{"id":"286f037c-7334-4010-9f8d-d5b103ead978","arxiv_id":"2504.19894","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"CineVerse combines LLM-based cinematic shot planning with a LoRA-fine-tuned FLUX model so a single scene description yields consistent, shot-labeled keyframes.","lead":"CineVerse reads a one-sentence movie scene description, has a large language model write a shot-by-shot plan, and fine-tunes FLUX with LoRA to draw consistent keyframes for each shot. The result is an honest, incremental improvement over the IC-LoRA baseline, carried by a new dataset and practical fixes that make frame counts and character consistency more reliable.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table 4 frame-count comparison is asymmetric: CineVerse boundaries are found by its own checkerboard cue, IC-LoRA boundaries by an unvalidated row-difference heuristic; the 95.45% vs 34.84% gap may be inflated by the metric.","rationale":"The central claim is that CineVerse is practical because it matches the plan in number, shot size, and content while keeping characters and setting consistent. The frame-count match is the only one of these that is quantified with a head-to-head automatic table, and Table 4 is the evidence for it. The asymmetry identified by the reader is a genuine measurement confound rather than a disagreement over metric choice: the two methods are counted by different detectors, one of which is designed around the method's own output format and the other of which is an unvalidated heuristic. This can inflate the reported gap without any difference in true frame-count accuracy. I considered the DreamSim contradiction in Table 6 as an alternative, but it is less decisive because DreamSim may legitimately penalize the varied camera angles that CineVerse intentionally produces, and the paper's dismissal, while terse, is at least a plausible research position; the frame-count issue has no such legitimate defense because it is purely a measurement problem. The concern does not by itself invalidate the pipeline or the human-preference results, so the reader's CONDITIONAL verdict remains appropriate; the authors should repair the frame-count measurement before the superiority claim is accepted.","tokens_in":16661,"tokens_out":6145,"duration_ms":65157,"concrete_test":"Manually annotate the true number of shots for a stratified sample of at least 50 of the 200 IC-LoRA outputs used in Table 4, then recompute IC-LoRA accuracy using human counts and report the row-difference heuristic's precision and recall. If accuracy at 3 shots rises substantially (e.g., from 34.84% toward the reported 95.45%), or if the heuristic has >10% error on boundary detection, the frame-count advantage is inflated by the metric. A complementary control: erase the checkerboard borders from CineVerse outputs and run the same row-difference heuristic; if accuracy drops materially, the comparison is method-dependent.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The claim that CineVerse produces 'the correct frame count' (Sec. 5.4) rests on Table 4, where CineVerse reports 95.45% accuracy at 3 shots versus IC-LoRA's 34.84%. The counting procedure is asymmetric. Appendix A.3 states that CineVerse boundaries are detected with Canny edges on the 16 px checkerboard borders that CineVerse itself inserts; IC-LoRA boundaries are inferred from the row pixel-difference heuristic. That heuristic is never validated against ground truth, and IC-LoRA's variable frame sizes make a fixed row-difference cue brittle. No precision/recall numbers are given for either detector. The concern is concrete: if the heuristic misses low-contrast cuts or over-segments gradual content, IC-LoRA's frame count will be systematically underestimated. Section 5.6 further admits that CineVerse outputs sometimes have missing borders, so even the Canny side is not guaranteed exact. Because 'correct frame count' is one of the three explicit superiority claims in Sec. 5.4, the headline quantitative advantage over IC-LoRA is not yet supported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"CineVerse is a two-stage framework for cinematic scene composition: a large language model converts a high-level scene description into a structured shot plan (setting, characters, and per-shot descriptions with shot sizes), and a LoRA-fine-tuned FLUX model generates a sequence of consistent keyframes. The paper introduces the CineVerse dataset, built from Storyboard20K with refined scene descriptions and new shot-level attributes, and adapts IC-LoRA by adding fixed frame heights, checkerboard borders between frames, and balanced training data. Experiments compare CineVerse against 1P1S, ConsiStory, StoryDiff, VideoStudio, and IC-LoRA using CLIP, DreamSim, GPT-4, LLaVa-OneVision, and human user studies, reporting consistent preferences for CineVerse on scene/shot alignment and some consistency metrics, as well as higher frame-count accuracy than