{"id":"6e26de7a-be73-4191-b605-a344359c47c8","arxiv_id":"2504.18318","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"STP4D directly denoises 4D Gaussian splatting attributes with DDIM conditioned on time-varying text embeddings, producing high-fidelity 4D assets in about 4.6 seconds.","lead":"STP4D is a new text-to-4D system that uses a diffusion model to directly generate 4D Gaussian scenes in about 4.6 seconds per asset. It adds time-varying prompt embedding, geometric enhancement, and temporal extension modules to improve consistency, and claims to surpass prior text-to-4D methods in speed and quality.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Quality comparison is compromised: STP4D is trained on the same CLIP and FVD objectives used for evaluation, while baselines are not, so the claimed SOTA quality is not established.","rationale":"The reader's weakest assumption identifies exactly the load-bearing concern: the evaluation metrics overlap with the training objectives, making the quality comparison unfair. My independent reading confirms this and finds no additional internal inconsistency that changes the verdict. The proposed architecture is plausible, the speed advantage is concrete and less affected by metric overlap, and the ablations are informative, so acceptance of the core method is not at issue. However, the claim of state-of-the-art quality should remain conditional on a fair, controlled comparison, as the reader concluded. No further verdict adjustment is needed.","tokens_in":15153,"tokens_out":4273,"duration_ms":46408,"concrete_test":"Retrain or fine-tune at least one strong baseline, such as Diffusion4D, with the same L_PC-CLIP and L_TC-FVD losses, or alternatively evaluate all methods using a CLIP variant not used in training and a video feature extractor other than the I3D used in Eq. 14, on the same 20-case test set. If STP4D's margins in Table I shrink below the reported gaps or reverse, the SOTA quality claim is an artifact of training/evaluation overlap.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that STP4D surpasses existing methods in quality rests on Table I, but the quality comparison is asymmetrical. STP4D is explicitly trained with L_PC-CLIP (Eq. 16), which maximizes cosine similarity between CLIP text and image embeddings of rendered frames, and L_TC-FVD (Eq. 14), which minimizes the Frechet distance between I3D features of generated and ground-truth sequences. These are the same quantities later reported as CLIP-F, CLIP-O, and FVD in Table I. The baselines (MAV3D, Animate124, 4DFY, Diffusion4D) were trained under their own objectives, which did not include optimizing these evaluation metrics, so the comparison gives STP4D an advantage that is not disclosed. The paper also does not report re-running the baselines under the same prompts, rendering pipeline, and 20-case test split; no per-case results or error bars are provided, so the reported margins (0.031 CLIP-F, 0.055 CLIP-O, 41.5 FVD) cannot be assessed for significance. The speed claim is less affected by this issue, but the headline quality conclusion is not supported by the current evidence.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes STP4D, a text-to-4D generation method based on Gaussian splatting. The method initializes Gaussian attributes from noise and uses a DDIM denoiser conditioned on CLIP text embeddings, with a Time-varying Prompt Embedding (TPE) module, a Geometric Information Enhancement (GIE) module built on K-Planes and a GroupFormer, and a Temporal Extension Deformation (TED) module that extends anchor frames to full-length video via cross-attention. Training employs five losses: SSIM-based spatial loss, local rigidity, FVD-based temporal loss, Savitzky-Golay smoothness, and CLIP similarity. Experiments on the Diffusion4D dataset report CLIP-F 0.841, CLIP-O 0.705, FVD 441.1, and an inference time of about 4.6 seconds per asset, which is claimed to surpass existing methods in both quality and speed. Ablation studies are provided for each module and each proposed loss.","tokens_in":15415,"tokens_out":4135,"duration_ms":45991,"significance":"If the reported results are validated, the main contribution is a feed-forward text-to-4D pipeline that avoids per-asset test-time optimization, together with a coherent module design for spatio-temporal-prompt consistency. The speed advantage over optimization-based baselines is plausible and potentially practically important. The paper is also honest about its limitations, noting in Section VIII that training data and model capacity restrict complex-scene quality. However, the quantitative quality comparison is not currently established because the evaluation metrics coincide with the training objectives, baseline numbers are quoted from the literature without re-running, and no error bars or significance tests are reported. The central quality claim therefore needs additional evidence before the state-of-the-art conclusion can be accepted.","major_comments":[{"comment":"The evaluation is asymmetric: STP4D is explicitly trained with L_TC-FVD (Eq. 14) and L_PC-CLIP (Eq. 16), which are the same quantities reported as FVD and CLIP-F/CLIP-O in Table I, while the baseline methods were not trained on these objectives. This gives STP4D a direct and undisclosed advantage when comparing against quoted baseline numbers. The reported margins are also small (0.031 in CLIP-F, 0.055 in CLIP-O, 41.5 in FVD) and are