{"id":"2aaa5fb8-d598-4a1f-be1a-783453058dd5","arxiv_id":"2508.08498","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":8.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"CObL uses multiple linked Stable Diffusion models to decompose an image into occlusion-ordered object layers, guided at inference so the layers reproduce the input.","lead":"This paper introduces CObL, a method that splits a single photo into an ordered stack of object pictures, each with hidden parts filled in, without needing user prompts or knowing how many objects are present. CObL is trained on synthetic tabletop scenes and appears to work on real photographs of tabletops, which could enable automatic photo editing and object-level scene understanding.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central claim is under-supported because the compositing-equivalence class makes layer stacks non-identifiable; the paper's own S3 admits many decompositions composite to the same image, and the reported LPIPS/CLIP metrics do not isolate amodal semantic correctness. A direct amodal IoU / human-j","rationale":"The reader's weakest assumption identifies the same load-bearing concern: the layer decomposition is underdetermined by the compositing loss, and the learned prior must select the semantically correct stack. This is indeed the most serious threat to the central claim because it attacks the meaning of the representation itself. The paper is honest about the equivalence class in Sec. S3 and shows a concrete example, which supports the reader's interpretation. I do not think this warrants rejection: the model is trained on synthetic ground-truth layers, so the prior can in principle resolve the ambiguity within the training distribution, and the reported qualitative results show plausible completions. However, the quantitative evidence is not decisive. LPIPS and CLIP are not targeted amodal correctness metrics; they do not measure whether the occluded shape and appearance are right. The 'best-of-seeds' reporting in Table 1 further weakens the quantitative case. The concrete test I propose would settle whether the prior actually selects the true layers: high amodal IoU on held-out synthetic scenes would demonstrate that the representation is identifiable in-distribution, while a human-study on real images would test zero-shot transfer. If those tests pass, the central claim is substantiated; if not, the 'zero-shot' claim should be softened. Therefore the reader's CONDITIONAL verdict is appropriate, and my read does not change it.","tokens_in":14270,"tokens_out":4448,"duration_ms":58035,"concrete_test":"On the 250 held-out synthetic validation scenes (exact ground-truth layers available) and on a random 30-scene subset of TABLETOP, compute amodal mask IoU and boundary F-score between each predicted alpha layer and its Hungarian-matched ground-truth layer, while also recording the final reconstruction error ||I - composite(x_hat)||. Then compare CObL's selected stack against a baseline that randomly permutes or replaces layers with any other stack that composites to the same image (e.g., by moving visible pixels between layers). If CObL's amodal IoU is not significantly above this compositing-equivalent baseline, or if multiple seeds with equally low reconstruction error have IoU varying by more than 0.2, then the prior is not resolving the underdetermination and the zero-shot claim is unsupported. For real images, supplement with human raters judging whether occluded regions are correct","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that CObL infers the correct occlusion-ordered, amodally-completed object stack from a single real image. This requires the inference-time losses to choose a semantically correct member of the many stacks that composite to the same input. The paper's Sec. S3 explicitly states: 'regardless of if a decomposition is accurate according to human perception, the object layers will generally composite back to the original scene,' and Fig. S3 shows two different stacks producing identical composites. The compositional loss Lc in Eq. (7) is invariant to this equivalence class, and the prior score matching term Lpsm in Eq. (8) only encourages each layer to be a natural image independently, not to be the true object decomposition. The model's only training signal that prefers the ground-truth layer structure comes from 2000 synthetic scenes; whether that transfers to novel real objects is exactly the zero-shot claim at issue. The reported evaluation uses LPIPS and CLIP after Hungarian matching, which reward visible-region fidelity and generic appearance similarity but do not isolate whether occluded portions are semantically correct. Combined with the acknowledged 'merging'/'splitting' failure modes (Sec. S4) and best-of-seeds reporting, the evidence does not yet rule out that outputs are arbitrary compositing-consistent stacks rather than true object layers.