{"id":"b33d8b41-d6cb-4c39-b5f8-a6e2e6c418b4","arxiv_id":"2504.15009","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Insert Anything is a single model fine-tuned on 159,908 prompt-image pairs that performs mask- or text-guided insertion of people, objects, and garments from reference images into target scenes.","lead":"Insert Anything trains one model that can copy a person, object, or garment from a reference photo into a new scene, guided by either a mask or a text description. It could replace several specialized editing tools with a single flexible pipeline for photo compositing and virtual try-on.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Text-prompt evaluation may be circular: source images in the text-prompt test set are FLUX.1 Fill outputs, and Insert Anything is a LoRA fine-tune of FLUX.1 Fill.","rationale":"The reader's weakest assumption identifies exactly this circularity: text-prompt test sources are FLUX.1 Fill outputs while the model is a FLUX.1 Fill LoRA. I agree that this is the most load-bearing concern for the text-prompt portion of the central claim. I do not find a comparably serious flaw in the mask-prompt evidence: mask-prompt results are evaluated on real AnyInsertion pairs, DreamBooth, and VTON-HD, and the reported gains over AnyDoor/MimicBrush/ACE++/CatVTON are large, though statistical significance and code release are still missing. Because the reader's verdict is already CONDITIONAL and conditioned on exactly this concern, my stress-test does not move the verdict. The recommendation stays CONDITIONAL/UNCHANGED pending an independent text-prompt evaluation and release of data/code. I mark agreement as 'partial' because the reader also emphasizes dataset/code availability and internal count inconsistencies, which I see as secondary to the circularity concern but still relevant to verifiability.","tokens_in":15815,"tokens_out":2310,"duration_ms":25265,"concrete_test":"Construct an independent text-prompt insertion benchmark: take real target images containing objects or garments, and create source images by removing the object without FLUX.1 Fill (e.g., manual inpainting, a different inpainting model, or unedited photographs with known ground-truth insertions). Run the released Insert Anything model and AnyEdit on identical reference-source-text triples and compare PSNR/SSIM/LPIPS/FID with paired significance tests. If the margin over AnyEdit collapses or reverses on non-FLUX-generated source images, the text-prompt performance claim is an artifact of matching the base model's edit distribution rather than a general text-guided insertion capability.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim includes text-guided insertion (mask and text control). The only text-prompt benchmark is the AnyInsertion test subset, but its source images were synthesized in §3.2 by applying FLUX.1 Fill [dev] to target images (\"Source Image Generation\"). Insert Anything is itself a LoRA fine-tune of FLUX.1 Fill (§5.1, Implementation Details). Thus the test pairs have the form (reference, FLUX-generated source, original target), which is exactly the distribution the model saw during training. The reported text-prompt improvements over AnyEdit (Table 3, Fig. 5) may largely reflect the model learning to invert its own base model's edits rather than learning a general text-guided insertion capability on arbitrary user scenes. This does not invalidate the mask-prompt results, which use real target/reference pairs and independent benchmarks, but it leaves the text-prompt half of the headline claim unsupported. The paper's own numbers also show internal inconsistency (120K in abstract/conclusion vs. 159,908 in §3.3; person test set 60 in §3.3 vs. 30 in §5.1), which compounds the difficulty of trusting the reported text-prompt evaluation without code or dataset release.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Insert Anything, a unified framework for reference-based image insertion built on FLUX.1 Fill by fine-tuning with LoRA. The method supports two control modes, mask-prompt and text-prompt, through two in-context prompting layouts: a mask-prompt diptych and a text-prompt triptych. The authors introduce a new dataset, AnyInsertion, containing mask- and text-prompt pairs for person, object, and garment insertion, and evaluate the trained model on AnyInsertion, DreamBooth, and VTON-HD against task-specific and unified baselines. The main claims are that a single model handles multiple insertion tasks under two control modes and consistently outperforms existing alternatives. The mask-prompt experiments use real paired images and independent benchmarks, while the text-prompt experiment relies on synthetic source images generated by FLUX.1 Fill, which is also the base model that the method fine-tunes. The manuscript