computed on only 20 test cases without error bars or significance testing. I therefore do not regard the claim that STP4D 'surpasses existing methods' in quality as supported. The authors should evaluate with metrics that are not part of the training objective, or re-run baselines under identical conditions, and report per-case results with confidence intervals or paired significance tests.","section":"Section III-E, Section IV-A, Eqs. (14) and (16), Table I"},{"comment":"The baseline numbers in Table I appear to be taken from prior publications rather than from re-running MAV3D, Animate124, 4DFY, and Diffusion4D on the same prompts, camera trajectories, rendering pipeline, and the same 20-case test split used for STP4D. Since CLIP-F, CLIP-O, and FVD all depend on the specific rendered sequences, cross-paper comparison of these numbers is unreliable. The authors should either run all methods under identical conditions or substantially weaken the quantitative comparison and explicitly state which numbers are quoted from the literature. Without this, the quantitative superiority claim is not established.","section":"Table I and Section IV-B"},{"comment":"The user study is reported only as percentages in Table I, with no description of the experimental protocol: whether participants saw paired or unpaired videos, whether ratings were forced-choice or Likert-scale, how prompts and assets were sampled, and whether any significance test was applied. The percentages appear to sum to 100 per column, suggesting forced choice, but this is not stated. Given that the user study is a central piece of evidence for the subjective quality claim, the protocol and statistical significance must be reported.","section":"Table I, User Study"},{"comment":"The ablation studies also use CLIP-F, CLIP-O, and FVD as the evaluation metrics, which are the same quantities minimized by L_PC-CLIP and L_TC-FVD during training. This confounds the interpretation of the module and loss ablations: removing a module or loss may change the training dynamics in ways that affect the reported metrics, but the direction of the effect is biased by the fact that the metric itself is an objective. I recommend reporting at least one additional independent metric (for example, LPIPS, user study, or a no-reference video quality metric) for the ablations.","section":"Tables II and III, Section IV-C"}],"minor_comments":[{"comment":"The statement that STP4D operates 'without the necessity of a prior guidance model' is slightly misleading because the method still uses a pretrained CLIP text encoder to obtain the conditioning signal; I suggest clarifying the distinction between a pretrained text encoder and the score-distillation guidance models used by prior work.","section":"Section II"},{"comment":"The preliminary on Gaussian splatting in Eq. (7) is not written in standard form; the exponent should involve (x - x_i)^T Sigma_i^{-1} (x - x_i) rather than a quadratic form without the data point. This is a presentation issue in the supplementary material but should be corrected for readability.","section":"Section VI.A, Eq. (9)"},{"comment":"FVD is a distribution-level metric defined between two sets of videos, while Eq. (14) applies it to a single generated video and a single ground-truth video. The authors should explain how the I3D features and Frechet distance are computed in this per-sample setting, for example whether a batch of frames is used and how the statistics are estimated.","section":"Section VI.C, Eq. (14)"},{"comment":"The hyperparameter study on the temporal extension ratio reports CLIP-F and FVD but omits CLIP-O; adding CLIP-O would make the table consistent with the other comparisons.","section":"Table IV"},{"comment":"The paper does not mention code or model release, which would be useful for reproducing the results and for the community to check the fairness of the baseline comparison.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The architectural contribution is interesting and the speed claim is credible, so I do not recommend rejection. The main issue is that the headline quality comparison is not yet proven due to the metric/training-objective overlap and the lack of re-run baselines. This can be fixed with additional experiments, so major revision seems appropriate. I would also encourage the editor to ask the authors to clarify provenance of all baseline numbers in Table I."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things worth knowing about STP4D. First, the core new idea is real: it treats 4D Gaussian generation as a direct DDIM denoising problem, with per-frame CLIP embeddings injected through cross-attention, and produces a 24-frame asset in 4.6 seconds. That is a genuine speed advance over SDS-based optimization, which takes minutes to hours. Second, the headline quality comparison is not established the way Table I suggests. The training losses include L_PC-CLIP and L_TC-FVD, which are the same quantities reported as CLIP-F/CLIP-O and FVD, while the baselines were not trained on those objectives. The paper never acknowledges that asymmetry.\n\nWhat is genuinely good: the pipeline is thoughtfully composed from known components (K-Planes, cross-attention, GroupFormer), and the ablations show each module contributes. The temporal extension ratio study and Gaussian group analysis are useful. The writing is clear, the related work is properly positioned, and the supplementary gives enough implementation detail to reproduce the architecture. The speed claim is credible and is the strongest contribution.