{"id":"bcc3c65a-acac-4922-8c65-4dc14c769eed","arxiv_id":"2504.12825","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A method that generates a temporally consistent, textured 4D mesh sequence from only two RGB images showing an object's initial and final poses.","lead":"TwoSquared creates a dynamic 3D animation from just two photos of an object at the start and end of a motion. It builds 3D models of both frames, then computes a smooth, physically inspired deformation between them, producing a texture-consistent animation at any frame rate.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central claim is conditional on Hunyuan3D producing two artifact-free, mutually consistent keyframe meshes; the paper acknowledges this in Sec.","rationale":"After reading the full manuscript, I agree with the reader that the load-bearing premise is the quality and mutual consistency of the Hunyuan3D keyframe meshes. The abstract's claim is broad ('physically plausible sequence ... in-the-wild images'), and Sec. 3.1 states the pipeline can be integrated with any 3D generation method, but every quantitative result in Tab. 1 is obtained with a single backbone on two human sequences. The paper's own robustness section confines success to cases where the endpoint meshes are free of significant artifacts, and its one failure-case visualization does not measure how often that condition fails or how sensitive the losses and thresholds are to backbone choice. This is not an internal contradiction—the deformation model is a sensible interpolation of whatever meshes it receives—but it means the central claims of generality and superiority are conditional on an external component that the paper neither controls nor benchmarks. Swapping the backbone is the cleanest falsifiable check because the paper explicitly claims flexibility in Sec. 3.1. I also considered whether the variational notation in Eqs. (3)-(7) is formally imprecise; that is a presentation issue that code release would resolve, but the backbone dependency is the more substantive risk to the headline claim. Since the reader already assigned CONDITIONAL with high confidence, my analysis does not change the verdict.","tokens_in":15761,"tokens_out":10192,"duration_ms":115060,"concrete_test":"Replace Hunyuan3D with a different image-to-3D backbone (e.g., InstantMesh or TripoSR) while keeping the rest of the pipeline fixed, and run it on the same 4D-DRESS pairs; compare CD, HD, and SAσ against the reported results. If the metrics degrade substantially (e.g., by more than 2x) or the pipeline requires retuned thresholds, loss weights, or additional alignment to avoid collapse, then the method is not backbone-agnostic as claimed and the paper should be revised to scope its claims to the specific backbone tested.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The method's central claim—that it generates physically plausible, texture- and geometry-consistent 4D sequences from arbitrary image pairs—rests on the premise that Hunyuan3D produces two keyframe meshes that are individually accurate and mutually compatible enough for Diff3F features and functional-map refinement to yield meaningful correspondences. The authors themselves state in Sec. 4.4 that 'as long as the starting and ending meshes are generated without significant artifacts, the intermediate shapes remain physically plausible,' making this an explicit load-bearing premise. This premise is not quantified anywhere: both quantitative sequences in Tab. 1 are human motions from 4D-DRESS, there is no experiment replacing Hunyuan3D with another image-to-3D model, no measurement of how often Hunyuan3D produces artifacts on the claimed in-the-wild categories, and the single failure case (Fig. 8) is qualitative. Therefore the headline claims of 'superior performance' and template-free generality are established only for the particular combination of Hunyuan3D plus 4D-DRESS humans; for other object classes or backbones they remain plausible but unverified. This does not make the paper's internal construction wrong—it makes the central claim conditional in exactly the way the reader's verdict states.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces TwoSquared, a method for generating a 4D (dynamic 3D) sequence from only two RGB images that depict the initial and final states of an object. The pipeline decomposes the problem into three components: (i) an image-to-3D generation block based on Hunyuan3D that produces keyframe meshes; (ii) a vertex registration block that extracts per-vertex features with Diff3F, refines them via functional-map optimization, and enforces bidirectional loop-closure consistency to obtain sparse correspondences with confidence weights; and (iii) a shape deformation block that learns a time-continuous velocity field parameterized by a small MLP, trained with a combination of smoothness, overlapping, normal, stretching, confidence-guided matching, and distortion losses. The authors report quantitative results on two sequences from 4D-DRESS, comparing against image-morphing