no user-supplied boxes or masks.","feed_headline":"Training-free loop fixes object counts in text-to-image diffusion","feed_subtitle":"It counts objects in an early denoising preview, then edits attention maps to add or remove instances.","key_machinery":"The load-bearing object is the one-step denoising preview: from a saved latent $z_{t_{\\text{mid}}}$ the method predicts $z_{0,\\text{mid}} = P(z_{t_{\\text{mid}}}, y, t)$, decodes it to an image, and lets Grounded SAM produce counts and masks. The correction step uses a universal-guidance loss on the cross-attention map $A_t$: with a mask $M$ from the detection stage, the loss is $\\pm \\frac{1}{P} \\sum \\mathrm{topk}(A_t \\odot M, P)$, taking the average of the top $P\\%$ attention values inside the region, and the sign determines whether objects are removed or added. The background is preserved by mixing the corrected latent with the saved original latent outside $M$. For multi-class prompts the organizing identity is $z_{\\text{mid}} = \\frac{1}{n} \\sum_{i=1}^{n} z_{\\text{mid},i}$, one single-class correction per class averaged at the intermediate step so the class-specific losses do not fight each other.","core_discovery":"The central claim is that the quantity signal needed for correction survives in a one-step denoising prediction, even though that prediction breaks the diffusion model's Markov-chain assumption and produces a blurry image. The paper argues that this preview still carries reliable object quantity and position information, so a counting tool can extract per-class counts and segmentation masks from it. The correction stage then rewrites the cross-attention map while denoising runs: it lowers attention inside a segmented region to remove objects, raises attention elsewhere to add objects, and smooths the attention map so the gradient lands on a continuous region rather than scattered pixels. For multi-class prompts, the method corrects each class separately and averages the resulting intermediate latents, avoiding competition between losses. The evidence reported is that this pipeline raises accurate-count generation from 34% to 59% with SDXL and from 40% to 60% with Pixart-Sigma on CoCoCount, with consistent improvements on the two newly constructed evaluation sets.","pith_inferences":["A testable extension would be to reuse the same preview-count-adjust loop for other measurable properties, such as object color, size, orientation, or spatial arrangement, by swapping the detector and the loss.","The real stress point is the one-step preview: on small, overlapping, or heavily occluded objects the preview's count is likely to diverge from the final image's count, which would make corrections miss exactly the hard cases.","Because the correction is applied locally through attention, CountDiffusion should compose naturally with layout-guided methods: a prompt-derived box layout could restrict the candidate regions and reduce false-positive detections before counting.","The fixed midpoint, $t_{\\text{mid}}=30$ of 40 steps, could become adaptive by monitoring the preview's count confidence and applying guidance only when the count signal is reliable."],"forward_implications":["Any diffusion-based text-to-image model can acquire counting ability simply by attaching CountDiffusion, so the improvement compounds as the base model improves.","Both single-class and multi-class counting accuracy improve, because the per-class latent averaging strategy prevents the correction losses for different object classes from competing.","The number of universal-guidance steps is a tunable trade-off: too few steps under-correct, and too many hurt image quality, with the optimum sitting near 30 guidance steps of 40 total in the paper's ablations.","Accuracy is bounded by the base model's and the counting model's ability, so very large requested counts remain a failure mode that the paper explicitly leaves to future work."],"supporting_citations":[{"why":"supplies the universal-guidance mechanism that back-propagates the attention loss during denoising, the core of the correction stage","marker":"[31]"},{"why":"Grounded SAM is the counting and segmentation model that produces the object counts and masks used in detection","marker":"[15]"},{"why":"CoCoCount is the public evaluation dataset and provides the CountGen comparison baseline","marker":"[34]"},{"why":"SDXL is the primary diffusion text-to-image base model the method is plugged into for experiments","marker":"[1]"},{"why":"Pixart-Sigma is the second base model used to show the framework transfers across architectures","marker":"[35]"},{"why":"defines the DDPM Markov-chain denoising assumption that the one-step preview intentionally violates","marker":"[16]"}],"fun_headline_variants":["Count objects in preview, fix attention to match","Training-free counting: count first, then adjust attention","CountDiffusion: accurate counts without retraining","One-step glance counts, attention edits fix quantity","Diffusion counts via preview, corrects via attention"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the image obtained by one-step denoising from an intermediate latent faithfully represents the object counts and positions of the final image; the paper asserts this from examples but does not measure how often it actually holds.","fun_headline_variants_meta":{"raw":{"variants":["Count objects in preview, fix attention to match","Training-free counting: count first, then adjust attention","CountDiffusion: accurate counts without retraining","One-step glance counts, attention edits fix quantity","Diffusion counts via preview, corrects via attention"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000251,"raw_usage":{"total_tokens":1544,"prompt_tokens":918,"completion_tokens":626,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":534,"completion_tokens_details":{"reasoning_tokens":553}},"tokens_in":534,"tokens_out":626,"duration_ms":6176,"temperature":1.0,"reasoning_tokens":553,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T23:30:59.807800+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a fixed set of prompts with small, overlapping, or similarly shaped objects, run the detection stage, and compare the Grounded SAM counts and object positions in the one-step preview $x_{0,\\text{mid}}$ against those in the final decoded image. If the preview and final image disagree on counts in a large share of cases where the final image is correct, then the documented correction mechanism is not what is fixing the counts; if they agree reliably, the counting signal is as claimed.","supporting_citations":[{"cited_title":"Universal guidance for diffusion models,","cited_arxiv_id":null,"evidence_quote":"supplies the universal-guidance mechanism that back-propagates the attention loss during denoising, the core of the correction stage"},{"cited_title":"Grounded sam: Assembling open-world models for diverse visual tasks,","cited_arxiv_id":null,"evidence_quote":"Grounded SAM is the counting and segmentation model that produces the object counts and masks used in detection"}],"review_version":1}