{"id":"70ebbca6-6bca-446c-996a-fb606808a54f","arxiv_id":"2607.08765","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Geometry-aware RGB–depth pretraining plus a 1M multi-task panoramic dataset yields a unified in-context 360° generator with stronger FAED and seam consistency than prior methods.","lead":"Canvas360 is a two-stage system that first pretrains a panoramic generator with parallel RGB–depth flow matching, then fine-tunes one model on a million paired samples for style transfer, inpainting, outpainting, and editing. It matters for VR/AR and immersive content because it targets the geometric distortions and seam breaks that still break most 360° generators.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"Pseudo-depth + synthetic pairs may not transfer spherical geometry once depth is dropped at fine-tuning.","rationale":"The reader correctly isolates the weakest link: reliance on DAP pseudo-depth and non-panorama-native synthetic pairs, plus the claim that geometry survives after depth is dropped. That is load-bearing for the “geometry-aware pretraining → unified in-context gains” narrative; the FAED lead, LRCE-RGB, and distortion-aware edits are the main positive evidence, but they do not isolate geometry transfer from data scale or backbone strength. My concrete test directly falsifies or supports that isolation. No stronger internal contradiction appears (methods are coherent, ablations exist, user study and multi-metric tables are present). Therefore the verdict stays CONDITIONAL—public data/code and an external geometry-isolation check would be needed for unconditional acceptance—rather than moving to REJECT or ACCEPT. Agreement with the reader is full on the weakest assumption; the stress-test only sharpens the same concern into a single decisive experiment.","tokens_in":23357,"tokens_out":675,"duration_ms":6491,"concrete_test":"Train three matched LoRA runs on identical 900K RGB pairs: (A) full Canvas360 pretrained with DAP depth + VCP + L_sim, (B) same pretraining but with depth maps replaced by random noise of matching statistics, (C) no geometry pretraining (FLUX.1-dev baseline). Evaluate FAED, LRCE-RGB, and a distortion-consistency score on held-out real Matterport edits/outpaints (not synthetic). If (A) does not beat (B) and (C) by a clear margin on FAED/LRCE and qualitative ERP distortion of added objects, the transfer claim fails.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim is that geometry-aware pretraining (parallel RGB–depth generation with DAP pseudo-depth, velocity circular padding, and L_sim) installs spherical priors that survive depth removal and improve FAED / seam consistency / distortion-aware edits across four unified in-context tasks. The least secure step is the transfer assumption in Secs. 3.2–3.3 and App. B: DAP depths are clipped (100 m outdoor / 10 m indoor) and noisy under ERP, and the bulk of the 900K pairs (style/edit especially) are produced by non-panorama-native models (FLUX.2-dev, NanoBanana) then lightly filtered. After depth is discarded, the model is trained only on appearance-level token concatenation; nothing rigorously shows that residual spherical geometry—not just better RGB texture statistics—is what drives the FAED lead and the qualitative “correct panoramic distortion” on edits. Ablations (Tab. 3 / Tab. 7 / Fig. 11) improve FAED with depth but also show instability (dark outputs) and mixed FID-family metrics; the cleaned-vs-random 20K edit check (Tab. 5) only tests residual noise, not geometric fidelity of the teacher. If the priors are mostly appearance, the two-stage story overclaims relative to simply fine-tuning a strong backbone on the same 1M pairs.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"Canvas360 is a two-stage framework for in-context panoramic generation. Stage 1 pretrains a Flow Transformer (FLUX.1-dev + LoRA) on ~100K RGB–depth ERP panoramas via parallel RGB–depth generation (token-wise concatenation with RoPE offset Td), velocity circular padding for 0°/360° continuity, and a squared-correlation similarity loss L_sim that discourages RGB–depth collapse. Stage 2 discards depth and fine-tunes a single model on a new 1M-scale Canvas360Dataset (style transfer, inpainting, outpainting, editing) using the same token-concatenation recipe with context offset Tc. The paper reports leading FAED and competitive FID-family / perceptual metrics on text-to-panorama (Table 1), strong LPIPS/FAED/PSNR on inpainting–outpainting–editing (Tables 9–10), LRCE-RGB boundary gains (Table 6), ablations (Tables 3/7, Figs. 9–11), and a 71-person preference study favoring boundary continuity, panorama awareness, and overall quality.","tokens_in":23735,"tokens_out":945,"duration_ms":18564,"significance":"If the geometry-transfer story holds, the work is a useful systems contribution: a large multi-task panoramic in-context dataset, a practical recipe for injecting spherical continuity and depth cues into a modern flow transformer, and a unified token-concatenation model that covers four editing-style tasks without task-specific