baselines (DiffMorpher, DreamMover) and the 3D-to-4D method 4Deform, and qualitative results on web images, sketches, and daily photographs. The central claims are that the method is template-free, class-agnostic, physically plausible, and that it outperforms prior approaches and even reconstruction from ground-truth intermediate images on the tested sequences.","tokens_in":16014,"tokens_out":6439,"duration_ms":62668,"significance":"If the claims hold, TwoSquared offers a simple and practical formulation for 4D generation from minimal input, with the appealing property of arbitrary-frame-rate inference at near-real-time cost after a short per-pair optimization. The decomposition into a mature image-to-3D backbone and a physics-inspired deformation module is conceptually clean, and the authors provide a code link, use a purely JAX implementation, and include an interesting control baseline that reconstructs the shape from ground-truth intermediate images. The physical loss terms (overlapping, normal, stretching, confidence-guided matching) are reasonable in intent, and the qualitative results on in-the-wild images are visually compelling. However, the evidence currently supports a much narrower claim than the abstract and introduction suggest: the quantitative evaluation is limited to two human-motion sequences, the method's success is explicitly conditional on the quality of the external Hunyuan3D backbone, and several implementation-critical hyperparameters are undisclosed.","major_comments":[{"comment":"The quantitative evaluation is restricted to two sequences (Take22 and Take7) from 4D-DRESS, both capturing human motions from the same dataset. No error bars, repeated runs, or statistical tests are reported, and the differences between methods in several rows are small (e.g., CD values around 0.074–0.126 in Take7). This is load-bearing for the claim that TwoSquared achieves 'lower error rates compared to other approaches' and 'even outperforms results generated directly from ground truth intermediate images.' Please either expand the evaluation to more sequences and object categories, report variance over multiple runs with different seeds, or explicitly scope the claim to the two tested sequences.","section":"§4.2, Table 1"},{"comment":"The total loss in Eq. (15) depends on six scalar weights λv, λo, λn, λs, λm, λd, and the correspondence filter in Eq. (1) depends on a distance threshold δd. None of these values is reported in the main paper or the supplementary material. Since the method trains a small network per image pair, these hyperparameters are essential for reproducibility, and the ablation in Table 2 shows that the final metrics are sensitive to which loss terms are active. Please disclose all hyperparameter values and, ideally, report a sensitivity study.","section":"§3.4, Eq. (15); §4.1"},{"comment":"The abstract and introduction assert that TwoSquared generates 'physically plausible' sequences from arbitrary in-the-wild image pairs without object-class priors. However, §4.4 states that plausibility holds 'as long as the starting and ending meshes are generated without significant artifacts,' making the quality of Hunyuan3D a load-bearing premise. There is no quantitative measurement of how often Hunyuan3D produces artifact-free keyframes on the claimed in-the-wild categories, no experiment that replaces Hunyuan3D with another image-to-3D backbone, and the only failure case (Fig. 8) is qualitative. Consequently, the method's generality is verified only for 4D-DRESS human motions. Please either add a quantitative robustness study or explicitly reframe the claims as conditional on the backbone's output quality.","section":"§4.4, with §1 and §5"},{"comment":"The distortion loss as written appears dimensionally inconsistent. Both 1/6 Tr(D)² and 1/2 Tr(D·D)² are scalars, so the Frobenius norm is simply the absolute value of their difference; squaring the trace of D·D is unusual, and the expression is likely a typo for something like Tr(D)² - Tr(D·D) or a norm of the matrix D·D. Since this term is part of the total loss and is claimed to follow [38], please correct the formula or clarify the intended expression.","section":"§3.3.2, Eq. (14)"}],"minor_comments":[{"comment":"The text refers to 'Kinect energy'; this should be 'kinetic energy.' Also, the constraint notation in Eq. (4), ∫Ω0 X(0) dx = P0, is confusing because Ω0 is described as a subset of the point domain; please clarify the measure used and the relation between Ω0 and the sampled point cloud.","section":"§3.3.1, Eq. (3)–(6)"},{"comment":"The confidence map Cij = (δd − D_j^i) / max_j(D_j^i) uses the index j both for the target shape and as the variable in the maximum, which is ambiguous. Please rewrite with explicit per-vertex notation, e.g., C(xi) = (δd − D(xi)) / max_{xk} D(xk) for correspondences that pass the loop-closure filter.","section":"§3.2, Eq. (2)"},{"comment":"The text says 'we extract five keyframes from two sequences ... and generate four 4D deformation sequences,' but it is unclear how many intermediate