{"id":"7e845d5e-cbad-4d31-806e-f3b7b98f856b","arxiv_id":"2608.12314","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"StateFlow constructs, evolves, and accesses a persistent 3D world state for previsualization, reporting higher consistency and controllability than one-shot video generation on VBench and user studies.","lead":"StateFlow turns a text prompt into an editable 3D world, keeps it as a persistent state, and lets creators evolve it and plan camera shots for video or game-like output. The system matters for filmmakers and game designers who need rapid iteration on scenes, motion, and camera work before final production.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'evolving' pillar of the central claim is asserted but never quantitatively tested; editability evidence is one sentence in Sec 4.5.","rationale":"The reader's weakest_assumption is the single-VLM dependency and lack of 3D verification. I share that concern, but I think the more immediately load-bearing gap is the absence of any quantitative test of the evolution stage, which is one of the three pillars in the title and contributions. Even a perfect State Construction and State Access pipeline would not support the central claim if State Evolution is just VLM-planned table edits with no measured benefit. This is an evidence gap rather than a demonstrated error, so it does not overturn the paper; it strengthens the case for the reader's CONDITIONAL verdict. I credit the paper for concrete quantitative comparisons on construction and access (VBench, CLIP-I/T, user/MLLM study) and for ablating BEV layout and conflict resolution; if those were the whole claim, the evidence would be materially stronger. The missing evolution evaluation, however, is exactly what the conditional should require before the advertised 'Building, Evolving, and Accessing' claim is accepted as supported. I therefore leave the reader's verdict unchanged rather than moving it, and propose one benchmark that would settle the question.","tokens_in":16723,"tokens_out":5776,"duration_ms":55204,"concrete_test":"Run a multi-turn edit benchmark: take 20-30 prompts, apply 3 sequential StateFlow edits per scene covering scene expansion, object pose/motion change, and event-level substitution (explosion/destruction). Compare against a regeneration baseline that rebuilds the world from the updated prompt, and against one-shot video generation with the prompt updated after each edit. For each turn, measure (a) identity preservation of untouched objects via CLIP feature similarity or re-ID over rendered frames, (b) spatial stability via 3D-box IoU for untouched objects, (c) edit success via human raters, and (d) per-turn time. Report means with confidence intervals and significance tests. If StateFlow does not beat regeneration on identity/layout stability or edit success, the evolution pillar of the central claim fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"StateFlow's central claim is that persistent object-centric 3D world states support construction, evolution, and access, and that structured state transitions outperform one-shot regeneration on coherence, controllability, and editability. The construction and access stages are evaluated (Tables 1-3), but the evolution stage (Sec 3.4, Eq. 6) has no quantitative evaluation anywhere. Sec 4.5's State Evolution ablation is described in a single sentence: 'replacing selective state updates with direct regeneration ... weakens identity preservation and spatial consistency across edits,' with no metric, table row, or figure. Table 3 contains no editability/evolution dimension, and the user study evaluates single-turn outputs. Since 'evolving' is one of the paper's three named contributions, the editability half of the central claim rests on an untested assertion. There is also a mechanism-level risk: event substitution replaces g_i while preserving p_i, and scene-style edits rewrite descriptors without any verification that the rendered geometry matches the updated semantic state, so 'persistent' may mean a stale box table rather than a consistent world. A code release alone would not fix this gap; the experiments themselves need to measure multi-turn edits.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes StateFlow, a state-centric framework for generative previsualization. Instead of one-shot image or video synthesis, StateFlow maintains a persistent object-centric 3D world state W_t = {o_i^t = (g_i^t, p_i^t, s_i^t)} with three stages: construction (prior-guided conflict-aware dual-view initialization), evolution (intent-guided structured state transitions), and access (world-state camera planning with render-feedback reflection). The framework is instantiated with off-the-shelf components: Gemini 3.1 for VLM reasoning, Nano Banana 2 for image generation, Hunyuan3D for lifting crops to 3D assets, and Seedance2 for video enhancement. Quantitative comparisons are reported against video-generation baselines (VBench, Table 1), scene-generation baselines (CLIP-I/T, HPS, Q-Align, Table 2), and a combined user study/MLLM evaluation (Table 3), along with ablations of construction and camera-planning components.","tokens_in":16961,"tokens_out":4830,"duration_ms":44873,"significance":"If the central claim is fully