heads. Strengths that should be credited include the scale and multi-task design of Canvas360Dataset, the explicit LRCE-RGB boundary metric, the user study, and the relatively thorough ablations of depth, positional offset, L_sim, and velocity circular padding. Even if residual gains partly come from data scale rather than pure geometry, the engineering package and public dataset would still be of interest to panoramic generation and immersive-content communities.","major_comments":[{"comment":"Secs. 3.2–3.3 and App. B: the central claim is that parallel RGB–depth pretraining installs spherical geometric priors that survive depth removal and drive FAED / seam / distortion-aware editing gains. After fine-tuning, supervision is appearance-only token concatenation. Fig. 7 (cubemap-side depth re-estimation) and Fig. 9 (backbone transfer) are suggestive but indirect. A load-bearing control is missing: same RGB pretraining data and schedule without the depth branch (and without L_sim / depth RoPE offset), then identical Stage-2 fine-tuning. Without that isolation, it remains possible that gains largely reflect better panoramic RGB statistics and the 900K pairs rather than retained spherical geometry. Please add this control (or an equivalent geometry-vs-appearance probe) and temper causal language in the abstract and conclusion accordingly.","section":null},{"comment":"Sec. 3.4 and App. B (pseudo-depth + synthetic pairs): DAP depths are clipped (100 m outdoor / 10 m indoor) and known to be noisy under ERP; style/edit pairs are largely produced by non-panorama-native models (FLUX.2-dev, NanoBanana, etc.) then filtered (50K manual check; Tab. 5 only compares residual noise on 20K edits, not geometric fidelity of teachers). Because Stage 2 discards depth, any teacher-induced ERP-inconsistent structure can be baked into the unified model. Please quantify geometric error of the synthetic edit/style pairs (e.g., line-drawing / seam / polar-distortion checks beyond the brief style filter) and discuss how much of the qualitative “correct panoramic distortion” on edits could be inherited from filtered teachers versus from Stage-1 priors.","section":null},{"comment":"Table 1 vs. Tables 3/7: Canvas360 leads FAED (2.33) and several perceptual metrics but trails DiT360 on FID / FID_pole / FID_equ and shows non-monotonic ablation behavior (depth alone can darken outputs; VCP helps BRISQUE/NIQE more than FID-family). The paper asserts “particularly strong” panorama-specific fidelity. Please reconcile when FAED should be preferred over FID-family for ERP evaluation, report confidence intervals or multi-seed variance for the main Table 1 numbers, and state more carefully which claims are supported by FAED/user study versus which remain mixed.","section":null}],"minor_comments":[],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"Punchline: this is a competent two-stage systems paper that makes panoramic in-context generation more usable. The new pieces that matter are velocity circular padding (ghost columns on the velocity field with continuous longitude indices, not just latent copy-paste), parallel RGB–depth generation with positional offsets and L_sim under spherical geometry, and Canvas360Dataset—1M paired samples spanning style, inpaint, outpaint, and edit. They then drop depth and train one token-concatenation model on all four tasks.\n\nWhat it does well is concrete. FAED is the cleanest win (2.33 vs DiT360/HunyuanWorld at 2.91). LRCE-RGB boundary numbers are best in the table. Qualitative edits put correct ERP distortion on added objects where FLUX/NanoBanana/SE360/Omni2 do not. Ablations (Tabs. 3/7, Figs. 9–11) show depth helps FAED, offsets + L_sim stop the dark-collapse mode, and the depth-aware backbone beats a same-recipe FLUX fine-tune on completion. User study (71 people) prefers them on seam continuity, panorama awareness, and overall quality. Citations are fair to DiT360, SE360, PanoDiffusion, etc.; free parameters are ordinary LoRA/FM knobs, not hidden fitting.\n\nSoft spots, in proportion: the stress-test is real but overstated. DAP depths are clipped and noisy; most of the 900K pairs come from non-panorama-native models with light filtering. After depth is removed, nothing isolates “spherical prior” from “better RGB texture stats on a big multi-task set.” Mixed FID vs DiT360, no error bars, internal preference study, and residual synthetic noise are real but secondary. The cleaned-vs-random 20K edit check only speaks to noise, not geometry fidelity. Still, the paper reports the dark-output failure and the mixed metrics; it does not hide them. The central claim—better geometric consistency under ERP—holds on the evidence they show.