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces CObL, a diffusion-based architecture that takes a single image and jointly generates an occlusion-ordered stack of amodally-completed object layers (RGBA with binary alpha), without user prompting and without prior knowledge of the number of objects. The model consists of several frozen Stable Diffusion UNets tied by learnable lateral attention and an input-conditioning adapter. It is trained on 2000 synthetic tabletop scenes generated via a novel pipeline combining Blender geometry, ControlNet-depth textures, and composited shadows. At inference, compositional guidance (Eq. 7) forces the stack to composite back to the input, and a prior-score-matching loss (Eq. 8) keeps each layer within the SD natural-image distribution. The authors evaluate on a new real-world dataset, TABLETOP (100 images), reporting LPIPS/CLIP scores against prior amodal-completion/inpainting models and ARI against Mask2Former. The central claim is that the model zero-shot generalizes to real tabletop photographs containing novel objects and variable object counts.","tokens_in":14594,"tokens_out":5359,"duration_ms":71701,"significance":"If the central claim is upheld, CObL is a meaningful step toward open-world perceptual organization: it addresses the full amodal, multiple-object, occlusion-ordered decomposition problem without any per-image prompting, and it demonstrates a practical way to adapt a large pretrained generative prior to a structured scene representation with limited synthetic data. The paper's strengths include a concrete and reproducible pipeline, a new benchmark (TABLETOP) with ground-truth layers, explicit reporting of both best and average runs in some comparisons, and a candid discussion of non-uniqueness and failure modes in the supplementary material. The central risk is that the evaluation does not cleanly separate 'true' amodal object decomposition from arbitrary compositing-equivalent stacks, and the use of best-of-seeds selection for headline numbers tempers confidence.","major_comments":[{"comment":"The LPIPS/CLIP metrics on the full layer images do not isolate amodal semantic correctness. Because the compositional loss Lc only enforces equality of the composite, and Sec. S3 explicitly states that many layer stacks composite to the same input (Fig. S3), a model could produce arbitrary occluded-region content while scoring well on visible-region fidelity. Table 1 therefore does not establish that the inferred layers are the true amodally-completed objects rather than some other compositing-consistent stack. The authors should add amodal-mask IoU (e.g., comparing alpha masks over occluded regions), separate visible/occluded-region metrics, or a human perceptual study of completion correctness.","section":"§5.1, Eq. (7), Sec. S3"},{"comment":"The paper reports best-of-four-seed results as the primary numbers (Table 1) and uses top-1 for the ARI claim (Sec. 5.2). Sec. S5 acknowledges that CObL's outputs are highly initialization-dependent and non-convex, and that in practice the authors 'choose the most likely output' without specifying a formal selection rule. This best-run selection can systematically overstate the method's reliability, which is load-bearing for the zero-shot generalization claim. The authors should report the full distribution over seeds (with variance or box plots) and define and validate a selection criterion (e.g., lowest compositional loss) that is shown to correlate with human quality.","section":"§5.1, §5.2, Sec. S5"},{"comment":"The inference-time guidance terms Lc and PSM do not themselves prefer the ground-truth object structure: Lc is invariant to merging/splitting and occlusion-order changes within the equivalence class, and PSM only encourages each layer to be a natural image. The only signal that prefers the true decomposition is the synthetic training objective (Eq. 3). Given the acknowledged non-uniqueness, the paper should directly quantify how often the learned decomposition transfers to real data: e.g., report merging/splitting rates on TABLETOP (as in Sec. S4) and show that they are not the dominant mode, and ideally provide a per-scene analysis across object counts, especially for the 5-6 object cases where performance degrades (Fig. 7). Without this, aggregate LPIPS/CLIP may hide a model that composites well but groups pixels arbitrarily.","section":"§3.3, Eq. (3), Sec. S3"},{"comment":"The comparison to inpainting/amodal-completion baselines is asymmetric: CObL is specifically trained on synthetic tabletop scenes, while baselines are frozen generic models, even though the baselines receive oracle masks. This asymmetry makes the quantitative comparison hard to interpret as a measure of the full task. At minimum, the paper should also compare against an unsupervised object-centric learning baseline (e.g., a slot-attention or DINO-based method) on the same TABLETOP benchmark, or ablate CObL's own components (as in Sec. 6) to isolate the contribution of the synthetic training prior. The current Table 1 does not let a reader assess whether the advantage comes from the architecture, the training data, or the evaluation protocol.","section":"§5.1, Table 1"}],"minor_comments":[{"comment":"The arXiv identifier for Stable Video Diffusion is mistyped as 