established, StateFlow is a meaningful step toward controllable previsualization: the formulation cleanly separates world state, evolution, and access; the construction and camera-planning stages are training-free and model-agnostic; and the paper includes a human study alongside automatic metrics. The manuscript is honest about its dependence on third-party models and explicitly frames the world state as an editable representation. However, the evidence is incomplete: the 'evolving' pillar of the claim has no quantitative evaluation, the main automatic evaluator is the same model family that performs planning, and the video-level comparison may conflate StateFlow's contribution with the Seedance enhancement backbone. Because these gaps bear directly on two of the three named contributions, the current evidence is not sufficient for acceptance, though the deficiencies appear fixable within the manuscript's scope.","major_comments":[{"comment":"The State Evolution ablation is asserted but not measured. Section 4.5 states that replacing selective state updates with direct regeneration 'weakens identity preservation and spatial consistency across edits,' and that results are reported alongside baselines in Table 3, but no row or column in Table 3 corresponds to this ablation, no metric is given, and no figure is referenced. Since 'evolving' is one of the paper's three core contributions, the claim that structured state transitions outperform full-scene regeneration for iterative editing is currently unsupported. The authors should add quantitative experiments on multi-turn edits (e.g., identity preservation, spatial consistency, layout plausibility before and after edits, with and without selective updates), plus a user study on edit sequences, not only single-turn outputs.","section":"§4.5 and §3.4 (Eq. 6)"},{"comment":"The video-level comparison is confounded by the Seedance2 enhancement. Implementation details state that Seedance2 is used as the video generation backbone for downstream outputs, while Table 1 lists Seedance2.0 as a baseline. It is therefore unclear whether the 'Ours' videos evaluated in Table 1, Table 3 (video-level), and Fig. 4 are pure geometry-grounded renders or Seedance2-enhanced videos. If the latter, the VBench scores for 'Ours' partly reflect the enhancement model's capabilities rather than StateFlow's state representation, making the comparison with the Seedance2.0 baseline unfair. The paper should report the rendered-only and enhanced results separately, or apply the same enhancement to all baselines, and state explicitly which output is used in each evaluation.","section":"§4.1, §4.3, Table 1"},{"comment":"The automatic MLLM evaluation uses Gemini 3.1, the same model family that performs scene planning, state transition prediction, and camera reflection. This creates a self-preference risk: the evaluator may systematically favor layouts and trajectories it proposed. The human study partially offsets this, but the paper reports both scores side by side as if they were equally independent, and does not report human-MLLM agreement or any analysis of divergence. The authors should either use an independent evaluator, report the MLLM scores as a secondary diagnostic with an explicit caveat, or provide evidence (e.g., per-prompt correlation) that the MLLM scores align with human judgments.","section":"§4.4 and §4.1"},{"comment":"The construction and camera-planning objectives are not specified in enough detail to be reproduced or audited. Equation (4) contains four loss terms with weights lambda_b, lambda_v, lambda_p that are never given, and the text does not describe the optimization algorithm, the box parameterization, the number of iterations, or how the front-view identity term is computed. Similarly, the trajectory selection score J in Eq. (9) is described only qualitatively. Given that these objectives are central to the method's claimed generality, the authors should provide the full objective definitions, weights, solver details, and, ideally, code or a detailed pseudo-code appendix.","section":"§3.3 (Eq. 4) and §3.5 (Eq. 9)"},{"comment":"No significance tests, confidence intervals, or per-prompt variance are reported for any quantitative comparison. In the user study, N=30 participants rate 12 prompts per setting, but only mean scores are shown, so it is impossible to tell whether differences between StateFlow and the strongest baselines (e.g., SynCity in Table 2, Seedance2.0 in Table 3) are meaningful relative to inter-participant or inter-prompt variance. The authors should add standard deviations or CIs and, where feasible, paired tests; otherwise the claimed overall superiority is not statistically grounded.","section":"Tables 2, 3"}],"minor_comments":[{"comment":"The state table in Fig. 2 contains the typo 'Solar Pbanel' in the rightmost table; it should read 'Solar Panel'.","section":"Fig. 2"},{"comment":"The user study description says attention-check items were inserted but does not report how many participants failed them or whether their responses were excluded; this information is needed to judge data quality.","section":"§4.4"},{"comment":"The VBench evaluation does not state the video resolution, duration, or frame count used for the comparison; since the paper compares methods that