\n\nWhere it is soft: the evaluation is the main weakness. Beyond the metric overlap, the 20-case test set has no error bars, and baseline numbers are quoted from the literature rather than re-run under the same prompts and rendering pipeline. The user study helps, but with 40 participants it cannot carry the quantitative claim. No code or models are released, so there is no external check. A minor presentation issue: Table I's 'Human Performance' row seems to actually be the user-study percentage for each method, which should be labeled more carefully.\n\nThe discussion section honestly acknowledges that the method struggles on complex scenes because of limited training data and the 40k Gaussian budget—that is a fair limitation statement and does not undercut the speed result.\n\nBottom line: this paper is for people working on fast text-to-4D generation, and it deserves a serious referee. But the quality comparison needs to be redone or reframed before the SOTA claim can stand. I would send it to peer review with a clear request: re-run the baselines, report per-case results and error bars, and either drop the overlapping metrics from the evaluation or train a version without those losses and show the margins survive.","headline":"STP4D is a genuinely fast diffusion-based 4D generator, but its quality lead over baselines is inflated because it trains on the same CLIP and FVD metrics it reports.","tokens_in":15941,"tokens_out":2554,"would_cite":true,"duration_ms":25036,"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":"The paper claims a single end-to-end diffusion model can generate 4D Gaussian Splatting content from text in 4.6 seconds while beating prior methods on consistency metrics.","keywords":["text-to-4D generation","Gaussian Splatting","Denoising Diffusion Implicit Models","time-varying prompt embedding","spatio-temporal consistency","temporal extension","local rigidity","4D generation"],"falsifier":"Re-train the strongest baseline on the same Diffusion4D split with the STP4D loss terms $L_{\\mathrm{PC-CLIP}}$ and $L_{\\mathrm{TC-FVD}}$ added to its own objective, then compare CLIP-F, CLIP-O, and FVD. If its scores reach or exceed 0.841, 0.705, and 441.1, the claimed quality advantage is an artifact of uneven objectives rather than the architecture.","tokens_in":14974,"feed_emoji":"🎬","tokens_out":7280,"duration_ms":61425,"temperature":0.7,"pith_summary":"STP4D is a text-to-4D generation method that uses a Denoising Diffusion Implicit Model (DDIM) to directly denoise the attributes of 4D Gaussian splat primitives from text, rather than optimizing a 3D or 4D representation through score distillation from separate image or video diffusion models. The paper argues that existing text-to-4D pipelines neglect combined spatial, temporal, and prompt consistency. Three modules - Time-varying Prompt Embedding, Geometric Information Enhancement, and Temporal Extension Deformation - are designed to enforce these consistencies together, with auxiliary losses for spatial rigidity, smooth motion, and CLIP-based prompt alignment. The result is a claimed state-of-the-art quality on the Diffusion4D benchmark with an inference time of roughly 4.6 seconds per 24-frame asset.","feed_headline":"Diffusion model draws a 4D scene from text in 4.6 seconds","feed_subtitle":"STP4D denoises Gaussian splats directly from text, adding prompt, spatial, and temporal consistency in one pass.","key_machinery":"The load-bearing object is the 4D Gaussian splat itself, treated as a denoising target: 40,000 anisotropic Gaussians with attributes (position, rotation, scale, color, opacity) are sampled from standard Gaussian noise and iteratively denoised by a 50-step DDIM. Three mechanisms act on this object. Time-varying Prompt Embedding maps a CLIP text embedding into per-frame features injected by cross-attention; Geometric Information Enhancement uses K-Planes to factorize the Gaussian tensor into three low-rank planes and applies GroupFormer (window multi-head self-attention plus sparse convolution) to extract inter-group and intra-group features; Temporal Extension Deformation uses a learnable weight pool as the query and anchor-frame Gaussians as key and value to produce all actual frames. The losses $L_{\\mathrm{SC-Rig}}$, $L_{\\mathrm{TC-FVD}}$, $L_{\\mathrm{TC-Smooth}}$, $L_{\\mathrm{SC-SSIM}}$, and $L_{\\mathrm{PC-CLIP}}$ jointly steer the representation.","core_discovery":"The central claim is that the full spatio-temporal-prompt consistency needed for high-fidelity text-to-4D generation can be captured in a single end-to-end diffusion model over Gaussian splat parameters. STP4D treats 40,000 Gaussians across 12 anchor frames as tokens, embeds per-frame text features through cross-attention at each denoising step, enhances geometry via low-rank K-Planes decomposition and a GroupFormer attention block, and extends the anchor frames to 24 output frames with a learned weight pool in a further cross-attention stage. The paper reports that this setup produces higher CLIP-F and CLIP-O scores and a lower FVD score than prior text-to-4D methods, and that per-asset inference is roughly two orders of magnitude faster than the Diffusion4D baseline.","pith_inferences":["Editorial inference: the reported margins likely overstate the architectural advantage, since STP4D is trained