reports several inconsistent dataset and test-set counts, and the central ablation for the dataset contribution uses a training-free base model rather than a controlled training comparison.","tokens_in":16064,"tokens_out":3718,"duration_ms":35581,"significance":"If the central claims hold, the contribution is practically valuable: a single DiT-based model covering person, object, and garment insertion with both mask and text control would simplify deployment, and the AnyInsertion dataset, with roughly 160K pairs, would be a useful resource for the community. The idea of using polyptych prompting to turn reference-based insertion into an in-context editing problem is interesting and is in line with recent DiT in-context work. The mask-prompt results on real paired data, including strong improvements on VTON-HD over specialized try-on methods, provide partial independent support. However, the text-prompt evaluation is weakened by the synthetic-source circularity, the reported numbers lack error bars on small test sets, and the dataset-ablation claim is not supported by a controlled experiment. The significance of the paper is therefore currently contingent on fixing these evaluation issues and on the promised dataset/code release.","major_comments":[{"comment":"The text-prompt evaluation is potentially circular. The source images for text-prompt pairs are generated with FLUX.1 Fill [dev] (Section 3.2, 'Source Image Generation'), and Insert Anything is itself a LoRA fine-tune of FLUX.1 Fill (Section 5.1). The AnyInsertion test set is a subset of the same dataset built by the same pipeline, so the model sees test pairs drawn from the exact distribution of its own base model's outputs. Consequently, the reported text-prompt gains over AnyEdit may largely reflect the ability to invert FLUX.1 Fill's own edits rather than a general text-guided insertion capability on arbitrary user scenes. To support the headline claim, the authors should evaluate text-prompt insertion on independently created source-target pairs (e.g., real images of the same scene with and without an object, or sources generated by a different model) and report results separately for FLUX-generated and non-FLUX sources.","section":"§3.2, §5.1, Table 3"},{"comment":"The test-set sizes are internally inconsistent, which makes the quantitative claims difficult to verify. Section 3.3 states that the mask-prompt test set has 120 pairs, including 60 person pairs (30 simple and 30 complex); Section 5.1 says the person test set has 30 simple-scene samples; and Section 5.4's weighting uses 40 object, 30 garment, and 30 person pairs, totaling 100. Table 5 reports person-insertion numbers without stating whether they are computed on 30 or 60 pairs, and Table 6 is ambiguous about which subset it is aggregated over. The authors must give exact per-table test-set sizes and recompute any weighted averages accordingly.","section":"§3.3 vs. §5.1, §5.4, Tables 5 and 6"},{"comment":"The 'w/o AnyInsertion' row is not a controlled ablation of the dataset. The text says the model relies on a 'training-free model for inference,' which is the base FLUX.1 Fill inpainting model, so this comparison conflates the effect of the AnyInsertion data with the effect of LoRA fine-tuning, the training loss, and the prompting format. The observed degradation therefore cannot be attributed to the training data. A valid ablation would train the same framework on an alternative dataset or ablate the data while keeping the training procedure identical; otherwise, the claim that the AnyInsertion dataset drives the improvement is unsupported.","section":"§5.4, Table 6"},{"comment":"The benchmark sizes are very small and no error bars or significance tests are reported: the text-prompt object test set has 16 pairs, the garment test sets have 30 pairs, the person test set has 30 or 60 pairs, and FID is computed on these tiny samples. The claim that the method 'consistently outperforms' all baselines on every metric is not statistically established. The authors should report per-metric confidence intervals, per-image standard deviations, or significance tests, and ideally release the test set and evaluation code so the numbers are independently reproducible.","section":"Tables 2–5"}],"minor_comments":[{"comment":"The dataset size is reported inconsistently: the abstract, introduction bullet, and conclusion say 120K prompt-image pairs, while Section 3.3 and Table 1 report 159,908 (58,188 mask + 101,720 text). The authors should use one consistent number throughout.","section":"Abstract, §1, §6, §3.3, Table 1"},{"comment":"The paper says the text-prompt test subset contains 38 pairs (16 object, 22 garment), but Table 3 reports only object insertion and no garment text-prompt results. It should be stated whether garment text-prompt pairs were