may generate different aspect ratios and lengths, these details should be reported for fairness.","section":"§4.3, Table 1"},{"comment":"PartCrafter is described as a part-aware 3D object generation method, while the comparison claims to be on scene generation; the task mismatch should be acknowledged or the baseline matched to the scene-generation setting.","section":"§4.2, Table 2"},{"comment":"The evaluation prompts used for the MLLM judge are not given; to let readers gauge possible bias and to support replication, the full evaluation prompt should be included in the appendix.","section":"§4.4"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the scope of the venue, and the formulation is timely. The main obstacle is that the 'evolution' contribution is essentially unvalidated, and the evaluation has a same-model circularity and an enhancement confound. These issues are addressable with additional experiments and transparency, so I recommend major revision rather than rejection. I would also ask the editors to encourage the authors to release code/data or at least detailed prompts and evaluation scripts, since the method depends on several proprietary models and would otherwise be difficult to reproduce."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"StateFlow is worth a serious look. The core idea—previs should be an explicit persistent 3D world state W_t = {o_i = (g_i, p_i, s_i)} rather than one-shot video—is a real reframing, and the three-stage decomposition (construct, evolve, access) is clean and useful. The dual-view initialization with VLM conflict resolution is clever, and the render-feedback camera loop is a nice two-system design: let the VLM propose, let rendering verify. The system is assembled from off-the-shelf parts, so it is a systems contribution, but a well-executed one. The qualitative results look meaningfully better on spatial and identity consistency than the video baselines.\n\nThe soft spots are real but not fatal. Most important: the 'evolving' half of the central claim is essentially untested. Section 4.5's evolution ablation is one sentence with no metric, no table row, no figure. The user study and MLLM eval cover single-turn outputs. There is no editability or evolution dimension in Table 3. Given that 'evolving' is one of the paper's three named contributions, this is a load-bearing gap. Code release would not fix it; the experiments need to measure multi-turn edits—identity preservation, spatial consistency, time-to-edit, etc.\n\nSecond, evaluation has circularity risk: Gemini 3.1 does the planning and also serves as the MLLM evaluator. The human study offsets this, but it is N=30, 12 prompts, no significance tests, and per-dimension means only. That is thin but acceptable for a systems paper if code and data accompany it. They do not. No code, no data, and several hyperparameters (Eq. 4 weights, camera search size M, max rounds) are unreported.\n\nThird, the final video uses Seedance2 enhancement, so Table 1's VBench results are partly for a post-processed output, not pure StateFlow. That should be disclosed more prominently. The limitation section only mentions inference speed, not the VLM-semantics reliance or the absence of 3D ground truth. That is a miss.\n\nOn citation pattern: related work is broad and fair, and self-citations are in the normal range. No red flag.\n\nBottom line: this is a serious, well-thought-out system paper that advances a genuine research direction. It deserves peer review, but a referee should require (1) quantitative evolution/editability evaluation, ideally a multi-turn editing protocol; (2) code/data release; (3) significance tests or at least error bars; (4) clearer separation between the planner VLM and the evaluator MLLM, or else a justification that the VLM-family overlap does not drive the gap. I would accept it conditionally.","headline":"A genuinely useful state-centric reframing for previs, but the 'evolve' pillar is under-tested and the evaluation needs cleaner separation from the VLM it uses.","tokens_in":17519,"tokens_out":1739,"would_cite":true,"duration_ms":13915,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["68T45"],"pacs":[],"model":"deepseek-v4-flash","headline":"StateFlow claims that previsualization is best modeled as a persistent 3D world state, not as one-shot image or video generation.","keywords":["previsualization","3D world state","object-centric representation","dual-view initialization","camera planning","render-feedback reflection","intent-guided state transition","game prototyping"],"falsifier":"Take a prompt in which the bird's-eye view implies an object count that the front view contradicts, for example text says 'three towers', the front view shows two, and the BEV shows three. If the vision-language model's conflict resolution keeps a hallucinated tower or drops one clearly visible in both views, and the lightweight box-optimization objective does not correct it, then the claim that StateFlow reconciles cross-view conflicts and builds physically plausible worlds is falsified. A second check is to render the constructed world from an unseen azimuth and look for floating, interpenetrating, or boundary-violating objects; any such artifact that persists after the physical-refinement term demonstrates that