{"id":"a82d31e7-309e-4120-87d7-a8b934e30617","arxiv_id":"2412.19533","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"P3S-Diffusion generates personalized images by using one or two user clicks on a reference image to select and preserve the target subject.","lead":"This paper presents P3S-Diffusion, a method that selects which subject in a reference image to regenerate from a point click instead of a detailed mask or text description. If it works, personalized image generation becomes cheaper and more reliable for images that contain several similar objects.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Central claim depends on CLIP point-similarity mask reliably excluding a same-class distractor, but no mask-quality or selection-accuracy measurement exists; this must be tested before accepting selective subject generation.","rationale":"The paper is internally coherent and the ablations are directionally consistent: RNGR-Encoder, Attention Consistency Loss, and the timestep-based weight scheduler each shift the reported aggregate fidelity metrics in the expected direction. The reader's weakest assumption is identical to mine: mask quality for same-class distractors is unmeasured. DreamBench mostly contains single-subject images, so Table I and Table II cannot validate the selective-subject claim. The paper's own admission in Sec. III-A that CLIP is not instance-discriminative makes this gap load-bearing rather than cosmetic. The proposed concrete test directly measures the bottleneck: whether the point-derived mask actually isolates the target instance when a same-category distractor is present. If that test fails, the central contribution is unsupported and the verdict should move toward REJECT or UNVERDICTED; if it passes, the current CONDITIONAL verdict is appropriate. I am not raising a novelty objection or questioning the authors' integrity; this is an evidential gap in the central mechanism.","tokens_in":9637,"tokens_out":3259,"duration_ms":47670,"concrete_test":"Build or reuse an instance-mask benchmark (e.g., COCO images containing two instances of the same category, or pair DreamBench subjects into composite images with SAM/GroundingDINO masks). For each image, place a positive point at target-instance center and a negative point at distractor-instance center, run the Eq. 2 pipeline with the authors' stated hyperparameters, and compute target-instance IoU, distractor-instance inclusion rate, and the fraction of masks intersecting both instances. Require mean target IoU >= 0.5 and distractor inclusion <= 10% over at least 100 images; also run the full generation on this set and have evaluators (or CLIP-I/DINOv2 against the target reference) judge whether outputs match the target rather than the distractor. If the mask test fails, the selective-generation claim is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that a point pair can select one of two similar subjects (e.g., two dogs). The mechanism that must make this true is Eqs. 1-3: CLIP patch similarities from the positive point, suppressed by the negative point, Gaussian-filtered, Otsu-thresholded, and connected-component-filtered, must yield a mask isolating the target instance. This is precisely where the paper is thinnest. CLIP patch features are trained with class-level image-text contrastive loss, and the paper itself says in Sec. III-A that CLIP 'fails to accurately recognize objects with different instances of the same class.' Eq. 2 multiplies M_P by (1-M_N), which only helps if the negative-point similarity is high on the distractor and low on the target; for two similar dogs, CLIP can assign high similarity to both, either suppressing the target or letting both pass Otsu's threshold. Table II's 'W/o RNGR-Encoder' ablation uses DreamBench, whose images mostly contain a single subject, so it does not measure whether the correct instance is selected; no selection-accuracy or mask-IoU metric appears anywhere. If the mask is wrong, the inpainted image contains the distractor or neither dog, and the subsequent injection (Eqs. 4-8) has no way to recover the intended identity. Thus the central contribution is currently an unsupported inference.