'arXiv:311.15127' (should be 2311.15127), and reference [6] contains a stray page number '1' at the end. Please correct these bibliographic entries.","section":"Sec. 2, references [5] and [6]"},{"comment":"There are minor typos: 'We we find' in Sec. 5.3, 'LPIPs' in Sec. 6, and inconsistent capitalization 'CoBL' vs 'CObL' in Sec. 5.2. These should be fixed in a final polish.","section":"Sec. 5.3 and Sec. 6"},{"comment":"The abstract claims CObL is 'not limited to the world it was trained in,' but the evaluation is confined to tabletop scenes. This is acceptable given the paper's scope, but the wording should be softened to avoid overclaiming, especially since Sec. 7 itself restricts the conclusions to tabletops.","section":"Abstract and Sec. 7"}],"recommendation":"major_revision","confidential_remarks":"The core idea is interesting and the engineering is sound, but the paper needs to address the identifiability/underdetermination concern more directly. The stress-test concern about compositing-equivalent stacks is real, but it is not a circularity issue; the training signal in Eq. (3) does prefer the ground-truth decomposition. The main problem is that the reported metrics are not sensitive enough to distinguish true amodal completion from arbitrary compositing-consistent outputs. I would like to see amodal IoU or a human study, plus seed-variance reporting and a defined selection rule. The paper is publishable after those additions. I would not reject it, because the limitations are acknowledged and the proposed fix is within the scope of a revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know about CObL. First, it does something new: it infers a whole occlusion-ordered stack of amodally-completed layers from a single real image, with no prompt and no known object count. Second, the central zero-shot claim is real but not yet fully supported; the paper's own supplement admits many layer stacks composite to the same image, and the metrics don't isolate whether the occluded parts are semantically right.\n\nThe strengths: the task definition is valuable, the pipeline of generating 2250 synthetic tabletop scenes with ControlNet and depth maps is resourceful, and the concurrent-UNet architecture with lateral attention plus compositional and PSM guidance is a sensible way to use Stable Diffusion as a prior. The ablations show each component matters. The authors are also honest about merging/splitting failures and the need to use best-of-seed sampling.\n\nThe soft spots: the identifiability problem is load-bearing. In S3 they write that 'regardless of if a decomposition is accurate according to human perception, the object layers will generally composite back to the original scene.' Since L_c only checks reconstruction and PSM only checks per-layer naturalness, the only bias toward the true decomposition comes from 2000 synthetic examples. The LPIPS/CLIP metrics after Hungarian matching reward visible fidelity and generic appearance, not true amodal shape. Best-of-seeds reporting over four runs adds uncertainty. A direct amodal IoU on occluded regions or a human study would be the right fix. The comparison with inpainting baselines is favorable but they get oracle masks, so it's a lower bound, not a proof of semantic correctness.\n\nNet: the paper deserves a serious referee. It defines a new task, ships a new dataset, and the method is clearly better than nothing and better than prompted baselines. But the abstract oversells the zero-shot generalization. I'd suggest a revise-and-resubmit that foregrounds identifiability, releases code/data, and adds a metric that targets occluded regions. If I worked on object-centric perception or amodal completion I'd cite it.","headline":"A genuinely new task and a cleverly engineered diffusion pipeline, but the evaluation doesn't yet nail the amodal-semantics claim because the supplement concedes the decomposition is non-identifiable from the image alone.","tokens_in":15002,"tokens_out":2806,"would_cite":true,"duration_ms":30854,"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":"CObL is a diffusion-based model that, trained only on a few thousand synthetic tabletop scenes, zero-shot decomposes a single real photograph into an occlusion-ordered stack of amodally completed object layers, with no user prompting and no","keywords":["object layers","amodal completion","zero-shot generalization","occlusion ordering","diffusion models","scene decomposition","synthetic-to-real transfer","tabletop scenes"],"falsifier":"Run CObL many times on the same real tabletop photo from different random noise initializations, keep only the runs whose composited output matches the input nearly perfectly, and have human raters judge whether the extracted layers correspond to actual objects and occlusion order. If equally-compositing outputs disagree on object boundaries, object counts, or stacking order at a rate near chance, then the compositing constraint alone is satisfied by arbitrary stacks and the learned prior is not resolving the underdetermination.","tokens_in":14196,"feed_emoji":"🧩","tokens_out":6867,"duration_ms":69613,"temperature":0.7,"pith_summary":"The