IC-LoRA.","tokens_in":16868,"tokens_out":9131,"duration_ms":84084,"significance":"If the central claims hold, CineVerse represents a practical step toward controllable storyboard generation from text, with a useful dataset and a clear two-stage design that separates planning from generation. The inclusion of multiple preference-based evaluations (human, GPT-4, LLaVa) is a strength, as is the thoughtful dataset construction that addresses pronoun ambiguity and adds shot-level annotations. The work also identifies a real limitation of IC-LoRA (incorrect frame counts) and proposes a concrete fix. However, the quantitative support for some headline claims, especially frame-count accuracy and the dismissal of DreamSim results, needs strengthening before the conclusions are fully supported.","major_comments":[{"comment":"The frame-count accuracy comparison is asymmetric and the reported gap is not yet supported. CineVerse boundaries are detected from its own 16-pixel checkerboard borders with Canny edges, while IC-LoRA boundaries are inferred from an unvalidated row pixel-difference heuristic. No precision/recall or manual validation is reported for either detector, so the 95.45% versus 34.84% difference at three shots may be inflated by the measurement. Please validate the IC-LoRA detector on ground-truth shot counts (e.g., manually annotate a sample) or use a common detection protocol, and report confidence intervals for both methods.","section":"§5.3/Table 4 and App. A.3"},{"comment":"The text states that \"baseline methods show slightly higher DS scores due to the metric's bias toward static perspectives,\" but Table 6 lists DreamSim as lower-is-better (DS↓) and shows CineVerse with values around 0.55 versus baselines in the 0.21–0.42 range. This misstates the direction of the result: CineVerse is substantially worse on DreamSim, not the baselines. Please correct the description and either provide a controlled analysis supporting the metric-bias claim or temper the conclusion about this metric.","section":"§5.3/Table 6"},{"comment":"The claim that the user study shows \"a preference for our method over the baselines across all metrics\" is stronger than the data support. Several per-criterion preferences are near chance, for example Ours versus StoryDiff on Character (50.00) and Background (50.00), and Ours versus ConsiStory on Shot (52.44). With 50 responses per pair and no statistical tests or confidence intervals, these cells provide no evidence of superiority. Please report binomial tests with confidence intervals and qualify the consistency claims accordingly.","section":"§5.3/Table 3"},{"comment":"The paper does not specify whether all baselines receive the same detailed shot-by-shot plan (the script) or only the high-level scene description. If the baselines use their native input formats rather than the same per-shot prompts, the reported preferences may reflect input differences rather than method quality. Please state the input protocol for each baseline and, ideally, report an ablation where all methods receive identical per-shot text prompts.","section":"§5.1/Baselines"}],"minor_comments":[{"comment":"The sentence \"Table 5 compares various prompting strategies\" appears to be a cross-reference error; the prompting-strategy comparison is in Table 2, while Table 5 reports ablation results.","section":"§5.3"},{"comment":"The user-study section reports 50 responses per (CineVerse, baseline) pair per survey but does not state the number of unique participants or whether responses involve repeated measures; please clarify the participant count and analysis design.","section":"§5.2"},{"comment":"Please clarify whether the 800 test scenes (or their movies) are disjoint from the 1000 training scenes; the phrase \"from different movies\" is ambiguous and could imply a training/test split at the movie level.","section":"§5.1"},{"comment":"There is a formatting inconsistency in the IC-LoRA row: \"08.33\" has a leading zero and \"16.66\" appears where \"16.67\" is used elsewhere; please make the formatting and rounding consistent.","section":"Table 4"},{"comment":"LLaVa-OneVision is used both to generate the dataset's shot descriptions and to evaluate shot alignment in Table 8; the paper does not acknowledge this overlap, which could introduce bias in the textual-alignment evaluation, even though GPT-4 and human studies are also reported.","section":"App. A.3"}],"recommendation":"major_revision","confidential_remarks":"The paper addresses a timely task and the two-stage design is sensible. The main methodological concern is the asymmetric frame-count evaluation, which directly affects a headline quantitative claim. The DreamSim misreporting and the near-chance user-study cells also need to be addressed with better statistical reporting. I do not see a fundamental flaw in the method itself, so revision rather than rejection seems appropriate."