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"P3S-Diffusion proposes a point-supervised framework for subject-driven generation. Given a reference image and positive/negative points, an RNGR-Encoder computes patch-level CLIP similarities, suppresses the negative-point region, applies Gaussian filtering and Otsu binarization to obtain a rough mask, and inpaints the masked region. The resulting latent is injected through a trainable copy of a simplified U-Net via self- and cross-attention condition injection, with a timestep-based weight scheduler and an attention consistency loss. The paper reports DreamBench quantitative results and qualitative examples, and claims that the method can select one of two visually similar subjects using only point annotations, without pixel masks or extra segmentation models.","tokens_in":9892,"tokens_out":5549,"duration_ms":52589,"significance":"If the selective-subject claim were fully supported, the paper would make a practical contribution: point annotation is substantially cheaper than pixel masks and more precise than text descriptions for disambiguating same-class instances. The architecture is simple enough to reproduce, and the paper includes ablations of the main components on DreamBench, which is a strength. However, the decisive evaluation for the central claim is missing: the selective scenario is only shown qualitatively, and the quantitative benchmark does not measure whether the correct instance was selected. As it stands, the contribution is a plausible but unvalidated engineering proposal rather than a demonstrated capability.","major_comments":[{"comment":"The paper's central claim is that two points allow the model to select one of two similar subjects, but this claim is supported only by qualitative examples. DreamBench, used for Table I, mostly contains single-subject images and no selection-accuracy metric; the RNGR-Encoder ablation in Table II measures overall CLIP-I/DINOv2 scores, not whether the correct instance was selected. Please add a dedicated evaluation: a multi-subject benchmark with ground-truth instance masks, reporting mask IoU for Eq. (2), selection accuracy after inpainting and after generation, and a comparison against SSR-Encoder and other baselines on the same multi-subject prompts. Without such a metric, a failure of Eqs. (1)–(3) could make the entire pipeline collapse while remaining invisible in the current tables.","section":"§IV-B, Fig. 1, Fig. 4, Table I"},{"comment":"The selection mechanism relies on CLIP patch similarity to separate two instances of the same class, yet the paper itself states that CLIP is trained on class-level annotations and fails at instance-level recognition. The negative-point suppression (1−M_N) only helps if the distractor has high similarity to the negative point and low similarity to the positive point; for two similar dogs, CLIP may assign high similarity to both, either suppressing the target or letting both pass the Otsu threshold. This is an empirical assumption that is never validated. Please report positive/negative similarity distributions on multi-instance images, provide mask-quality statistics, or justify why a different patch-feature source (e.g., DINO) would not be more appropriate.","section":"§III-A, Eqs. (1)–(3)"},{"comment":"The sentence 'P3S-Diffusion basically outweighs previous method' overstates the quantitative results. In Table I, SSR-Encoder is higher than Ours on CLIP-I (0.7793 vs 0.7748), CLIP-Iv (0.0032 vs 0.0045) and DINO-v (0.0140 vs 0.0187); Ours is higher only on DINOv2 and CLIP-T. Please state the comparison precisely, report confidence intervals or significance tests across the 30 classes, and avoid the claim of overall superiority unless it is statistically supported.","section":"§IV-B, Table I"},{"comment":"The ablation 'W/o RNGR-Encoder' is not informative for the selective-subject claim because DreamBench does not contain the two-similar-subjects scenario. Removing RNGR-Encoder may degrade single-subject fidelity while leaving the selective capability entirely untested. All ablations should be repeated on the proposed multi-subject benchmark so that the contribution of each component to correct selection can be assessed.","section":"§IV-C, Table II"}],"minor_comments":[{"comment":"The notation Cos(patch_{x′,y′}, patch_{i,j}) is undefined; please specify which patch contains the point p, how point coordinates map to patch indices, and how the similarity values are arranged into the matrix M.","section":"§III-A, Eq. (1)"},{"comment":"The text describes removing outliers that are not connected to the specified points, but this connected-component filtering is not present in Eq. (2). Please include it in the equation or define it as a separate preprocessing step.","section":"§III-A, Eq. (2)"},{"comment":"The metrics are described inconsistently: the text mentions 'DINO Scores [30]' but then refers to 'GroundingDINOv2', and the reference [30] is the DINOv1 paper. Please clarify which model is used, which layer/patch tokens are compared, and give a citation for the exact DINOv2 or GroundingDINO variant.","section":"§IV-A, Evaluation Metrics"},{"comment":"The normalization in Lac is unclear: avg is defined as the average of the sums of the two attention maps, not the average of the maps themselves. Please clarify the intended