with the exact CLIP and FVD objectives used to score all methods; re-training baselines with the same losses would quantify the true gain.","Editorial inference: the Temporal Extension Deformation module is a generic anchor-to-full-frame cross-attention mechanism that could be reused for temporal interpolation or frame-rate upsampling of any 4D Gaussian scene, independent of the diffusion front end.","Editorial inference: the paper's own discussion predicts that scaling the Gaussian budget beyond 40,000 and training on richer multi-object datasets, without architecture changes, should extend the same consistency benefits to complex scenes; this is directly testable.","Editorial inference: because DDIM sampling is deterministic given the initial noise, diversity across generated assets will come from the noise prior and the learned sampler; controlling or editing the noise seed is an unexplored handle for content variation."],"forward_implications":["If the central claim is right, text-to-4D generation no longer needs a separate score-distillation stage; the diffusion model itself is the generator, so inference drops from hours to seconds.","Removing Geometric Information Enhancement hurts all metrics most in the ablation, which suggests local and global geometric modeling, not just diffusion conditioning, is what keeps dynamic scenes coherent.","Removing Time-varying Prompt Embedding reduces CLIP-F and CLIP-O, supporting the claim that frame-specific prompt features improve text alignment beyond a single global text embedding.","The 2:1 anchor-to-actual temporal extension ratio is reported to generalize from 24-frame to 50-frame generation, implying the efficiency gain does not come at a quality cost at that ratio.","Because the whole pipeline is end-to-end and runs in 4.6 seconds per asset, it is a direct candidate for interactive or real-time 4D content creation."],"supporting_citations":[{"why":"Supplies the Diffusion4D dataset, the 20-case test split, and the CLIP-F, CLIP-O, and FVD protocol used for all quantitative comparisons, and is the strongest baseline.","marker":"[4]"},{"why":"Provides the 4D-FY hybrid score distillation baseline whose outputs are compared qualitatively and quantitatively.","marker":"[7]"},{"why":"Supplies the deterministic DDIM formulation that STP4D adapts to denoise Gaussian attributes.","marker":"[13]"},{"why":"Provides the 4D Gaussian Splatting representation that lets the generated Gaussians be rendered in real time.","marker":"[14]"},{"why":"Provides the CLIP text encoder used in Time-varying Prompt Embedding and in the prompt-consistency loss and metrics.","marker":"[23]"},{"why":"Provides the K-Planes low-rank decomposition used in Geometric Information Enhancement.","marker":"[24]"},{"why":"Supplies the Fréchet Video Distance used both as a temporal-consistency training loss and as the FVD evaluation metric.","marker":"[29]"},{"why":"Provides the 3D Gaussian Splatting rendering equation that turns the generated Gaussians into images.","marker":"[32]"}],"fun_headline_variants":["Text to 4D scene in 4.6 seconds via diffusion","STP4D: 4D Gaussian splats from text, faster than ever","4D from text: one diffusion pass, 4.6 seconds","Diffusion model crafts 4D scenes from text in 4.6s","STP4D: Text-driven 4D Gaussians with spatio-temporal prompts"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The state-of-the-art quality claim rests on comparing methods under evaluation metrics that the proposed model is explicitly trained to optimize, while the baselines were not trained on those same objectives.","fun_headline_variants_meta":{"raw":{"variants":["Text to 4D scene in 4.6 seconds via diffusion","STP4D: 4D Gaussian splats from text, faster than ever","4D from text: one diffusion pass, 4.6 seconds","Diffusion model crafts 4D scenes from text in 4.6s","STP4D: Text-driven 4D Gaussians with spatio-temporal prompts"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000471,"raw_usage":{"total_tokens":2329,"prompt_tokens":916,"completion_tokens":1413,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":532,"completion_tokens_details":{"reasoning_tokens":1309}},"tokens_in":532,"tokens_out":1413,"duration_ms":10323,"temperature":1.0,"reasoning_tokens":1309,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T10:19:23.599965+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-train the strongest baseline on the same Diffusion4D split with the STP4D loss terms $L_{\\mathrm{PC-CLIP}}$ and $L_{\\mathrm{TC-FVD}}$ added to its own objective, then compare CLIP-F, CLIP-O, and FVD. If its scores reach or exceed 0.841, 0.705, and 441.1, the claimed quality advantage is an artifact of uneven objectives rather than the architecture.","supporting_citations":[{"cited_title":"4d-fy: Text-to-4d gener- ation using hybrid score distillation sampling,","cited_arxiv_id":null,"evidence_quote":"Provides the 4D-FY hybrid score distillation baseline whose outputs are compared qualitatively and quantitatively."},{"cited_title":"4d gaussian splatting for real-time dynamic scene rendering,","cited_arxiv_id":null,"evidence_quote":"Provides the 4D Gaussian Splatting representation that lets the generated Gaussians be rendered in real time."},{"cited_title":"K-planes: Explicit radiance fields in space, time, and appearance,","cited_arxiv_id":null,"evidence_quote":"Provides the K-Planes low-rank decomposition used in Geometric Information Enhancement."}],"review_version":1}