{"id":"7171df2a-ff8f-47e1-aa9d-1903404ea31f","arxiv_id":"2507.15365","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Models trained only on a 300K procedurally generated human dataset achieve accuracy comparable to 2B-parameter foundation models on depth, surface normals, and matting, at a fraction of the compute.","lead":"DAViD trains depth, surface-normal, and portrait-mask prediction models entirely on 300,000 computer-generated human images, and shows they match or beat much larger models trained on millions of real photos. The work suggests that high-quality synthetic data can replace huge real-image collections for human-centric vision, cutting training and inference cost while preserving accuracy.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'synthetic-only' claim depends on an undisclosed backbone initialization: the paper never states whether the DPT/ViT encoder starts from ImageNet/MiDaS-pretrained weights, which would make the data-efficiency and compute comparisons non-apples-to-apples.","rationale":"The reader's weakest assumption is the synthetic-to-real domain gap. That is a reasonable concern, but the paper's own matting results (PhotoMatte85 SAD 5.85 vs. MODNet 13.94, Table 3) and cross-dataset depth results provide substantial evidence that the renderer transfers; the domain gap is therefore not the least secure link. The normals ground-truth quality issue (App. B.1, Table 5) is real but affects the evaluability of the benchmark rather than the source of the model's capability. The most load-bearing unexamined assumption is attribution: the paper claims synthetic-only training but never discloses backbone initialization. DPT is an architecture with commonly used pretrained weights; if those weights are loaded, then the central 'synthetic data alone' claim is false, and the compute comparison in Sec. 7 undercounts pretraining. The absence of training code makes this impossible to verify from the paper. A random-init retraining experiment would settle it directly. The reader's CONDITIONAL verdict remains appropriate, but the condition should explicitly include disclosure and verification of initialization, not only renderer realism.","tokens_in":19996,"tokens_out":9615,"duration_ms":115895,"concrete_test":"Release or inspect the training script and model checkpoints from https://aka.ms/DAViD, and retrain the Large depth model from random initialization under the Sec. 4.1 hyperparameters. If the random-init model's Goliath and Hi4D RMSE degrades materially below Table 1's 0.253 and 0.072 (or fails to converge), then the headline result depends on real-image pretrained DPT weights, and the 'synthetic-only' claim must be conditioned on that initialization.","verdict_should_be":"UNCHANGED","load_bearing_attack":"DAViD's central claim is that a single 300K synthetic dataset is sufficient, and the paper repeatedly says models are 'trained entirely on synthetic data' (Sec. 3.1, Abstract) and that the approach 'eliminates the need for data mixing' (Sec. 2). Section 3.2 states the encoder 'follows DPT's encoder' with Readproj, and Sec. 4.1 gives optimizer, epochs, and resolution, but neither section states the initialization. DPT is commonly released with ViT weights pretrained on ImageNet and MiDaS, i.e., on real images and real-derived depth. If those weights are used as the starting point, the models are not trained solely on synthetic data: the visual representation inherits a large real-data pretraining, and the attribution of the results to SynthHuman alone is not supported as stated. The compute comparison in Sec. 7 ('we can train ~800 models with the compute used to train a single Sapiens-2B') also omits pretraining cost if the backbone was pretrained. Because the training code is not released, an independent reader cannot determine whether the observed accuracy comes from SynthHuman or from the DPT initialization. This is more load-bearing than the renderer domain gap: the matting results (PhotoMatte85 SAD 5.85, Table 3) already give strong evidence that the synthetic renderer transfers to real images, so the weakest link is whether the claimed 'synthetic-only, no-pretraining' framing is actually what was trained.