normalization and why this particular scaling was chosen.","section":"§III-B, Eq. (11)"},{"comment":"The timestep scheduler is controlled by hyperparameters α=0.5, β=0.2, k=2, but no sensitivity study is reported; Table III only tests two hand-picked curves. A small sweep or a statement of robustness across these parameters would strengthen the claims about balancing identity and editability.","section":"§IV-C, Table III and Fig. 5"},{"comment":"The limitation that the method 'does not perform well on non-salient images' is not defined or quantified. Please specify what 'non-salient' means and provide examples or failure statistics.","section":"§V, Conclusion"}],"recommendation":"major_revision","confidential_remarks":"The core gap is the absence of any quantitative evaluation of the selective-subject scenario, which is the paper's main advertised capability. I would ask the authors to add a multi-instance benchmark with mask/selection metrics and to correct the overstated comparison in Table I before reconsidering the manuscript. The engineering design is reasonable and potentially useful, so I would not desk-reject on this basis."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my take. What's actually new is the point-supervision interface for selective subject-driven generation: a positive and a negative click, turned into a coarse mask via CLIP patch similarity, negative suppression, Otsu, and connected components, then inpainted and injected into a trainable U-Net copy with multi-layer self-attention conditioning plus an attention consistency loss. That specific combination is not in the cited prior work; SSR-Encoder uses text or masks. The ablations are directionally consistent and the training/inference details are concrete. The paper deserves credit for proposing a lower-annotation-cost interface.\n\nThe soft spots are exactly where the stress-test lands. The central advertised capability—picking one of two similar dogs—is only demonstrated qualitatively in Figs. 1 and 4. There is no quantitative selection-accuracy or mask-IoU metric. Table I reports CLIP-I 0.7748, slightly under SSR-Encoder's 0.7793, yet the text says the method 'basically outweighs previous method'; that is a misreading. No error bars appear anywhere. Hyperparameters (alpha, beta, k, lambda, gamma) are empirically tuned on DreamBench, and DreamBench mostly has single-subject images, so Table II's 'W/o RNGR-Encoder' ablation does not test the selective case. The mechanism itself is plausible but not guaranteed: CLIP patch features are class-level, and the paper admits CLIP fails at instance discrimination; Eq. (2) relies on the negative similarity being high only on the distractor, which is exactly what needs validation. The failure mode is not just cosmetic—if the mask is wrong, the whole pipeline degenerates. The paper's own conclusion lists limitations (rich detail, non-salient images) that are honest but partial.\n\nNone of this is fatal to the idea. It is an engineering contribution with a real gap between claim and evidence. A serious referee could fix this by asking for code, error bars over seeds, and a dedicated two-similar-subjects benchmark with mask and selection metrics. I would send it to review, not desk-reject. I might not cite the selection claim until that evidence exists, but I would cite the point-supervision interface as an alternative annotation paradigm.\n\nVerdict: worth referee time, likely major revision.","headline":"Point-supervised subject selection is genuinely new, but the central claim—reliable selection among similar instances—is not yet quantitatively supported.","tokens_in":10454,"tokens_out":2806,"would_cite":true,"duration_ms":26901,"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":"P3S-Diffusion shows that a single point-click in a reference image is enough to tell a diffusion model which subject to generate, even when a nearly identical distractor sits in the same photo.","keywords":["subject-driven generation","point supervision","diffusion models","selective subject representation","RNGR-Encoder","attention consistency loss","image personalization","CLIP patch similarity"],"falsifier":"Create reference images containing two same-breed dogs of different individual identity, click one as positive and the other as negative, and measure the RNGR-Encoder mask's intersection-over-union against a human-annotated segmentation of the clicked dog; if the mask routinely includes or swaps in the negative dog (IoU clearly below, say, 0.5), the selective-selection claim fails. A second check: run the pipeline with the negative point moved from the distractor to plain background; if the selected subject changes dramatically, the selection is driven mainly by the negative-point suppression