frames are evaluated per sequence and what the temporal spacing is. Please specify the exact keyframe indices and the number of compared intermediate steps.","section":"§4.2, Table 1"},{"comment":"The supplementary reports that Diff3F feature extraction takes 5 minutes per frame and the velocity net trains in under 2 minutes, but the total 'roughly 17 minutes to train' does not include the 3 minutes per Hunyuan3D mesh generation; please make the pipeline timing consistent and state the total wall-clock time for one pair.","section":"Supplementary, S0"},{"comment":"The qualitative evaluation on web images is presented only as images, with no quantitative or user study. Since the paper claims robustness to 'varying input quality,' a simple user study or at least a larger set of diverse inputs would strengthen the claim.","section":"§4.1, Dataset"},{"comment":"The comparison against 4Deform uses the same potentially noisy correspondences as TwoSquared, which is a good ablation of the deformation module, but it is not a comparison of the full methods in 4Deform's intended setting (high-quality correspondences). Please state this explicitly to avoid overgeneralizing the comparison.","section":"§4.2, Baselines"}],"recommendation":"major_revision","confidential_remarks":"The paper leans heavily on the authors' own prior work ([37] and [38]), and the only quantitative evaluation uses human-motion sequences that are the same category as those works. The novelty is mainly the integration of an image-to-3D backbone with the existing deformation machinery, and the 'first to address 4D generation from image pairs' claim is plausible but would benefit from a careful positioning against [37]/[38]. The missing hyperparameters and the narrow evaluation are the main obstacles to acceptance; if the authors provide a supplementary table of all hyperparameters, add error bars or more sequences, and moderate the claims to match the conditional nature of the results, the paper could become a solid contribution."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nHere's my read on TwoSquared. The genuinely new thing is the task: generating a continuous 4D sequence from just two RGB images of start and end states. That's not in prior work. The decomposition is clean — image-to-3D with Hunyuan3D, registration via Diff3F features refined with functional maps and loop closure, then deformation with a velocity field MLP under physically inspired losses. The integration is sensible, and the speed advantage is real: after the 3D generation step, the deformation trains in under two minutes and infers arbitrary frame rates in under a second.\n\nThe paper does several things well. It compares against 4Deform, its closest relative, and the robustness to noisy correspondences looks real — the loop-closure filtering and confidence-weighted matching are reasonable engineering. The qualitative results, including the sketch and web-image examples, support the claim of generality. The ablation study is standard, and the supplementary gives concrete timings. The authors also clearly state the load-bearing assumption in Sec. 4.4: as long as the start and end meshes are artifact-free and mutually consistent, the intermediates are plausible. That honesty is worth crediting.\n\nNow the soft spots, in proportion. The quantitative evaluation is the big one: two sequences from 4D-DRESS, no error bars, no statistical testing, and the reported gain over 4Deform on Take7 is small in absolute terms. The loss weights and the correspondence distance threshold are not disclosed, so the method as described is not fully reproducible. The claim of superiority over reconstruction from ground-truth intermediate images rests on two examples and could reflect Hunyuan3D's per-frame inconsistency rather than a general property. And the 'physically plausible' language is doing a lot of work — the constraints are geometric (smoothness, normals, stretch), not actual physics. That is fine, but it should be labeled as such.\n\nNone of this makes the paper wrong. The central premise — that the deformation module is only as good as the two keyframe meshes — is acknowledged, not hidden. But it means the headline claim of 'physically plausible, texture- and geometry-consistent 4D from arbitrary pairs' is established only for the particular backbone and two human sequences; for other object classes it remains plausible but unverified.