evaluated and, if so, where the numbers are.","section":"§3.3, Table 3"},{"comment":"The claim that AnyInsertion is 'the first open-source dataset for image insertion' is not substantiated in the manuscript, which provides no download link, license, or release plan. Either provide the release details or soften the claim.","section":"§1, §6"},{"comment":"The ablation study is limited to mask-prompt insertion; there is no ablation of the text-prompt triptych format, the text template, or the contribution of the text branch. Given that text-prompt control is half of the central claim, an ablation for that mode should be added.","section":"§5.4, supplementary Table 7"},{"comment":"The adaptive crop strategy introduces hyperparameters β and T, but the paper does not report sensitivity to these values or show that the chosen values are robust across object, person, and garment insertion. A small sensitivity analysis would strengthen the appendix.","section":"Appendix A, Eq. (7)"},{"comment":"The resolution entry 'Mainly 1–2K' is vague; the paper should report the actual resolution distribution of the dataset, including how many samples fall below 1K and how resizing to 768×768 affects the higher-resolution claims.","section":"Table 1"},{"comment":"The related work entry for OmniControl appears as 'Ominicontrol' in the text; the reference list should be checked for consistency and typos.","section":"References, related work"}],"recommendation":"major_revision","confidential_remarks":"The paper has a useful core idea and the mask-prompt results on real paired data are promising, but the text-prompt evaluation design, the inconsistent test-set counts, and the uncontrolled dataset ablation are load-bearing issues. I believe these are fixable within the scope of a revision: the authors can re-evaluate text-prompt insertion on non-circular sources, correct the reported counts, and replace the 'w/o AnyInsertion' row with a training-controlled comparison. If the authors are unable to obtain non-circular text-prompt evaluation data, they should explicitly limit their text-prompt claim to the FLUX-distribution setting. The manuscript would also benefit from releasing the dataset and evaluation code, since the current paper promises open-source data but gives no artifact."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You asked about arXiv:2504.15009. Read it carefully. The real contribution is the AnyInsertion dataset and the mask-prompt insertion results; the text-prompt evaluation has a load-bearing circularity problem that the authors do not address.\n\nWhat is actually new: a 159,908-pair dataset covering person, object, and garment insertion with both mask and text control at 1–2K resolution, assembled from e-commerce, video, and multi-view sources. That is genuinely useful and fills a gap—prior datasets are task-specific, lower-resolution, or single-mode. The unified DiT-based model with diptych/triptych in-context prompting is a sensible extension of recent in-context generation work, and the mask-prompt experiments on real paired images (AnyInsertion, DreamBooth, VTON-HD) show consistent gains over AnyDoor, MimicBrush, ACE++, CatVTON, and OOTDiffusion. The supplementary details on hybrid masking and adaptive crop are thoughtful. On the mask-prompt side, the paper delivers.\n\nThe soft spot is the text-prompt evaluation. Section 3.2 generates source images with FLUX.1 Fill, and the model is itself a LoRA fine-tune of FLUX.1 Fill. The text-prompt test pairs are therefore in-distribution: (reference, FLUX-generated source, original target) is exactly the training shape. The reported improvements over AnyEdit may largely reflect inverting the base model's own edits rather than a general text-guided insertion ability. This does not sink the mask-prompt results, but it leaves the 'text prompt' half of the headline claim unsupported as written.\n\nThere are also smaller problems: the abstract and conclusion say 120K pairs while §3.3 says 159,908; the person test set is 60 in §3.3 but 30 in §5.1 and §5.4; test sets are small with no error bars; and the 'w/o AnyInsertion' ablation compares against a training-free base model, which is not a controlled ablation of dataset value. Code and dataset are not released, so these inconsistencies cannot be checked independently.\n\nWho should read this: anyone building reference-based editing systems or designing insertion benchmarks. The dataset alone is worth citing once released. The paper deserves a serious referee, but only if the authors fix the text-prompt evaluation, clear up the number inconsistencies, and release the dataset/code. I would not cite it in my own work until the dataset is available.