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces DAViD, a family of DPT-style vision models trained on SynthHuman, a procedurally generated dataset of 300K high-fidelity synthetic human images with dense ground truth for relative depth, surface normals, and soft foreground segmentation. Using a single architecture and a single synthetic dataset, the authors report accuracy competitive with or better than much larger foundation models such as Sapiens-2B on real benchmarks (Goliath, Hi4D, PhotoMatte85, PPM-100), while requiring far fewer FLOPs and parameters. The paper also provides ablations on data source, dataset size, model size, and multi-task training, and it releases the dataset and trained models.","tokens_in":20287,"tokens_out":4786,"duration_ms":56098,"significance":"If the central claim holds, the paper demonstrates a qualitatively different scaling story for human-centric dense prediction: a relatively small, carefully designed synthetic dataset can replace massive real-data pretraining and produce state-of-the-art results on multiple tasks. The contribution is potentially significant for privacy, provenance, and compute efficiency. The paper's strengths include cross-dataset evaluation on real benchmarks with no training-to-test circularity, careful ablations isolating data source and model size, transparent reporting of benchmark construction details (Appendix B.7), and an honest discussion of failure cases and of noise in the surface-normal ground truth (Appendix B.1). The main risk is that the paper's strongest claims are stated more broadly than the evidence supports, and one load-bearing implementation detail—the initialization of the encoder—is not disclosed.","major_comments":[{"comment":"The manuscript never states whether the DPT/ViT encoder is initialized from pretrained weights (e.g., ImageNet, MiDaS, or DPT checkpoints) or trained from random initialization. This is load-bearing for the abstract's claim of training 'entirely on synthetic data' and for the compute comparison in Sec. 7, because if the encoder starts from a real-data-pretrained backbone, the models are not trained solely on synthetic data and the pretraining compute is omitted. Please disclose the initialization and, ideally, provide an ablation training from random initialization; without this, the attribution of the results to SynthHuman is not established.","section":"Sec. 3.2 and Sec. 4.1"},{"comment":"The abstract's 'no loss in accuracy' is contradicted by the surface-normal results in Table 2: on Hi4D, Ours-Large has mean angular error 15.37° versus 12.14° for Sapiens-2B, and on Goliath UpperBody and FullBody the gaps are 13.96° vs. 10.65° and 14.60° vs. 11.49°, respectively. The depth numbers in Table 1 also show a small aggregate gap on the averaged-over-all row (Ours-Large RMSE 0.191 vs. Sapiens-2B 0.170), although Ours-Large is better on several Goliath subsets. The accuracy claim should be qualified to reflect which tasks, subsets, and metrics are actually on par or better.","section":"Abstract and Table 2"},{"comment":"Table 2's caption states 'All results on the Hi4D dataset are taken from [17],' yet the table includes rows for Ours-Base and Ours-Large, which cannot come from [17]. Please clarify which Hi4D numbers were computed by the authors, which were copied from Sapiens, and whether the same preprocessing and foreground masks were used for all entries. The current wording makes the comparison protocol ambiguous and affects the validity of the surface-normal claims.","section":"Table 2 caption and Sec. 4.2"},{"comment":"The claimed efficiency advantage—'we can train ~800 models with the compute used to train a single Sapiens-2B'—is not substantiated with a concrete compute accounting. No total GPU-hours are reported for DAViD training, and the Sapiens reference includes only the 18-day/1,024-A100 pretraining cost, explicitly excluding fine-tuning. If the DAViD backbone is itself pretrained (see the first major comment), that pretraining cost must be included. Please provide an apples-to-apples training-cost table.","section":"Sec. 7"}],"minor_comments":[{"comment":"The observation that blurring the predicted normals improves all Goliath metrics (Table 5) is an important caveat for the surface-normal comparisons and should be stated in the main text near the normals results, not only in the supplementary material.","section":"Appendix B.1 and Sec. 4.3"},{"comment":"The text cites 'Depth-Pro [46]' when discussing the depth comparison, but reference [46] in the bibliography is Depth Anything v2; the intended citation appears to be [2]. Please correct the citation.","section":"Sec. 4.3"},{"comment":"References [33] and [34] are the same paper (Ranftl et al., 'Towards robust monocular depth estimation'); please merge or differentiate them appropriately.","section":"References"},{"comment":"The implementation details report a batch size of 24 'on each GPU of a A100 ×4 compute node'; please state the total batch size and whether all four GPUs were used for each task, and report the number of training runs or seeds for the main results.","section":"Sec. 4.1"}],"recommendation":"major_revision","confidential_remarks":"The core empirical finding—that a 300K-image high-fidelity synthetic dataset can yield competitive dense-prediction models—is credible and well supported by the cross-dataset evaluations and ablations. The main correctness risk is the undisclosed encoder initialization; if the authors confirm a random-init baseline and provide the corresponding ablation, the paper's central claim becomes much stronger. If they cannot, the 'synthetic-only' and compute-efficiency claims would need to be substantially revised. I recommend major revision rather than rejection because the issue is local and fixable within the paper's scope."