term, localizing the mechanism.","tokens_in":1903,"feed_emoji":"🎯","tokens_out":2285,"duration_ms":71652,"temperature":0.7,"pith_summary":"This paper proposes P3S-Diffusion, a method for subject-driven image generation in which the user selects the subject of a reference image with a single point (plus one negative point), rather than writing a detailed text prompt or drawing a pixel mask. The paper's central claim is that this minimal annotation is enough: the model derives an approximate mask from CLIP patch-level similarities, inpaints away any similarly looking distractor (for example, a second dog), and then injects the selected subject's features into a frozen text-to-image diffusion model through a trainable U-Net copy. If the claim holds, selective subject generation—knowing which of several similar objects to copy—becomes nearly free in annotation cost and can be combined with other controllers such as ControlNet. The method is evaluated on DreamBench with CLIP-I, CLIP-T, and DINOv2 metrics; the reported numbers place it at or near the top in subject fidelity while retaining diversity.","feed_headline":"One click tells an AI image model which subject to copy","feed_subtitle":"Point-guided P3S-Diffusion separates two similar dogs without pixel masks or long text prompts.","key_machinery":"The load-bearing mechanism is the RNGR-Encoder plus the multi-layer condition injection. RNGR-Encoder turns one positive and one negative click into a rough binary mask via Eq. 1–3: patch cosine similarity from CLIP features, suppression by the negative map, Gaussian filtering, Otsu binarization, and connected-component cleanup, followed by inpainting of the masked region. The cleaned image is then encoded and cross-attended with CLIP's last hidden state (Eq. 4–5) to bias the latent toward the selected subject. Multi-layers Condition Injection copies a simplified U-Net and delivers its self-attention hidden features into the original U-Net's self-attention layers by concatenation and addition with a small weight $\\lambda=0.2$ (Eq. 6–8), using zero convolutions; the timestep-based weight scheduler $\\epsilon_t = 1 - \\alpha(t/T)^k + \\beta$ adjusts control strength over the denoising trajectory, and the attention consistency loss $L_{ac}$ (Eq. 11–13) matches the last cross-attention map of the trainable copy to the frozen model's map.","core_discovery":"On the paper's own terms, P3S-Diffusion's discovery is that point supervision is a sufficient interface for selective subject-driven generation. Given a reference image I and a point p on the target subject, the RNGR-Encoder computes the cosine similarity between the patch containing p and every other patch in CLIP's image feature map, suppresses it with the similarity to a user-provided negative point, Gaussian-smooths, Otsu-binarizes, and keeps the connected component around p to produce a rough mask. The mask is then inpainted, so the negative subject disappears; the completed image is encoded to latent form and cross-attended with CLIP features, and the result is fed through a trainable simplified U-Net whose hidden self-attention states are concatenated into the frozen original U-Net. A timestep-based weight scheduler reduces control early in denoising to preserve editability, and an attention consistency loss aligns the trainable copy's cross-attention with the frozen model's. The paper reports that this lets the model faithfully generate the clicked subject while ignoring a visually similar second subject in the same reference image, without any segmentation model.","pith_inferences":["The approximate mask produced by point clicks could be used as a free by-product for other tasks, such as weakly supervised segmentation or dataset annotation, since it requires no segmentation model and only two clicks.","Because CLIP is explicitly trained at class level, the method's success on two similar dogs likely depends more on the negative-point suppression and the connected-component cleanup than on CLIP's fine discrimination; stress tests with identical twins or same-breed dogs would expose the true ceiling.","The same attention-consistency idea could stabilize other multi-adapter or multi-subject conditioning schemes, where a trainable adapter drifts from the frozen base model during fine-tuning.","Nothing in the architecture ties the mask to diffusion; the RNGR-Encoder alone could serve as a point-to-mask proposer for any image-editing pipeline that needs a rough subject mask."],"forward_implications":["A point click can replace pixel masks and per-subject text descriptions for subject-driven generation, cutting annotation cost.","The pipeline can be applied to frozen diffusion backbones, and the paper states it is compatible with ControlNet-style controllers.","The