{"id":"1f564902-dce7-4815-8782-0c3513cf2cec","arxiv_id":"2606.24232","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":5.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":0,"one_line_summary":"FiCA creates real-time drivable 3D Gaussian avatars from one image using diffusion-based generative mapping from partial observations plus feed-forward mesh refinement.","lead":"FiCA generates photorealistic, drivable 3D Gaussian avatars from a single portrait image via a feed-forward pipeline combining vision foundation models, a diffusion model for mesh reconstruction, and a refinement network. A smart generalist might care because it removes per-person test-time optimization, potentially enabling instant avatar creation for VR, gaming, and video applications.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.3","headline":"Diffusion model's generative mapping from single portrait to identity-preserving 3D mesh remains the least-secured step; no explicit verification of generalization across pose/ethnicity diversity is described.","rationale":"Reader's weakest assumption matches the load-bearing technical risk exactly; the abstract-only review correctly flags the unverified generalization step. No other internal inconsistency is visible from the provided description.","tokens_in":1700,"tokens_out":296,"duration_ms":10642,"concrete_test":"Run the released model on a 50-subject test set containing 45°+ yaw angles and underrepresented ethnicities; compute ArcFace cosine similarity between rendered novel views and ground-truth multi-view images. If mean identity similarity falls below the best competing single-image baseline by >0.05, the mapping assumption does not hold at the claimed fidelity.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The strongest claim requires that the diffusion model (plus refinement network) produces complete, authentic meshes that preserve identity for diverse subjects without any test-time optimization. Single-image input is fundamentally under-constrained for full-head geometry and unseen appearance; success therefore hinges on the training distribution and the model's ability to hallucinate consistent back/side views. The abstract states this mapping is learned, yet provides no quantitative evidence (e.g., identity similarity scores on held-out extreme poses or cross-ethnicity splits) that the learned prior actually generalizes rather than memorizing frontal-biased statistics.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.3","summary":"The paper introduces FiCA, a feed-forward pipeline for generating photorealistic, drivable 3D Gaussian avatars from a single portrait image. It combines human-centric vision foundation models with a diffusion model that learns a generative mapping from partial observations to complete 3D mesh reconstructions, followed by a feed-forward mesh refinement network and a universal prior decoder to produce 3D Gaussians. The system claims to eliminate person-specific test-time optimization while producing avatars that faithfully represent diverse identities, support novel expressions in real time, and surpass the visual quality of recent competing methods.","tokens_in":1823,"tokens_out":512,"duration_ms":11965,"significance":"If the central claims hold with rigorous validation, the work would be significant for enabling instant, optimization-free avatar creation suitable for real-time applications in AR/VR and animation, addressing a key bottleneck in single-image 3D head reconstruction.","major_comments":[{"comment":"Experiments section: the central claim that the feed-forward approach 'surpass[es] the visual quality of avatars produced by recent competing methods' and 'faithfully represent[s] diverse identities' rests on unspecified experiments; no quantitative metrics (e.g., identity similarity scores, pose generalization error), dataset splits, ablation studies, or error bars are described to substantiate generalization of the diffusion model across pose, ethnicity, or extreme viewpoints.","section":"Experiments"},{"comment":"Method section (diffusion model description): the assertion that the diffusion model 'learns a generative mapping from these partial observations to complete and authentic 3D mesh reconstruction' is load-bearing for the no-optimization claim, yet the manuscript provides no verification (e.g., held-out extreme-pose or cross-ethnicity quantitative results) that the learned prior produces consistent back/side geometry rather than frontal-biased hallucination.","section":"Method"}],"minor_comments":[{"comment":"Abstract and introduction: the phrase 'universal prior model that decodes a generated mesh into a set of 3D Gaussians' would benefit from a brief citation or reference to the specific prior work being reused.","section":"Abstract"},{"comment":"Figure captions and results: visual comparisons would be clearer if they explicitly labeled the input portrait, generated