{"id":"fa955a4a-b48f-49c3-adff-1200c815ad96","arxiv_id":"2412.14963","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A single-image feed-forward model trains on a 100K-generated-subject multi-view dataset and reconstructs animatable 3D Gaussian human avatars in under one second.","lead":"IDOL builds a photorealistic, animatable 3D human avatar from a single photograph in under a second, using a feed-forward transformer trained on a new 100K-subject synthetic multi-view dataset. The significance is the speed and the scale of generated training data for human avatars, though real-world validation is limited.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported metrics are not end-to-end on real images: all quantitative tests supply ground-truth SMPL-X/camera, so the single-image generalization claim remains unmeasured.","rationale":"The reader's conditional verdict is well placed: the paper's central claim is that a single RGB image yields an instant, animatable, photorealistic 3D human, and the strongest quantitative evidence is on generated data with ground-truth SMPL-X and cameras supplied. My stress test agrees with the reader's concern about synthetic-data transfer, but sharpens it: even the independent 2K2K evaluation and the user study do not exercise the full end-to-end system from a real photograph with estimated body parameters. This is a correctness-risk issue, not an internal inconsistency, so it does not change the verdict from CONDITIONAL; it strengthens the conditions. I credit the paper for including the 2K2K experiment, the WE comparison, the user study, and the runtime figures, which are real evidence in the right direction. The missing piece is a quantitative real-capture benchmark with the actual deployment pipeline, including a standard SMPL-X estimator and no ground-truth camera or body parameters. The suggested test would settle whether the reported quality transfers to real photos or whether the gain is partly an artifact of the evaluation protocol.","tokens_in":20897,"tokens_out":6374,"duration_ms":59658,"concrete_test":"Evaluate IDOL on a real multi-view capture benchmark (e.g., MVHumanNet or HuMMan held-out subjects) using the exact deployment pipeline: single reference RGB image -> standard SMPL-X estimator (e.g., SMPLer-X/NLF) -> IDOL -> multi-view rendering, then compute PSNR/LPIPS/MSE against the captured views. Compare these numbers with Tab. 2. If PSNR drops by more than ~1.5 dB or if the rank order against SIFU/GTA reverses, the single-image generalization claim is not supported. Also report the SMPL-X error on the same inputs.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that IDOL turns a single RGB image into an animatable photorealistic avatar. The quantitative support (Tab. 2, Tab. 4, and the supplement's 2K2K table) is produced under a protocol that explicitly supplies ground-truth SMPL-X parameters and camera intrinsics to the method. That is not the deployment condition of the claim, and it matters because IDOL's Gaussian maps are defined relative to SMPL-X vertices (Sec. 4.1) and animation via LBS needs a SMPL-X pose; the model cannot be exercised 'from a single image' without an accurate front-end SMPL-X estimate. The paper's own supplementary Fig. 16b shows that imperfect SMPL-X produces leaning/bent avatars. Moreover, the main held-out test set (HuGe100K) is generated by MVChamp, the same generator that produced the training data, so even the provided SMPL-X/camera are for synthetic views. The 2K2K numbers and the small qualitative in-the-wild and user-study examples go in the right direction, but they do not measure the full RGB-to-avatar pipeline on real captures with estimated body parameters. Thus the load-bearing assumption is not just MVChamp realism; it is the unstated coupling of IDOL to external SMPL-X estimation whose accuracy is never quantitatively evaluated.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces IDOL, a feed-forward transformer that reconstructs an animatable 3D Gaussian avatar from a single image, trained on a new large-scale synthetic multi-view dataset, HuGe100K, generated by a pose-conditioned video diffusion model (MVChamp). The method represents the avatar as Gaussian attribute maps in a SMPL-X UV space, which supports animation and editing. The authors report quantitative improvements over existing baselines on held-out synthetic data and a small real-capture benchmark, along with qualitative results on in-the-wild images.","tokens_in":21179,"tokens_out":5932,"duration_ms":51450,"significance":"If the result holds, the paper offers two contributions: a fast (sub-second) single-image human avatar reconstruction model and a large-scale generated dataset that enables training of such models. The use of generated data at scale is a timely direction with potential to address the scarcity of photorealistic 3D human data. The UV-space Gaussian representation is also practical for downstream editing and animation. However, as detailed below, the evaluation protocol currently limits the strength of the central claim, and the paper's comparative evidence is weakened by metric and baseline issues. The work is nevertheless a plausible and useful step, provided the evaluation gaps are addressed.","major_comments":[{"comment":"The quantitative evaluation provides ground-truth SMPL-X parameters and camera parameters to all methods (quote: 'The ground-truth camera parameters and SMPL-X parameters are provided for all methods'), so Tables 2 and 4 measure conditional