\n\nThis paper deserves a serious referee. It opens a new task and ships a working pipeline, but the revision should include the missing hyperparameters, a larger evaluation with variance, a backbone swap, and a failure-rate analysis on in-the-wild inputs. I would send it out.","headline":"A genuinely new two-image-to-4D task with a clean, honest pipeline, but the quantitative evidence is thin and the central claim is conditional on the 3D backbone.","tokens_in":16547,"tokens_out":2113,"would_cite":true,"duration_ms":20699,"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 full 4D motion sequence can be generated from just two 2D photographs: one of the object at the start of an action and one at the end.","keywords":["4D generation","image pair","velocity field","shape deformation","functional maps","diffusion features","texture consistency","physically plausible motion"],"falsifier":"Take a pair of images for which the image-to-3D backbone produces one mesh with a missing limb or large hole, run TwoSquared, and check the intermediate meshes: if the artifact appears in every intermediate frame and Chamfer distance to ground-truth scans grows accordingly, the load-bearing premise — that two good keyframe meshes are required — is violated.","tokens_in":15592,"feed_emoji":"🎬","tokens_out":6758,"duration_ms":62815,"temperature":0.7,"pith_summary":"TwoSquared claims that a complete 4D motion sequence — a continuously deforming 3D object — can be generated from just two ordinary 2D photographs of the same object, one at the start and one at the end of an action. The paper's method avoids generating intermediate images or relying on category templates: it lifts each input image to a 3D mesh with an existing image-to-3D model, finds correspondences between the two meshes, and learns a time-continuous velocity field that physically deforms the first mesh into the second. The authors report that the resulting sequences stay texture-consistent and geometry-consistent on real scanned human motions, and that the approach beats image-morphing baselines, the 4Deform baseline, and even direct reconstruction of ground-truth intermediate frames. A sympathetic reader would care because this turns 4D content creation into a two-image problem with no class-specific priors and arbitrary frame-rate output.","feed_headline":"TwoSquared turns a start and end photo into a 4D motion sequence","feed_subtitle":"No video, no templates: the method lifts two images to 3D, then interpolates a physically plausible path between them.","key_machinery":"The load-bearing mechanism is a time-continuous velocity field $V: \\mathbb{R}^3 \\times [0,T] \\to \\mathbb{R}^3$, implemented as an 8-layer MLP (the Velocity Net), which maps each point of the source mesh to its trajectory so that intermediate shapes are obtained by Euler integration. The field is trained against six losses: a kinetic-energy smoothing term, an overlapping Chamfer loss pulling the endpoint of the trajectory onto the target mesh, a normal-alignment loss using the deformation gradient, a stretching penalty, a confidence-weighted matching loss, and a distortion term. The correspondences that feed the matching loss come from per-vertex diffusion features (Diff3F) refined by a functional-map optimization and a bidirectional close-loop consistency check, which filters unreliable matches.","core_discovery":"The central discovery is that 4D generation from two RGB images is best approached by decomposition rather than end-to-end synthesis: first hallucinate the two endpoint geometries in 3D, then solve for a physically plausible path between them. On the 4D-DRESS real-scanned human dataset, TwoSquared reports lower Chamfer and Hausdorff distances than per-frame reconstruction from morphing baselines and than 4Deform, and it also outperforms lifting ground-truth intermediate images with the same backbone, which the paper attributes to the inconsistency of independent per-frame generation. The deformation is trained with losses for kinetic-energy smoothness, endpoint overlap, normal alignment, stretching, confidence-weighted matching, and distortion, so that the interpolated meshes preserve surface properties rather than merely blending shapes.","pith_inferences":["A testable prediction of the two-stage design is that its advantage over per-frame reconstruction grows with pose change and with the inconsistency of the image-to-3D backbone across unrelated viewpoints; this is not directly measured in the paper.","The velocity-field scaffold is, in principle, agnostic to the correspondence source and the 3D backbone; the paper states this flexibility but only demonstrates Hunyuan3D with Diff3F features.","The 'physically plausible' label comes from an optimization objective, not from a physics simulation; a stricter test would compare predicted trajectories against motion capture or optical-flow-derived motion rather than mesh distances.","Symmetry and featureless regions remain a likely failure zone, as the paper itself flags hair as an open problem where thousands of simultaneous small deformations exceed the tracking quality available."],"forward_implications":["Since the velocity field is continuous in time, the same trained model outputs 4D sequences at any frame rate without retraining; the paper recommends $T' < 2T$ for high frame rates.","The pipeline needs no templates or object-class priors, so it applies to humans, animals, sketches, and everyday photographs, and even to pairs of different subjects of the same species for pose transfer.","The closed-loop correspondence check and confidence weighting let the method keep intermediate shapes physically plausible even when registration fails completely