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague, here's my honest read.\n\nThis paper's real result is a 300K-image procedural human dataset, SynthHuman, with strand-level hair and clean dense labels, and a compact DPT-style model trained on it that beats large foundation models on several human depth and matting benchmarks, and comes close on normals. The matting numbers—SAD 5.85 on PhotoMatte85—are the strongest evidence; cross-dataset transfer to real photos is credible. The ablations in Table 4, showing data source and size effects, support the claim that fidelity and diversity, not scale, drive accuracy. The unified architecture across three tasks is a useful contribution.\n\nNow the soft spots. The abstract says 'no loss in accuracy,' but Table 2 shows Ours-Large's Hi4D mean normal error is 15.37° versus Sapiens-2B's 12.14°—a real gap. The authors themselves show blurring their normals output improves Goliath scores (Appendix B.1), meaning the normals benchmark cannot measure fine detail. There are no error bars, and the Resizer module has no quantitative ablation. The custom Goliath subset and re-evaluated baseline numbers complicate comparison with published scores, though the choices are documented.\n\nThe bigger issue is backbone initialization. The paper says 'trained entirely on synthetic data' and claims to 'eliminate the need for data mixing,' but it never states whether the ViT/DPT encoder starts from ImageNet- or MiDaS-pretrained weights. If it does, that is real-data pretraining, and the compute comparison against Sapiens—'~800 models'—omits that cost. This is not about the renderer domain gap; matting results already show the renderer transfers. It is a missing fact that changes how the synthetic-only claim should be read.\n\nThis deserves a serious referee. The direction—small high-fidelity synthetic datasets replacing massive real-data pretraining—is important and testable, and releasing SynthHuman will help the field. Send it to review, but the authors need to state the initialization, soften the normals claim, and report variance. I would cite the dataset if those details hold up, but not yet.","headline":"Powerful synthetic-data result, but hidden pretrained weights and overclaimed normals stop it short of the abstract's promise.","tokens_in":20909,"tokens_out":3616,"would_cite":true,"duration_ms":38193,"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 high-fidelity synthetic dataset of 300K human images trains depth, surface-normal, and matting models that match or beat 2B-parameter foundation models on real benchmarks.","keywords":["synthetic data","dense prediction","depth estimation","surface normals","soft foreground segmentation","procedural generation","data efficiency","human-centric vision"],"falsifier":"Measure the depth and normal models on a real test set whose ground truth is captured at finer fidelity than Goliath and Hi4D (for example, light-stage scans that resolve hair strands and wrinkles). If the models' fine-detail advantage over Sapiens-2B does not appear there, then the paper's attribution of accuracy to high-fidelity synthetic ground truth is refuted; if blurring the predictions still helps on that finer benchmark, the benchmark itself is the bottleneck, not the model.","tokens_in":19764,"feed_emoji":"🧍","tokens_out":9483,"duration_ms":94779,"temperature":0.7,"pith_summary":"The paper claims that a single high-fidelity synthetic dataset of 300,000 procedurally generated human images is enough to train accurate models for relative depth, surface normals, and soft foreground segmentation on real photographs, with no real training data at all. Its models, built on one shared architecture, match or beat foundation models of up to 2 billion parameters on real benchmarks while running roughly 16 times faster and using a fraction of the training compute. The point is a change in philosophy: carefully rendered synthetic data with pixel-perfect ground truth can replace vast real-image collections and multi-stage