reconstruction given an exact body model rather than the advertised single-image pipeline. Because the model's representation and animation are defined relative to SMPL-X vertices (Sec. 4.1), and because supplementary Fig. 16b shows that imperfect SMPL-X input causes leaning/bent avatars, the actual RGB-to-avatar claim is not quantitatively validated. Please add an end-to-end evaluation using an off-the-shelf SMPL-X estimator on real captures, or explicitly reframe the claim and protocol to state that the method assumes an accurate upstream body estimate.","section":"Sec. 5.2 (Dataset and Metrics)"},{"comment":"The main evaluation is on held-out HuGe100K images produced by MVChamp, the same generator used to create the training set, so the reported metrics partly measure self-consistency within the synthetic distribution. The 2K2K results in the supplementary (Tab. 4) provide useful external grounding, but the absolute performance drops substantially (PSNR 16.7 vs 21.7), and that real-data evaluation is small. To support the generalization claim, please report a more extensive real-capture benchmark (e.g., a larger subset of 2K2K or a comparable multi-view human dataset) with estimated SMPL-X parameters.","section":"Sec. 5.2 and Table 2"},{"comment":"The reported LPIPS values for SIFU and GTA (≈1.6) are far above the typical range for image reconstruction metrics, and the authors attribute this to an orthographic-vs-perspective projection mismatch. This makes the quantitative comparison against these baselines difficult to interpret. Please report results under a matched projection/camera setting (e.g., re-render all methods with the same perspective camera or provide orthographic equivalents) and specify the exact LPIPS computation. In addition, the main quantitative table omits strong feed-forward human-specific baselines such as HumanLRM, which appear only in the supplementary; including such a baseline in the unified protocol would make the comparison more convincing.","section":"Table 2 and Sec. 5.2"}],"minor_comments":[{"comment":"The word 'Spaiens' in the architecture figure appears to be a typo for 'Sapiens'.","section":"Fig. 4"},{"comment":"Reference [25] contains a malformed URL (a space in 'black-forest labs/flux'); please correct it.","section":"References, [25]"},{"comment":"The use of FaceFusion to enhance facial details may alter identity; please discuss whether and how identity consistency is verified after face swapping.","section":"Sec. 3.2"},{"comment":"The limitation paragraph lists fixed viewpoints and facial optimization but does not mention the method's dependence on an accurate SMPL-X estimate; given the evaluation protocol, this should be explicitly acknowledged.","section":"Sec. 6 (Limitations)"},{"comment":"The 'under 1 second' claim should specify whether the time includes SMPL-X estimation or only the IDOL forward pass.","section":"Sec. 1 and Sec. 5.2"},{"comment":"Setting opacity to 1 for all Gaussians may limit modeling of transparency in hair and loose clothing; a brief comment on this choice would be helpful.","section":"Sec. 4.1"}],"recommendation":"major_revision","confidential_remarks":"The paper presents a plausible method and a large dataset, but the evaluation protocol needs strengthening before the central claim can be considered substantiated. In addition to the major comments, I would urge the editors to verify that the dataset and code are made available, as the dataset construction is a large part of the contribution. The generated-data approach also raises questions about data licensing and bias that the authors may wish to address in a broader discussion."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a serious systems paper with a genuinely large new generated dataset, and the feed-forward design is clean. But every quantitative test supplies ground-truth SMPL-X and camera, and the held-out test set comes from the same generator as the training set. The headline 'from a single image' claim is not actually measured end to end.\n\nWhat is genuinely new is HuGe100K: 100K identities x 24 views, built from Flux images and a fine-tuned Champ (MVChamp) with temporal shift denoising and face/hand refinements. That pipeline is well thought out, and the paper shows honest ablations of its components. The model itself is a fairly direct combination of known pieces - frozen Sapiens encoder, UV-alignment transformer, Gaussian maps on SMPL-X - but the scale of training data makes the system non-trivial, and the authors do the right thing in ablating Sapiens vs DINOv2 and with/without the generated data. They also state limitations plainly: fixed viewpoints, secondary facial detail, half-body failure, and in the supplement they own that imperfect SMPL-X leads to leaning/bent avatars.\n\nThe soft spot is proportional and real. On the main table (Tab. 2), the test images are held-out HuGe100K frames from MVChamp - the very generator that produced the training data. On 2K2K, which is real, the metrics are better than SIFU/GTA, but again the method is handed the ground-truth SMPL-X and camera. The in-the-wild comparisons are qualitative. So the paper demonstrates a strong conditional reconstruction: given an accurate body fit and known camera, IDOL produces good animatable Gaussians on synthetic distribution, and promising results on real scans. It does not demonstrate the full RGB-to-avatar pipeline under deployment conditions. The authors' own Fig. 16b says the model is sensitive to the front-end fitting. That's a missing experiment, not a fatal flaw.