paper tries to establish that perceptual organization into objects and their depth ordering can be done zero-shot by a model that never sees real training data. CObL takes one photograph of a tabletop scene and concurrently generates a stack of RGB-alpha layers, each containing one object completed behind occluders, ordered from background to front. The model uses several frozen copies of Stable Diffusion as a natural-object prior, tied together by learned cross-layer attention, and is trained on about two thousand synthetic scenes made by combining 3D geometry with generated textures. At inference, guidance forces the layers to composite back to the input photo while a prior-score-matching term keeps each layer natural. The reported results indicate generalization to real tabletop photos containing novel objects and varying object counts, which would matter because it suggests amodal scene decomposition can be learned in a closed synthetic world and transferred without per-object prompts.","feed_headline":"Zero-shot model turns one photo into occlusion-ordered object layers","feed_subtitle":"Trained on a few thousand synthetic tabletop renders, it amodally completes unseen objects in real photos.","key_machinery":"The central device is a bank of N concurrently denoising latent diffusion UNets: N frozen copies of Stable Diffusion 2.1, each responsible for one object layer, connected by learnable lateral cross-attention so layers can communicate while being generated together. The other load-bearing mechanism is the guidance loss $L_g = L_c + \\lambda L_{\\mathrm{psm}}$, where $L_c$ is the compositing loss forcing the decoded layers, composited back-to-front, to match the input image, and $L_{\\mathrm{psm}}$ is the prior score matching loss that keeps each layer close to Stable Diffusion's native distribution. The fixed-size stack with empty layers is what lets the model handle unknown object counts.","core_discovery":"The central claim is that an occlusion-ordered stack of amodal object layers—not just visible segmentation—can be inferred from a single image by a concurrent diffusion process. CObL instantiates N frozen Stable Diffusion UNets, one per layer, coupled by trainable lateral attention and conditioned on the input image through a lightweight adapter that first estimates depth with MiDaS. Training uses only 2,250 synthetic tabletop scenes assembled from 600 3D assets, with 2,000 for training; the layers are padded with empty layers so the stack has fixed size. During DDIM sampling, the model applies compositional guidance, which measures the $\\ell^2$ error between the input and the composite of t","pith_inferences":["Because the paper itself notes in Sec. S3 that different layer stacks can composite to the same image, the real test of 'correct' layering is whether the prior-score-matching term, not the compositing loss, picks human-plausible objects; an experiment that varies random seeds and asks human raters to choose among equally-compositing stacks would separate those two effects.","The fixed-size stack with empty layers effectively lets the model discover object count from the image; a natural extension is to read off a count from the alpha channels and to evaluate count accuracy separately from shape accuracy.","The synthetic-data pipeline decouples geometry from texture, so the same machinery should transfer to other scene geometries such as shelves or ground scenes as long as a matching depth-conditioned generator is available; the paper explicitly says the idea extends beyond tabletops.","Merging and splitting failures suggest the unit of decomposition is dictated as much by the Stable Diffusion prior as by the image; using a prior trained on objects rather than scenes might reduce these failures, and measuring the rate of merge/split errors directly would be a sharper metric than LPIPS."],"forward_implications":["Amodal completion can be performed for all objects in a scene at once, without user-provided masks or prompts, whereas current alternatives complete one object at a time and require an oracle mask.","A model trained only on synthetic tabletop renders transfers to real photographs with novel objects and variable object counts, indicating that a generative prior plus compositing guidance can bridge a substantial sim-to-real gap.","Because output layers composite back to the input by construction, visible regions such as text and fine texture are preserved instead of hallucinated, a property the paper demonstrates against inpainting baselines.","Performance degrades when scenes contain more than four objects, so the method's practical range is limited by clutter and by the memory cost of running many UNets.","The model also produces a valid visible panoptic segmentation as a by-product, reaching a top-1 ARI of 83.5% on the introduced TABLETOP dataset without finetuning."],"supporting_citations":[{"why":"Stable Diffusion's frozen UNet is the generative backbone; each output layer is denoised by a copy of it.","marker":"[48]"},{"why":"ControlNet-depth generates per-layer images from rendered