\n\nMy recommendation: send to peer review, but mark it as needing major revision. The mask-prompt contribution is real; the text-prompt claims need to be re-run on independently created sources.","headline":"The dataset and mask-prompt results are solid and worth referee time; the text-prompt half of the claim rests on a circular benchmark that should be redone before publication.","tokens_in":16627,"tokens_out":1891,"would_cite":false,"duration_ms":17520,"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 single diffusion-transformer model, trained once on 120K prompt-image pairs, can insert any object from a reference photo into a target scene under mask or text control, and outperforms task-specific and unified alternatives on person…","keywords":["image insertion","reference-based editing","in-context editing","diffusion transformer","polyptych prompting","AnyInsertion dataset","virtual try-on","multi-modal control"],"falsifier":"Assemble a text-prompt test set of 50 to 100 pairs in which the source image is an unmodified real photograph, with edit instructions written by humans rather than synthesized by FLUX.1 Fill, and compute Insert Anything's LPIPS and FID against AnyEdit on this set; if the reported advantage disappears or reverses, the claim that the model generalizes across text-prompt insertion scenes is refuted.","tokens_in":15616,"feed_emoji":"🖼️","tokens_out":10640,"duration_ms":80447,"temperature":0.7,"pith_summary":"The paper seeks to establish that image insertion—taking an object, person, or garment from one photograph and placing it into another—can be solved by one model rather than by a collection of task-specific systems. It introduces Insert Anything, a framework trained once on a new dataset of 120K prompt-image pairs (AnyInsertion) that covers person, object, and garment insertion under both mask and text control. On the AnyInsertion, DreamBooth, and VTON-HD benchmarks, the authors report that this single model outperforms all compared alternatives, including specialized try-on models and unified editing frameworks, across every reported metric. A sympathetic reader would take the paper's contribution to be a shared architecture and prompting scheme for the whole insertion problem, with a public dataset to support it.","feed_headline":"One model handles person, object, and garment insertion in any scene","feed_subtitle":"Trained once on 120K pairs, it beats task-specific and unified baselines under both mask and text control.","key_machinery":"The load-bearing mechanism is polyptych in-context editing within a DiT: the reference image, with background removed, is concatenated in the channel dimension with the masked source (diptych) or with the source plus an empty panel (triptych), and the binary mask is appended as an extra channel. These visual tokens enter the image branch of the transformer while CLIP image features from the reference and T5 text features enter the text branch; multimodal attention then lets the generated region draw on both. This design treats the reference as context rather than as a separate conditioning stream, which the paper argues preserves high-frequency identity details while harmonizing style, color, and texture with the target scene. The whole model is a LoRA fine-tune of the FLUX.1 Fill inpainting model, so the training and inference pipeline is comparatively light.","core_discovery":"The paper's central claim is that a diffusion transformer (DiT) can act as a universal insertion engine when it is fine-tuned on a mixed insertion dataset and fed a 'polyptych' of images. For mask-guided insertion, the input is a diptych that places the background-segmented reference image beside the partially masked source image; for text-guided insertion, it is a triptych of reference, source, and an empty target panel. Multimodal attention inside the DiT jointly attends to text, mask, and image tokens, so the same weights handle both control modes. The paper reports that this approach, without auxiliary modules such as ControlNet, beats AnyDoor, MimicBrush, and ACE++ on mask-guided object and person insertion, AnyEdit on text-guided insertion, and CatVTON and OOTDiffusion on garment insertion, including on the standard VTON-HD benchmark.","pith_inferences":["If the text-prompt results hold up on non-synthetic sources, the polyptych formulation could be transferred to other in-context editing problems, such as multi-object insertion or exemplar-based style transfer, by simply rearranging the panels.","The paper presents adaptive crop-and-zoom only as an inference-time trick; one could test whether training the same model with a scale-augmented objective would make the trick unnecessary.","Since the model is built on FLUX.1 Fill, its strong text-prompt performance might partly stem from reproducing that base model's own editing behavior; a clean ablation would retrain the LoRA on a different DiT inpainting base and compare.","The