\n\nThe baseline comparisons are also confounded by the orthographic/perspective mismatch they acknowledge, which makes the margin over SIFU/GTA in Tab. 2 less decisive than it looks. And no code or data are released, so independent verification is stalled.\n\nWho should read it: anyone working on human avatar reconstruction or synthetic training data. The dataset contribution is the real asset, and if it ships, it could seed further work. For peer review, I'd send it, but with a concrete ask: either release the data/models, or add a small real-image end-to-end evaluation with an off-the-shelf SMPL-X estimator and report the drop. Without that, the central claim stays under-supported. Conditional accept, not clean accept.","headline":"A serious, well-engineered avatar system with a genuinely large synthetic dataset, but the end-to-end 'single image' claim is unmeasured because all quantitative tests provide ground-truth SMPL-X/camera and the main test set comes from the same generator as training.","tokens_in":21713,"tokens_out":3959,"would_cite":true,"duration_ms":32107,"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":"A single photo yields an animatable, photorealistic 3D human in under a second","keywords":["single-image 3D human reconstruction","animatable avatar","3D Gaussian Splatting","SMPL-X UV space","synthetic multi-view dataset","feed-forward transformer","human avatar editing","pose-controllable image animation"],"falsifier":"Take a set of real people photographed simultaneously by 24 calibrated cameras, feed one view into IDOL, and compare rendered novel views against the true captured views at the same resolution; if the error on these real captures is substantially worse than the error on held-out HuGe100K samples (for example, LPIPS more than twice as high) or worse than a strong optimization baseline, the generated-data assumption is falsified. The 2K2K experiment in the paper is a small step in this direction but uses only a few subjects.","tokens_in":20715,"feed_emoji":"🧍","tokens_out":7534,"duration_ms":61686,"temperature":0.7,"pith_summary":"This paper is trying to establish that single-image 3D human reconstruction can be made instant, animatable, and generalizable by replacing scarce real multi-view captures with a large generated dataset. To that end it introduces HuGe100K, a dataset of 100K diverse subjects with 24-view posed image sets, and IDOL, a feed-forward transformer that maps one 1024x1024 photo to a 3D Gaussian avatar defined in SMPL-X UV space. The stated result is reconstruction at 1K resolution in under a second on one GPU, with avatars that animate without post-processing. If true, this would remove the per-subject optimization and manual rigging that currently block practical avatar creation from casual photos.","feed_headline":"One photo becomes a poseable 3D human in under a second","feed_subtitle":"A 100K-subject generated multi-view dataset lets a single forward pass create an editable, animatable avatar.","key_machinery":"The load-bearing object is the canonical UV Gaussian avatar: every reconstructed person is expressed as Gaussian attribute maps registered to SMPL-X vertices, so pose, shape, clothing geometry, and texture are predicted as structured 2D maps rather than as a free 3D field. A pretrained high-resolution human-image encoder, kept frozen, supplies tokens to a UV-alignment transformer that fuses them with learnable UV tokens; a decoder lifts the fused tokens to Gaussian parameter maps, and differentiable rendering trains the whole pipeline against multi-view images. The dataset side is carried by MVChamp, a re-trained pose-controllable image animation model that turns a reference photo plus SMPL-X pose sequences into 24 consistent views, with temporal shift denoising to close the first-to-last-frame gap and face swapping to sharpen facial detail.","core_discovery":"The central claim, stated on the paper's own terms, is that a single feed-forward pass over one input image can predict a full animatable human avatar as a set of Gaussian primitives parameterized in a canonical SMPL-X UV space, and that this becomes possible when the model is trained on HuGe100K's 2.4M generated multi-view frames. The model predicts per-vertex offsets, colors, and opacities as UV attribute maps; linear blend skinning then animates the result, and editing is done by modifying UV textures or SMPL-X shape parameters. On held-out generated data the method reports lower error than optimization-based and generic large reconstruction baselines on MSE, PSNR, and LPIPS, and qualitative in-the-wild tests show it generalizes to cartoon inputs and large-angle views. The paper also reports a user study in which participants preferred its outputs on face, clothing, back-view consistency, and overall quality.","pith_inferences":["Beyond the paper's own evaluation, the quantitative gains are measured mainly on held-out generated data, so real-world performance may be weaker than reported until a larger real-capture benchmark is run; the paper's own 2K2K evaluation suggests the gap is not zero.","Because every avatar is anchored to SMPL-X vertices, very loose clothing, asymmetric garments, and extreme topologies are likely to stress the representation, and a focused benchmark on garments with large offsets from the body would test this ceiling.","The data-generation recipe (LLM-designed prompts, pose-conditioned multi-view video model, face refinement) is portable to other subject categories; testing