VLM semantic priors alone cannot guarantee physical plausibility.","tokens_in":16537,"feed_emoji":"🎬","tokens_out":6572,"duration_ms":55170,"temperature":0.7,"pith_summary":"The paper argues that previsualization, the iterative planning of scenes, cameras, and motion before final production, is poorly served by one-shot generators that entangle scene content, layout, motion, and camera in a single prompt and expose no persistent state to edit. Its proposal, StateFlow, maintains an explicit object-centric 3D world state, written as a set of object entities each with geometry, placement, and semantic attributes, and updates that state through three stages: construction from dual-view references with conflict resolution, evolution through structured transitions that preserve world memory, and access through camera trajectories refined by rendered feedback. The point of the state is reuse: different frames are local modifications or recombinations of shared geometry, pose, and semantics, so creators can iterate without regenerating the whole world. The paper reports that this state-centric approach outperforms video-generation and scene-generation baselines on consistency, controllability, and previsualization usefulness in both automatic metrics and user studies.","feed_headline":"Persistent 3D world state beats one-shot video for previs","feed_subtitle":"Editable 3D state keeps objects, layout, and cameras consistent across edits and viewpoints.","key_machinery":"The load-bearing object is the structured 3D world state, written as a set of object entities, each a triple of geometry, placement, and semantic attributes, together with three operations on it. State construction uses prior-guided conflict-aware dual-view initialization: a front view supplies appearance and asset crops, a bird's-eye view supplies ground-plane layout, and a vision-language model resolves count and placement conflicts before an inference-time objective refines the 3D boxes. State evolution applies intent-guided structured state transition: a vision-language model reads the state table and emits a compact transition plan, which is applied selectively over geometry, placement, or semantics. State access performs world-state camera planning with render-feedback reflection: the vision-language model proposes a trajectory, rendering exposes visibility, framing, and collision failures, and local repairs are scored and selected. The state itself is what carries reuse, because geometry, placement, and semantics are separate fields that can be updated independently.","core_discovery":"The central claim, stated on the paper's own terms, is that previsualization should be formulated as persistent 3D world-state modeling rather than one-shot visual synthesis. Concretely, a world is represented at time t as a set of object entities, each with geometry for rendering and physical reasoning, placement for spatial layout, and semantic attributes for editable state. Construction initializes this world from user conditions by combining a front view that supplies appearance and asset crops with a bird's-eye view that supplies ground-plane layout, using a vision-language model to resolve cross-view conflicts before a lightweight optimization refines 3D boxes. Evolution translates user intent into compact transitions over the state table, so edits are selective and memory-aware rather than full regenerations. Access plans camera trajectories that are proposed semantically and then repaired through rendered feedback. Against video-generation baselines and scene-generation baselines, StateFlow reports the best average VBench score and the best user-study and MLLM ratings for spatial and identity consistency, layout plausibility, and previsualization usefulness, with ablations showing the contribution of BEV grounding, conflict resolution, and render-feedback reflection.","pith_inferences":["Beyond the paper: StateFlow's transition plan could be driven by a learned dynamics model or physics simulator rather than a vision-language model, making state evolution testable against physical ground truth.","Beyond the paper: the object-centric state table could serve as an interchange format between generative models and game engines, since it already exposes geometry, pose, and semantic attributes as editable records.","Beyond the paper: a natural stress test is edit locality, measuring whether a one-line intent change alters only the intended object records and leaves unrelated scene elements bit-identical, which the paper reports qualitatively but does not quantify.","Beyond the paper: the dual-view construction could be extended to additional views to see whether conflict resolution scales beyond front-and-bird's-eye views or whether error rates grow with more views."],"forward_implications":["Edits become local: expanding a scene, changing a style, or substituting a destroyed asset modifies only the affected object records, so the rest of the world keeps its identity and layout.","Camera work is verifiable before production: proposed trajectories are rendered and repaired against occlusion, framing, and collision, so director intent is checked against geometry rather than guessed.","The same world state serves multiple outputs, including video creation, shot