training pipelines, and it carries guarantees on consent, provenance, and diversity that real data cannot offer.","feed_headline":"300K synthetic images match or beat 2B-parameter vision models","feed_subtitle":"A 300K-image procedural dataset matches foundation-model accuracy on depth, normals, and matting at a fraction of the compute.","key_machinery":"The engine is the SynthHuman data-generation pipeline and its perfect labels: procedural sampling of bodies, expressions, poses, textures, hair, accessories, clothing, and environments, with two technical choices that make the ground truth trainable. Hair normals are re-defined by voxelizing the strand geometry and running marching cubes to build a coarse proxy mesh, so ground-truth normals follow hair volume and clumps instead of aliasing from individual strands; translucent surfaces such as glasses lenses are rendered so the depth and normals behind them can be included or ignored for different applications. On the model side, a DPT-style ViT encoder runs at a fixed 384×384 resolution while a light fully convolutional \"Resizer\" carries full-resolution features into the decoder, so fine detail is predicted without the quadratic self-attention cost of high-resolution tokens; the same architecture serves all three tasks, with only the output channels and losses changing.","core_discovery":"The central discovery is that the fidelity of the ground truth, not the size of the model or dataset, is what limits accuracy for dense human-centric prediction. Scans used by earlier synthetic pipelines (THuman, RenderPeople) represent hair as a coarse mesh and miss fine structure, and their ground truth carries scanning artifacts; the paper's SynthHuman pipeline instead renders strand-level hair, expressive faces with dynamic wrinkle maps, artist-created clothing and accessories, and varied HDRI and 3D environments, producing depth, normal, and alpha ground truth that resolves details real benchmarks cannot even record. A model trained only on these 300K images achieves depth root-mean-square error of 0.072 on Hi4D versus 0.095 for the 2.16B-parameter Sapiens model, on-par surface-normal accuracy with a fraction of the parameters, and matting SAD of 5.85 on PhotoMatte85 versus 13.94 for MODNet, using a single architecture and no task-specific add-ons.","pith_inferences":["The same recipe may transfer to other domains where dense ground truth is physically hard to capture, such as hands, animals, or transparent and reflective objects: strand-level or artist-authored geometry with physically motivated rendering could supply the fine-detail labels that real sensors and scans miss.","The paper's Appendix B.1 shows that blurring its normal predictions improves scores on Goliath, a sign that the real normals benchmark is too coarse to measure fine detail; if so, the published normal-error gaps probably understate the synthetic-data advantage, and a finer real benchmark is needed to settle it.","If synthetic fidelity continues to rise, the economic case for collecting and pseudo-labelling hundreds of millions of real images for dense prediction weakens; these results put a concrete data point on that trade-off, with 300K images standing in for 62M.","The paper restricts its claim to human-centric relative depth and does not show synthetic-only training for metric depth or general scenes; testing whether the same pipeline survives open-world geometry would define the boundary of the paradigm."],"forward_implications":["A 0.3B-parameter model trained only on synthetic data matches the depth accuracy of a 2.16B-parameter foundation model on Goliath and Hi4D, so parameter scale is not the binding constraint for these tasks.","Training cost drops by orders of magnitude: the paper estimates that the compute used to train one Sapiens-2B model could train roughly 800 of its own models.","Because the training data is procedurally generated, downstream models inherit guarantees on consent, licensing, and data provenance that large real-image collections cannot provide.","Diversity in pose, body shape, and environment is an explicit sampling dial, giving a direct mechanism to probe and correct unfairness across demographic groups.","The same dataset and architecture handle three dense prediction tasks at once, and a single multi-task model with three heads performs close to three task-specific models."],"supporting_citations":[{"why":"supplies the procedural data-generation pipeline that renders the SynthHuman dataset.","marker":"[11]"},{"why":"the face model with eyelid-fold consistency incorporated into SynthHuman.","marker":"[31]"},{"why":"the Sapiens foundation baselines whose depth and