\n\nWho it is for: anyone building 360 generators, VR/AR authoring, or multi-task ERP editing. Not a theory paper. Deserves a serious referee; I would engage, cite the dataset and the padding trick, and ask for public data/code plus a cleaner isolation of the prior after depth drop.","headline":"Solid systems paper: velocity circular padding + RGB–depth pretraining + a real 1M multi-task 360 dataset give a unified model that actually respects ERP geometry; the transfer story after dropping depth is the softest link, not a collapse.","tokens_in":24366,"tokens_out":593,"would_cite":true,"duration_ms":6872,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"Geometry-aware RGB–depth pretraining plus a 1M-task dataset yields one model that edits and completes 360° panoramas with better spherical fidelity.","keywords":["panoramic image generation","in-context generation","geometry-aware pretraining","flow matching","equirectangular projection","depth supervision","Canvas360Dataset"],"falsifier":"Retrain the identical architecture without any depth branch or velocity circular padding on the same 100 K panoramas, then re-evaluate FAED, left–right boundary error, and human preference for seam continuity and object distortion on the same held-out set; if the geometry-aware model no longer wins on those three measures, the central claim fails.","tokens_in":24206,"feed_emoji":"🌐","tokens_out":663,"duration_ms":6156,"temperature":0.7,"pith_summary":"Canvas360 argues that ordinary image generators fail on equirectangular panoramas because they never learn spherical geometry or left–right seam continuity. The fix is a two-stage recipe: first pretrain a flow transformer to generate RGB panoramas and radial depth maps in parallel, using velocity circular padding and a similarity penalty so the two modalities stay distinct yet mutually informative; then discard depth and fine-tune one unified model on a new 1-million-pair dataset covering style transfer, inpainting, outpainting, and object editing. Token-level concatenation of context and target latents lets the same weights handle every task. The paper shows this transfers geometry priors into cleaner seams, correct latitude distortion on inserted objects, and the best reported panorama-specific FAED scores, while remaining competitive on standard fidelity metrics.","feed_headline":"One model learns 360° geometry then edits any panorama","feed_subtitle":"RGB–depth pretraining plus a 1M-pair dataset yield cleaner seams and correct spherical distortion","key_machinery":"Parallel RGB–depth generation inside a Flow Transformer: post-VAE depth tokens are sequence-concatenated with RGB tokens under a constant 3D-RoPE offset, trained with independent flow-matching losses, velocity circular padding that synchronizes ghost longitude columns, and a squared-correlation similarity penalty that keeps the two velocity fields from collapsing.","core_discovery":"Geometry-aware pretraining—parallel RGB–depth generation regularized by positional offsets, similarity loss, and velocity circular padding—supplies spherical spatial priors that, once transferred by unified in-context fine-tuning on Canvas360Dataset, produce a single token-concatenation model whose panoramic fidelity (especially FAED) and geometric consistency exceed prior specialized and general baselines across text-to-panorama, style transfer, inpainting, outpainting, and editing.","pith_inferences":[],"forward_implications":[],"fun_headline_variants":["Canvas360 learns spherical geometry then edits any panorama","RGB–depth pretraining plus 1M pairs unlock unified 360° edits","Geometry-aware priors let one model handle all panoramic tasks","Parallel depth and circular padding yield consistent 360° fidelity","Token-level fine-tuning transfers spherical priors across edits"],"cache_read_input_tokens":16512,"weakest_assumption_plain":"That pseudo-depth maps estimated by an off-the-shelf panoramic depth model, plus synthetic edit pairs produced by non-panorama-native generators and then filtered, still carry accurate enough spherical geometry for the learned priors to survive after depth is removed at fine-tuning time.","fun_headline_variants_meta":{"raw":{"variants":["Canvas360 learns spherical geometry then edits any panorama","RGB–depth pretraining plus 1M pairs unlock unified 360° edits","Geometry-aware priors let one model handle all panoramic tasks","Parallel depth and circular padding yield consistent 360° fidelity","Token-level fine-tuning transfers spherical priors across edits"]},"model":"grok-4.5","effort":"low","cost_usd":0.004172,"raw_usage":{"total_tokens":1286,"prompt_tokens":790,"num_sources_used":0,"completion_tokens":67,"cost_in_usd_ticks":41720000,"prompt_tokens_details":{"text_tokens":790,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":429,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":790,"tokens_out":67,"duration_ms":4093,"temperature":1.0,"reasoning_tokens":429,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-14T15:28:19.735708+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Retrain the identical architecture without any depth branch or velocity circular padding on the same 100 K panoramas, then re-evaluate FAED, left–right boundary error, and human preference for seam continuity and object distortion on the same held-out set; if the geometry-aware model no longer wins on those three measures, the central claim fails.","supporting_citations":[],"review_version":2}