it on animals or objects with articulated skeletons would show whether the scaling story generalizes beyond humans.","Chaining MVChamp-style longer sequences with IDOL reconstruction could extend the method from single images to video reenactment with persistent identity, but temporal consistency over long sequences remains an open question and the paper only demonstrates short reenactments."],"forward_implications":["Reconstructing an animatable, textured human avatar from a casual photo becomes a one-second operation rather than a minutes-long optimization.","Because the output lives in SMPL-X UV space, animation and shape or texture editing come directly from the representation, with no manual rigging or post-processing.","Scaling up generated multi-view data appears to be the main driver of reconstruction quality, suggesting further synthetic-data scaling will keep improving generalization.","Avatar creation is no longer limited by the high cost and small size of scanned human datasets, sidestepping scanning infrastructure and many privacy or copyright constraints."],"supporting_citations":[{"why":"Defines 3D Gaussian Splatting, the rendering primitive the avatar representation builds on and the differentiable renderer used for supervision.","marker":"[36]"},{"why":"Supplies the frozen high-resolution human-image encoder that extracts fine-grained patch tokens from the input photo.","marker":"[37]"},{"why":"Establishes the UV-space Gaussian attribute map representation that IDOL adapts for animatable human avatars.","marker":"[81]"},{"why":"Base human image animation model that the paper retrains into MVChamp, the multi-view generator behind HuGe100K.","marker":"[89]"},{"why":"Text-to-image model used to synthesize the 90K diverse human reference images that anchor the dataset.","marker":"[25]"},{"why":"THuman scans provide the 3D consistency fine-tuning data for MVChamp and part of the training and test supervision for IDOL.","marker":"[86]"},{"why":"Source of 10K real-life full-body images that diversify the generated dataset with real captured appearance.","marker":"[47]"},{"why":"Side-view-conditioned implicit function baseline that IDOL compares against and outperforms in quantitative and qualitative tests.","marker":"[85]"}],"fun_headline_variants":["Single photo to animatable 3D human in one forward pass","One image to a poseable 3D avatar instantly","Instant photoreal 3D human from a single snapshot","One photo yields an editable, poseable 3D human","From photo to full 3D human avatar in under a second"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole pipeline assumes that the views generated by MVChamp are consistent, photorealistic, and diverse enough to stand in for real multi-view captures; if generated images have systematic artifacts or a domain gap from real photos, the model's quality on real inputs will not match the numbers reported on held-out generated data.","fun_headline_variants_meta":{"raw":{"variants":["Single photo to animatable 3D human in one forward pass","One image to a poseable 3D avatar instantly","Instant photoreal 3D human from a single snapshot","One photo yields an editable, poseable 3D human","From photo to full 3D human avatar in under a second"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00084,"raw_usage":{"total_tokens":3684,"prompt_tokens":991,"completion_tokens":2693,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":607,"completion_tokens_details":{"reasoning_tokens":2607}},"tokens_in":607,"tokens_out":2693,"duration_ms":19739,"temperature":1.0,"reasoning_tokens":2607,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T11:44:33.422230+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a set of real people photographed simultaneously by 24 calibrated cameras, feed one view into IDOL, and compare rendered novel views against the true captured views at the same resolution; if the error on these real captures is substantially worse than the error on held-out HuGe100K samples (for example, LPIPS more than twice as high) or worse than a strong optimization baseline, the generated-data assumption is falsified. The 2K2K experiment in the paper is a small step in this direction but uses only a few subjects.","supporting_citations":[{"cited_title":"Sapiens: Foundation for human vision mod- els","cited_arxiv_id":null,"evidence_quote":"Supplies the frozen high-resolution human-image encoder that extracts fine-grained patch tokens from the input photo."},{"cited_title":"Champ: Controllable and consistent human image animation with 3d parametric guidance","cited_arxiv_id":null,"evidence_quote":"Base human image animation model that the paper retrains into MVChamp, the multi-view generator behind HuGe100K."},{"cited_title":"Deephuman: 3d human reconstruction from a single image","cited_arxiv_id":null,"evidence_quote":"THuman scans provide the 3D consistency fine-tuning data for MVChamp and part of the training and test supervision for IDOL."},{"cited_title":"Deepfashion: Powering robust clothes recognition and retrieval with rich annotations","cited_arxiv_id":null,"evidence_quote":"Source of 10K real-life full-body images that diversify the generated dataset with real captured appearance."},{"cited_title":"Sifu: Side-view conditioned implicit function for real-world us- able clothed human reconstruction","cited_arxiv_id":null,"evidence_quote":"Side-view-conditioned implicit function baseline that IDOL compares against and outperforms in quantitative and qualitative tests."}],"review_version":1}