planning, keyframes, and game-like interaction, without regenerating the scene for each use.","Video fidelity can be layered on: an off-the-shelf video model renders the geometry-grounded observations, so visual quality improves without altering the underlying world state.","Object identity persists across time and viewpoints, reducing the subject drift and background inconsistency typical of one-shot video synthesis."],"supporting_citations":[{"why":"Supplies the vision-language model that grounds objects, resolves cross-view conflicts, plans transitions, and reflects on camera failures.","marker":"[9]"},{"why":"Generates the front-view and bird's-eye-view reference images that StateFlow reconciles during construction.","marker":"[10]"},{"why":"Lifts object crops from the front view into 3D geometry for the state's geometry term.","marker":"[42]"},{"why":"Serves as both the off-the-shelf video model that enhances rendered observations and a video-generation baseline.","marker":"[33]"},{"why":"The training-free 3D-world baseline used for scene-generation comparison.","marker":"[6]"},{"why":"The benchmark on which StateFlow reports its video-consistency and flicker scores.","marker":"[16]"},{"why":"The prior conversational camera-control method that lacked geometric feedback and that render-feedback reflection is designed to surpass.","marker":"[25]"},{"why":"Used to measure image-text alignment of generated scenes in the scene-generation comparison.","marker":"[28]"}],"fun_headline_variants":["Persistent 3D world state beats one-shot video for previs","Editable 3D state keeps previs consistent across edits and views","Previs as persistent 3D state, not one-shot synthesis","StateFlow: 3D world state for controllable previsualization","From one-shot video to editable 3D world states for previs"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Everything hinges on a single vision-language model: the same model decides which objects are real versus hallucinated, where they sit in 3D, how they change over time, and which camera fixes are needed, with no learned geometric verification or 3D ground truth to check its physical predictions.","fun_headline_variants_meta":{"raw":{"variants":["Persistent 3D world state beats one-shot video for previs","Editable 3D state keeps previs consistent across edits and views","Previs as persistent 3D state, not one-shot synthesis","StateFlow: 3D world state for controllable previsualization","From one-shot video to editable 3D world states for previs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000611,"raw_usage":{"total_tokens":2915,"prompt_tokens":1087,"completion_tokens":1828,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":703,"completion_tokens_details":{"reasoning_tokens":1734}},"tokens_in":703,"tokens_out":1828,"duration_ms":10559,"temperature":1.0,"reasoning_tokens":1734,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T00:08:16.348350+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a prompt in which the bird's-eye view implies an object count that the front view contradicts, for example text says 'three towers', the front view shows two, and the BEV shows three. If the vision-language model's conflict resolution keeps a hallucinated tower or drops one clearly visible in both views, and the lightweight box-optimization objective does not correct it, then the claim that StateFlow reconciles cross-view conflicts and builds physically plausible worlds is falsified. A second check is to render the constructed world from an unseen azimuth and look for floating, interpenetrating, or boundary-violating objects; any such artifact that persists after the physical-refinement term demonstrates that VLM semantic priors alone cannot guarantee physical plausibility.","supporting_citations":[{"cited_title":"Syncity: Training-free gen- eration of 3d worlds","cited_arxiv_id":null,"evidence_quote":"The training-free 3D-world baseline used for scene-generation comparison."},{"cited_title":"Gemini 3.1 Pro Preview.https://ai.google","cited_arxiv_id":null,"evidence_quote":"Supplies the vision-language model that grounds objects, resolves cross-view conflicts, plans transitions, and reflects on camera failures."},{"cited_title":"Nano Banana 2: Google’s latest AI image genera- tion model.https://blog.google/innovation- and-ai/technology/ai/nano-banana-2/, 2026","cited_arxiv_id":null,"evidence_quote":"Generates the front-view and bird's-eye-view reference images that StateFlow reconciles during construction."},{"cited_title":"Hunyuan3d 2.5: Towards high- fidelity 3d assets generation with ultimate details, 2025","cited_arxiv_id":null,"evidence_quote":"Lifts object crops from the front view into 3D geometry for the state's geometry term."},{"cited_title":"Vbench: Comprehensive bench- mark suite for video generative models","cited_arxiv_id":null,"evidence_quote":"The benchmark on which StateFlow reports its video-consistency and flicker scores."},{"cited_title":"Chat- cam: Empowering camera control through conversational ai.Advances in Neural Information Processing Systems, 37: 54483–54506, 2024","cited_arxiv_id":null,"evidence_quote":"The prior conversational camera-control method that lacked geometric feedback and that render-feedback reflection is designed to surpass."}],"review_version":1}