normal accuracy the paper matches or beats.","marker":"[17]"},{"why":"Depth Anything v2, the 62M-real-image teacher-student baseline the paper replaces with 300K synthetic images.","marker":"[46]"},{"why":"Depth Pro, the sharp monocular depth baseline trained on mixed real and synthetic data.","marker":"[2]"},{"why":"the DPT encoder-decoder architecture that the paper adapts for all three tasks.","marker":"[35]"},{"why":"the scale-and-shift-invariant depth loss the models are trained with.","marker":"[33]"},{"why":"the Goliath real capture dataset used to benchmark depth and surface normals.","marker":"[26]"},{"why":"the Hi4D close-interaction dataset used to benchmark depth and surface normals.","marker":"[49]"},{"why":"the prior demonstration that synthetic data alone suffices for face-analysis tasks.","marker":"[43]"}],"fun_headline_variants":["Synthetic data beats billion-parameter vision models","300K synthetic images outmatch 2B-parameter models","Strand-level synthetic data slashes vision model size","Perfect labels from synthetic data rival giant models","Small synthetic dataset defeats massive vision models"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the rendered synthetic humans in SynthHuman are realistic enough that a model trained only on them transfers to real photographs; if the synthetic-to-real gap were large, the accuracy gains over data-hungry baselines would not materialize.","fun_headline_variants_meta":{"raw":{"variants":["Synthetic data beats billion-parameter vision models","300K synthetic images outmatch 2B-parameter models","Strand-level synthetic data slashes vision model size","Perfect labels from synthetic data rival giant models","Small synthetic dataset defeats massive vision models"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000207,"raw_usage":{"total_tokens":1397,"prompt_tokens":941,"completion_tokens":456,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":557,"completion_tokens_details":{"reasoning_tokens":384}},"tokens_in":557,"tokens_out":456,"duration_ms":4575,"temperature":1.0,"reasoning_tokens":384,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T15:33:55.114941+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure the depth and normal models on a real test set whose ground truth is captured at finer fidelity than Goliath and Hi4D (for example, light-stage scans that resolve hair strands and wrinkles). If the models' fine-detail advantage over Sapiens-2B does not appear there, then the paper's attribution of accuracy to high-fidelity synthetic ground truth is refuted; if blurring the predictions still helps on that finer benchmark, the benchmark itself is the bottleneck, not the model.","supporting_citations":[{"cited_title":"Cashman, and Jamie Shotton","cited_arxiv_id":null,"evidence_quote":"the prior demonstration that synthetic data alone suffices for face-analysis tasks."},{"cited_title":"Look ma, no markers: holistic perfor- mance capture without the hassle","cited_arxiv_id":null,"evidence_quote":"supplies the procedural data-generation pipeline that renders the SynthHuman dataset."},{"cited_title":"Eyelid fold consistency in facial modeling","cited_arxiv_id":null,"evidence_quote":"the face model with eyelid-fold consistency incorporated into SynthHuman."},{"cited_title":"Sapiens: Foundation for human vision mod- els","cited_arxiv_id":null,"evidence_quote":"the Sapiens foundation baselines whose depth and normal accuracy the paper matches or beats."},{"cited_title":"Depth anything v2","cited_arxiv_id":null,"evidence_quote":"Depth Anything v2, the 62M-real-image teacher-student baseline the paper replaces with 300K synthetic images."},{"cited_title":"Richter, and Vladlen Koltun","cited_arxiv_id":null,"evidence_quote":"Depth Pro, the sharp monocular depth baseline trained on mixed real and synthetic data."},{"cited_title":"Vi- sion transformers for dense prediction","cited_arxiv_id":null,"evidence_quote":"the DPT encoder-decoder architecture that the paper adapts for all three tasks."},{"cited_title":"Towards robust monocular depth estimation: Mixing datasets for zero-shot cross-dataset transfer","cited_arxiv_id":null,"evidence_quote":"the scale-and-shift-invariant depth loss the models are trained with."},{"cited_title":"Jewett, Simon Venshtain, Christopher Heilman, Yueh-Tung Chen, Sidi Fu, Mohamed Ezzeldin A","cited_arxiv_id":null,"evidence_quote":"the Goliath real capture dataset used to benchmark depth and surface normals."},{"cited_title":"Hi4d: 4d instance segmentation of close human interaction","cited_arxiv_id":null,"evidence_quote":"the Hi4D close-interaction dataset used to benchmark depth and surface normals."}],"review_version":1}