authors' future-work list (physics-aware constraints, mixture-of-experts, more conditioning types) suggests the next practical ceiling is physical plausibility rather than identity fidelity; a concrete extension would add a contact or depth consistency loss during training."],"forward_implications":["A single trained model can replace separate systems for person, object, and garment insertion, so downstream applications such as virtual try-on, scene composition, and content creation need not train or maintain several editing models.","The same model accepts either a mask or a text description as control, giving users the choice of specifying where to insert an element geometrically or describing the change in language.","Because the method requires no auxiliary networks like ControlNet, its weight footprint and inference graph are simpler than those of prior insertion systems while the reported quality is higher.","The AnyInsertion dataset, with roughly 160K training pairs at 1–2K resolution across both prompt types, provides a common training and evaluation ground for future unified insertion research.","On VTON-HD, a general insertion model outperforms dedicated virtual try-on models, suggesting that unification does not sacrifice task-specific quality."],"supporting_citations":[{"why":"Supplies the base inpainting model that Insert Anything fine-tunes and that generates the source images for text-prompt data.","marker":"[18]"},{"why":"Provides the DiT backbone whose multimodal attention lets text, mask, and image tokens be modeled jointly.","marker":"[31]"},{"why":"Establishes the diptych prompting idea that the paper extends into polyptych in-context editing.","marker":"[37]"},{"why":"Primary mask-guided baseline for object and person insertion that the method must outperform.","marker":"[3]"},{"why":"Reference-imitation baseline for mask-guided editing, compared on object and person tasks.","marker":"[2]"},{"why":"Strongest specialized garment insertion baseline, especially on the VTON-HD benchmark.","marker":"[6]"},{"why":"Unified editing framework used as a baseline across object, person, and garment tasks.","marker":"[25]"},{"why":"The text-prompt editing baseline against which Insert Anything evaluates text-guided insertion.","marker":"[49]"},{"why":"One of the two external benchmarks, used for object and person insertion evaluation.","marker":"[34]"},{"why":"The standard virtual try-on benchmark used for garment insertion evaluation.","marker":"[5]"}],"fun_headline_variants":["DiT-based model unifies all image insertion tasks","One DiT handles mask and text guided insertion","In-context editing lets a single DiT insert anything","Unified insertion framework beats task-specific baselines","Trained once on 120K pairs, works for any insertion"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the text-prompt training and evaluation data—whose source images were generated by FLUX.1 Fill, the same model Insert Anything is a LoRA fine-tune of—are representative of real user scenes; if they are not, the reported text-prompt gains may measure how well the model reproduces its base model's outputs rather than how well it generalizes.","fun_headline_variants_meta":{"raw":{"variants":["DiT-based model unifies all image insertion tasks","One DiT handles mask and text guided insertion","In-context editing lets a single DiT insert anything","Unified insertion framework beats task-specific baselines","Trained once on 120K pairs, works for any insertion"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000212,"raw_usage":{"total_tokens":1414,"prompt_tokens":934,"completion_tokens":480,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":550,"completion_tokens_details":{"reasoning_tokens":402}},"tokens_in":550,"tokens_out":480,"duration_ms":4991,"temperature":1.0,"reasoning_tokens":402,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T11:34:59.920810+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Assemble a text-prompt test set of 50 to 100 pairs in which the source image is an unmodified real photograph, with edit instructions written by humans rather than synthesized by FLUX.1 Fill, and compute Insert Anything's LPIPS and FID against AnyEdit on this set; if the reported advantage disappears or reverses, the claim that the model generalizes across text-prompt insertion scenes is refuted.","supporting_citations":[{"cited_title":"Anydoor: Zero-shot object-level im- age customization","cited_arxiv_id":null,"evidence_quote":"Primary mask-guided baseline for object and person insertion that the method must outperform."},{"cited_title":"Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation","cited_arxiv_id":null,"evidence_quote":"One of the two external benchmarks, used for object and person insertion evaluation."}],"review_version":1}