mesh, and final Gaussian rendering for each competing method.","section":"Figures"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":{"model":"grok-4.3","summary":"We thank the referee for the thorough review and for highlighting the need for stronger quantitative support of our central claims. We agree that the current manuscript would benefit from expanded experimental details and additional verification results. We address each major comment below and will revise the manuscript accordingly.","responses":[{"response":"We acknowledge that the submitted manuscript presents the experimental claims primarily through qualitative comparisons and does not include the requested quantitative metrics, dataset splits, ablations, or error bars. This is a valid observation. In the revised version we will add: (1) identity similarity scores using a standard face recognition model such as ArcFace, (2) pose generalization error measured on held-out extreme yaw/pitch angles, (3) explicit train/test splits and cross-ethnicity evaluation, (4) ablation studies isolating the diffusion prior and mesh refinement, and (5) error bars across multiple random seeds. These additions will directly substantiate the generalization claims.","revision_made":"yes","referee_comment":"[Experiments] Experiments section: the central claim that the feed-forward approach 'surpass[es] the visual quality of avatars produced by recent competing methods' and 'faithfully represent[s] diverse identities' rests on unspecified experiments; no quantitative metrics (e.g., identity similarity scores, pose generalization error), dataset splits, ablation studies, or error bars are described to substantiate generalization of the diffusion model across pose, ethnicity, or extreme viewpoints."},{"response":"We agree that quantitative verification of non-frontal geometry consistency is essential to support the claim that the diffusion model produces authentic 3D reconstructions without frontal bias. The current manuscript relies on qualitative examples for this aspect. In the revision we will report quantitative metrics (e.g., surface reconstruction error on back/side regions) on held-out extreme-pose and cross-ethnicity test sets to demonstrate that the learned prior generalizes beyond frontal observations.","revision_made":"yes","referee_comment":"[Method] Method section (diffusion model description): the assertion that the diffusion model 'learns a generative mapping from these partial observations to complete and authentic 3D mesh reconstruction' is load-bearing for the no-optimization claim, yet the manuscript provides no verification (e.g., held-out extreme-pose or cross-ethnicity quantitative results) that the learned prior produces consistent back/side geometry rather than frontal-biased hallucination."}],"tokens_in":1380,"tokens_out":509,"duration_ms":15953,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"The main thing to know is that this paper describes a complete feed-forward system: vision foundation models feed a diffusion model that produces a 3D mesh from one photo, a refinement network cleans it up, and a fixed decoder turns the mesh into drivable 3D Gaussians. The selling point is that everything runs instantly with no per-person optimization at inference time.\n\nWhat is new is the specific end-to-end arrangement that removes the optimization step common in earlier Gaussian avatar work. The authors correctly identify that single-image input is under-constrained and try to solve it by learning a generative prior rather than relying on explicit reconstruction.\n\nThe pipeline itself is described clearly enough that someone could re-implement the high-level flow. The choice to decode through a universal prior model is sensible for keeping the output compatible with real-time rendering.\n\nThe soft spot is the lack of any visible evidence for the central claim. The abstract states that the avatars faithfully represent diverse identities and beat recent methods, yet supplies no identity similarity scores, no cross-pose or cross-ethnicity splits, no ablation on the diffusion or refinement stages, and no error bars. Without those, the stress-test concern stands: the diffusion model must hallucinate consistent back and side views, and nothing here shows it does so reliably rather than averaging over frontal-biased training data.\n\nThis is for graphics and VR groups that need quick avatar pipelines. A reader already working on feed-forward 3D generation or Gaussian splatting would find the architecture useful to examine, even if the results section needs more scrutiny.