timestep-based weight scheduler lets one trained model trade identity fidelity against prompt editability at inference time by changing $\\alpha$, $\\beta$, and $k$.","Attention consistency loss on the last cross-attention layer improves subject fidelity; applying it to all layers causes underfitting and worse scores, so the paper recommends the last layer only.","On DreamBench the method reports the best DINOv2 score among compared methods and competitive CLIP-I, meaning it preserves fine subject features while keeping diversity."],"supporting_citations":[{"why":"Supplies the CLIP image encoder whose patch features drive the point-patch similarity mask, and the text encoder used by the base diffusion model.","marker":"[4]"},{"why":"Provides the Stable Diffusion base model, the latent denoising objective, and the inpainting model used to fill the masked distractor region.","marker":"[7]"},{"why":"Establishes the trainable copy of the U-Net and zero-convolution injection scheme that Multi-layers Condition Injection adapts.","marker":"[2]"},{"why":"Supplies the DreamBench dataset and evaluation protocol, the unique-identifier training trick, and the main per-subject fine-tuning baseline DreamBooth.","marker":"[17]"},{"why":"Is the closest selective-subject-encoding baseline that the method compares against and extends with point supervision.","marker":"[3]"},{"why":"Provides DINOv2, the self-supervised ViT features used as an evaluation metric for subject fidelity.","marker":"[30]"},{"why":"Motivates the timestep-based weight scheduler with the earlier finding that early denoising stages fix the rough image structure.","marker":"[23]"},{"why":"Supports the choice of self-attention and cross-attention layers as the most impactful injection points, and is a comparison baseline.","marker":"[18]"}],"fun_headline_variants":["Point at the subject, AI copies it precisely","One click chooses which similar subject to copy","Point supervision beats masks for subject selection","Click a point, AI ignores the similar twin"],"cache_read_input_tokens":12544,"weakest_assumption_plain":"The method assumes that the positive-point patch similarity, after negative-point suppression and Otsu binarization, reliably isolates the clicked subject and not the visually similar distractor; if CLIP's class-level features cannot separate the two, the mask and everything after it collapses.","fun_headline_variants_meta":{"raw":{"variants":["Point at the subject, AI copies it precisely","One click chooses which similar subject to copy","Point supervision beats masks for subject selection","Click a point, AI ignores the similar twin"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00064,"raw_usage":{"total_tokens":2954,"prompt_tokens":960,"completion_tokens":1994,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":576,"completion_tokens_details":{"reasoning_tokens":1938}},"tokens_in":576,"tokens_out":1994,"duration_ms":15287,"temperature":1.0,"reasoning_tokens":1938,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T00:14:17.347734+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Create reference images containing two same-breed dogs of different individual identity, click one as positive and the other as negative, and measure the RNGR-Encoder mask's intersection-over-union against a human-annotated segmentation of the clicked dog; if the mask routinely includes or swaps in the negative dog (IoU clearly below, say, 0.5), the selective-selection claim fails. A second check: run the pipeline with the negative point moved from the distractor to plain background; if the selected subject changes dramatically, the selection is driven mainly by the negative-point suppression term, localizing the mechanism.","supporting_citations":[{"cited_title":"Adding conditional control to text-to-image diffusion models","cited_arxiv_id":null,"evidence_quote":"Establishes the trainable copy of the U-Net and zero-convolution injection scheme that Multi-layers Condition Injection adapts."},{"cited_title":"Ssr-encoder: Encoding selective subject representation for subject-driven generation, 2024","cited_arxiv_id":null,"evidence_quote":"Is the closest selective-subject-encoding baseline that the method compares against and extends with point supervision."},{"cited_title":"Emerging properties in 9 self-supervised vision transformers","cited_arxiv_id":null,"evidence_quote":"Provides DINOv2, the self-supervised ViT features used as an evaluation metric for subject fidelity."},{"cited_title":"Multi-concept customization of text-to-image diffusion","cited_arxiv_id":null,"evidence_quote":"Supports the choice of self-attention and cross-attention layers as the most impactful injection points, and is a comparison baseline."}],"review_version":1}