depth maps and prompts, producing the synthetic training textures.","marker":"[59]"},{"why":"Supplies the lateral cross-attention design that ties the concurrent UNets together so layers can communicate.","marker":"[6]"},{"why":"One of the guidance sources for enforcing constraints during diffusion sampling, used to shape the compositional guidance.","marker":"[30]"},{"why":"Score distillation sampling provides the intuition behind the prior score matching loss that keeps latents in Stable Diffusion's native distribution.","marker":"[41]"},{"why":"MiDaS monocular depth estimation gives the conditioning adapter its pseudo-depth map of the input scene.","marker":"[46]"},{"why":"T2I-Adapter provides the adapter mechanism that injects image and depth cues into each frozen UNet.","marker":"[37]"},{"why":"U2-Net foreground segmentation supplies the masks used during inference-time compositing guidance.","marker":"[44]"},{"why":"pix2gestalt is the amodal completion baseline that CObL must beat; it requires an oracle mask while CObL does not.","marker":"[38]"},{"why":"LaMa is the inpainting baseline used for comparison; it also receives an oracle mask of the occluded region.","marker":"[51]"}],"fun_headline_variants":["Zero-shot amodal layers from one photo, no user prompts","Synthetic-only training yields real-world occlusion layers zero-shot","One photo in, ordered amodal layers out—no prompts, no counts","Diffusion model turns a single image into an occlusion-ordered stack"],"cache_read_input_tokens":2816,"weakest_assumption_plain":"The central claim depends on the guidance and learned prior picking out a human-plausible object decomposition, since many different layer stacks can composite to the exact same input image and the paper acknowledges that its layers will generally recomposite correctly even when the decomposition is wrong.","fun_headline_variants_meta":{"raw":{"variants":["Zero-shot amodal layers from one photo, no user prompts","Synthetic-only training yields real-world occlusion layers zero-shot","One photo in, ordered amodal layers out—no prompts, no counts","Diffusion model turns a single image into an occlusion-ordered stack"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000352,"raw_usage":{"total_tokens":1740,"prompt_tokens":715,"completion_tokens":1025,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":459,"completion_tokens_details":{"reasoning_tokens":963}},"tokens_in":459,"tokens_out":1025,"duration_ms":11211,"temperature":1.0,"reasoning_tokens":963,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T21:31:03.497107+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run CObL many times on the same real tabletop photo from different random noise initializations, keep only the runs whose composited output matches the input nearly perfectly, and have human raters judge whether the extracted layers correspond to actual objects and occlusion order. If equally-compositing outputs disagree on object boundaries, object counts, or stacking order at a rate near chance, then the compositing constraint alone is satisfied by arbitrary stacks and the learned prior is not resolving the underdetermination.","supporting_citations":[{"cited_title":"Adding conditional control to text-to-image diffusion models","cited_arxiv_id":null,"evidence_quote":"ControlNet-depth generates per-layer images from rendered depth maps and prompts, producing the synthetic training textures."},{"cited_title":"Align your latents: High-resolution video synthesis with la- tent diffusion models","cited_arxiv_id":null,"evidence_quote":"Supplies the lateral cross-attention design that ties the concurrent UNets together so layers can communicate."},{"cited_title":"Syncdiffusion: Coherent montage via synchronized joint diffusions","cited_arxiv_id":null,"evidence_quote":"One of the guidance sources for enforcing constraints during diffusion sampling, used to shape the compositional guidance."},{"cited_title":"Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer","cited_arxiv_id":null,"evidence_quote":"MiDaS monocular depth estimation gives the conditioning adapter its pseudo-depth map of the input scene."},{"cited_title":"T2i- adapter: Learning adapters to dig out more controllable abil- ity for text-to-image diffusion models","cited_arxiv_id":null,"evidence_quote":"T2I-Adapter provides the adapter mechanism that injects image and depth cues into each frozen UNet."},{"cited_title":"U2-net: Going deeper with nested u-structure for salient object detection","cited_arxiv_id":null,"evidence_quote":"U2-Net foreground segmentation supplies the masks used during inference-time compositing guidance."},{"cited_title":"pix2gestalt: Amodal segmentation by synthesizing wholes","cited_arxiv_id":null,"evidence_quote":"pix2gestalt is the amodal completion baseline that CObL must beat; it requires an oracle mask while CObL does not."},{"cited_title":"Resolution-robust large mask inpainting with fourier convolutions","cited_arxiv_id":null,"evidence_quote":"LaMa is the inpainting baseline used for comparison; it also receives an oracle mask of the occluded region."}],"review_version":1}