\n\nIt deserves peer review because the feed-forward claim is practically relevant and the components are standard enough that referees can evaluate the combination directly.","headline":"FiCA assembles a feed-forward pipeline for single-portrait Gaussian avatars without test-time optimization, but the abstract gives no numbers or ablations to check whether the diffusion step actually delivers identity-preserving meshes on diverse inputs.","tokens_in":2371,"tokens_out":434,"would_cite":false,"duration_ms":14634,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.3","headline":"A single portrait image produces a drivable photorealistic 3D Gaussian avatar in one feed-forward pass.","keywords":["Gaussian avatars","feed-forward generation","single-image avatar","diffusion model","3D mesh reconstruction","photorealistic rendering","real-time animation","codec avatars"],"falsifier":"Run the pipeline on single portraits of subjects with rare head shapes or extreme lighting, then compare the generated avatar's rendered novel views against multi-view ground-truth captures of the same person; large deviations in identity or geometry would falsify the mapping claim.","tokens_in":2604,"feed_emoji":"👤","tokens_out":574,"duration_ms":11677,"temperature":0.7,"pith_summary":"The paper introduces FiCA, a pipeline that combines human-centric foundation models with a diffusion model to infer complete 3D head geometry and appearance from one portrait. A feed-forward refinement network then improves fidelity, after which a universal prior converts the mesh into 3D Gaussians that support real-time animation with new expressions. A sympathetic reader would care because prior avatar methods typically demand multiple views or slow per-person optimization at test time, restricting use in consumer applications. The system claims to deliver identity-preserving results that match or exceed the visual quality of slower recent approaches while remaining instantaneous.","feed_headline":"Single portrait yields instant drivable 3D Gaussian avatar","feed_subtitle":"Diffusion mapping plus feed-forward refinement produces photorealistic heads that animate in real time without per-person tuning.","key_machinery":"Diffusion model that maps partial visual observations to complete 3D mesh reconstruction, followed by a feed-forward refinement network and a universal prior that decodes the mesh into 3D Gaussians.","core_discovery":"FiCA learns a generative mapping from partial single-portrait observations to complete and authentic 3D mesh reconstructions via a diffusion model, augments this with a feed-forward mesh refinement network that removes the need for person-specific test-time optimization, and decodes the resulting mesh through a universal prior into a set of 3D Gaussians that render as photorealistic, expression-drivable avatars.","pith_inferences":["The same diffusion-plus-refinement structure could be tested on single-image reconstruction of full bodies or hands.","Real-time performance might allow direct integration into live video or mobile AR without cloud processing.","If the universal Gaussian decoder generalizes across identities, it could reduce the need for large per-avatar training datasets in future work."],"forward_implications":["Avatars faithfully represent diverse identities from single images.","Generated avatars surpass the visual quality of recent competing methods.","No person-specific test-time optimization is required.","Photorealistic 3D Gaussian avatars support real-time driving with novel expressions.","The full pipeline operates in a single feed-forward pass."],"fun_headline_variants":["FiCA turns one portrait into drivable 3D Gaussian avatar","Single image to Gaussian avatar without per-person tuning","Diffusion plus refinement yields instant Gaussian heads","Feed-forward Gaussian avatars from portrait in real time"],"cache_read_input_tokens":2112,"weakest_assumption_plain":"The diffusion model can learn a reliable mapping from one partial portrait view to a full, identity-preserving 3D head mesh without any person-specific optimization.","fun_headline_variants_meta":{"raw":{"variants":["FiCA turns one portrait into drivable 3D Gaussian avatar","Single image to Gaussian avatar without per-person tuning","Diffusion plus refinement yields instant Gaussian heads","Feed-forward Gaussian avatars from portrait in real time"]},"model":"grok-4.3","cost_usd":0.005535,"raw_usage":{"total_tokens":2650,"prompt_tokens":657,"num_sources_used":0,"completion_tokens":59,"cost_in_usd_ticks":55349500,"prompt_tokens_details":{"text_tokens":657,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":1934,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":657,"tokens_out":59,"duration_ms":13157,"temperature":1.0,"reasoning_tokens":1934,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-06-26T00:39:03.456487+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"Run the pipeline on single portraits of subjects with rare head shapes or extreme lighting, then compare the generated avatar's rendered novel views against multi-view ground-truth captures of the same person; large deviations in identity or geometry would falsify the mapping claim.","supporting_citations":[],"review_version":1}