in some regions, such as a leg with no matched correspondences.","The method is cheaper than morphing-based pipelines at inference: after about 17 minutes per pair, generating a mesh takes under one second, whereas baselines regenerate images and meshes per frame.","Deforming one consistent source mesh outperforms reconstructing each ground-truth intermediate image with the same image-to-3D backbone, indicating that per-frame lifting is the main source of texture and geometry inconsistency."],"supporting_citations":[{"why":"Supplies the image-to-3D backbone that generates the two keyframe meshes from the input images.","marker":"[46]"},{"why":"Provides per-vertex diffusion features used to compute the cosine-similarity correspondence maps between the two meshes.","marker":"[10]"},{"why":"Introduces the functional map representation used to treat and refine the noisy correspondence matrix.","marker":"[31]"},{"why":"Delivers the smooth discrete optimization algorithm that refines point-to-point correspondences before the loop-closure check.","marker":"[27]"},{"why":"Contributes the velocity-field formulation and the smoothing operator that the Shape Deformation Block is built on.","marker":"[37]"},{"why":"Is the 3D-to-4D baseline whose stretching and distortion losses and surface-area metric (SAσ) are reused in TwoSquared.","marker":"[38]"},{"why":"Provides the 4D-DRESS real-scanned human motion dataset with ground-truth intermediate meshes used in quantitative evaluation.","marker":"[48]"},{"why":"Serves as a 2D image-morphing baseline whose intermediate frames are lifted to 3D for comparison.","marker":"[59]"},{"why":"Serves as the second 2D image-morphing baseline, also lifted per frame to 3D.","marker":"[40]"}],"fun_headline_variants":["Two images in, 4D motion out: TwoSquared","From two stills to a 4D scene: TwoSquared","TwoSquared: 4D from just two photos","No video needed: TwoSquared builds 4D from two RGB images","TwoSquared: start and end photos to 4D motion"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The entire 4D sequence inherits the quality of the two keyframe meshes produced by the external image-to-3D model: if either mesh contains significant artifacts, the deformation stage will faithfully interpolate that error rather than repair it.","fun_headline_variants_meta":{"raw":{"variants":["Two images in, 4D motion out: TwoSquared","From two stills to a 4D scene: TwoSquared","TwoSquared: 4D from just two photos","No video needed: TwoSquared builds 4D from two RGB images","TwoSquared: start and end photos to 4D motion"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000565,"raw_usage":{"total_tokens":2652,"prompt_tokens":891,"completion_tokens":1761,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":507,"completion_tokens_details":{"reasoning_tokens":1669}},"tokens_in":507,"tokens_out":1761,"duration_ms":12647,"temperature":1.0,"reasoning_tokens":1669,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T12:20:40.921586+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a pair of images for which the image-to-3D backbone produces one mesh with a missing limb or large hole, run TwoSquared, and check the intermediate meshes: if the artifact appears in every intermediate frame and Chamfer distance to ground-truth scans grows accordingly, the load-bearing premise — that two good keyframe meshes are required — is violated.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides per-vertex diffusion features used to compute the cosine-similarity correspondence maps between the two meshes."},{"cited_title":"Functional maps: a flexible representation of maps between shapes","cited_arxiv_id":null,"evidence_quote":"Introduces the functional map representation used to treat and refine the noisy correspondence matrix."},{"cited_title":"Smooth non-rigid shape matching via effective dirichlet energy optimization","cited_arxiv_id":null,"evidence_quote":"Delivers the smooth discrete optimization algorithm that refines point-to-point correspondences before the loop-closure check."},{"cited_title":"Implicit neural surface deformation with explicit velocity fields","cited_arxiv_id":null,"evidence_quote":"Contributes the velocity-field formulation and the smoothing operator that the Shape Deformation Block is built on."},{"cited_title":"4deform: Neural surface deformation for robust shape interpolation","cited_arxiv_id":null,"evidence_quote":"Is the 3D-to-4D baseline whose stretching and distortion losses and surface-area metric (SAσ) are reused in TwoSquared."},{"cited_title":"4d-dress: A 4d dataset of real-world human clothing with se- mantic annotations","cited_arxiv_id":null,"evidence_quote":"Provides the 4D-DRESS real-scanned human motion dataset with ground-truth intermediate meshes used in quantitative evaluation."},{"cited_title":"DiffMorpher: Unleashing the Capability of Diffusion Models for Image Morphing","cited_arxiv_id":"2312.07409","evidence_quote":"Serves as a 2D image-morphing baseline whose intermediate frames are lifted to 3D for comparison."}],"review_version":1}