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a competent systems paper with a genuinely new dataset and a sensible two-stage pipeline. The headline superiority claims rest on evaluation choices that are not all fair, but the core contribution does not collapse without them. If you work on storyboard or multi-image generation, it is worth a read.\n\nThe new thing is the CineVerse dataset: 46K shots with shot descriptions, character appearances, setting, and shot-size labels, built on Storyboard20K with MovieNet plot context. That is a real artifact and likely useful for future work. The method itself is honestly derivative: IC-LoRA on FLUX plus three preprocessing tweaks (16-px checkerboard borders, fixed 272-px height, no cropping) and balanced sampling. That is not a criticism; the paper says so plainly, and the ablations show each tweak matters.\n\nWhere it gets soft: Table 4's frame-count accuracy (95.45% vs 34.84% at 3 shots) is measured asymmetrically. CineVerse boundaries are found with Canny edges on its own checkerboard cue; IC-LoRA boundaries are inferred from a row pixel-difference heuristic that is never validated. The appendix admits CineVerse sometimes misses borders too, so even the Canny side is not clean. The real gap is probably large, but those exact numbers are not load-bearing until measured fairly.\n\nAlso, Table 6 has CineVerse with the worst DreamSim on every row (0.548 vs 0.267 for 1P1S), and the text waves it away as 'slightly higher' and 'metric bias.' That is not a fatal issue; DreamSim may indeed reward static layouts. But the dismissal is too quick, especially since DS is reported as 'lower is better' for consistency.\n\nThe user studies and GPT-4/LLaVa preferences are consistent and directionally convincing. No uncertainty intervals, though. The circularity burden is low; the frame counts etc. are not baked into the objective. LLaVa-OneVision annotated the training data and is also used as an evaluator, which is mild self-reference but not a definitional circularity.\n\nBottom line: this paper is for people building controllable storyboard generation. It deserves a serious referee. I would send it to review, ask for fair frame-count measurement, uncertainty on preference percentages, and a real attempt at explaining DreamSim.","headline":"A solid, honest systems paper with a genuinely useful dataset; the headline frame-count and DreamSim claims are over-sold, but the core contribution survives.","tokens_in":17461,"tokens_out":2420,"would_cite":true,"duration_ms":23791,"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":"CineVerse claims that a one-line scene description can be converted, through LLM-based shot planning and a fine-tuned text-to-image model, into a coherent storyboard whose keyframes match the planned shot count and shot sizes while…","keywords":["cinematic scene composition","keyframe synthesis","shot planning","large language model","text-to-image generation","character consistency","diffusion transformer","storyboard generation"],"falsifier":"Run a blinded frame-count benchmark on the same test scenes with a single, ground-truth-validated boundary detector for both methods; if three-shot accuracy for IC-LoRA rises toward CineVerse's 95.45% or CineVerse falls toward IC-LoRA's 34.84%, the reported superiority is largely a measurement artifact. A second check: have human raters count the number of shots in each generated sequence without seeing borders, and compare those counts.","tokens_in":16437,"feed_emoji":"🎬","tokens_out":10343,"duration_ms":93034,"temperature":0.7,"pith_summary":"CineVerse claims that cinematic scene composition can be broken into two stages: a large language model turns a one-line scene description into a structured script—setting, characters, and a numbered sequence of shots with explicit sizes (wide, medium, close-up)—and a fine-tuned text-to-image model turns that script into a matching set of keyframes. The paper builds a movie-keyframe dataset with shot-level annotations, trains on it, and reports that the generated keyframes follow the planned shot count and sizes while keeping characters and settings consistent. A sympathetic reader would care because this makes storyboard generation practical: instead of hoping a single prompt yields a coherent set of images, users get a plan they can inspect and edit before any image is rendered. If the claim holds, everyday users could pre-visualize multi-character scenes in a filmmaking grammar without manual shot-by-shot prompting.","feed_headline":"One scene description becomes a consistent shot-by-shot storyboard","feed_subtitle":"A language model plans the shots; a fine-tuned image model keeps characters and settings consistent.","key_machinery":"The object that carries the argument is a two-stage pipeline ending in a fine-tuned text-to-image model. Stage one uses in-context prompting to make a large language model output a structured movie script—setting, character descriptions, and a list of shots each labeled wide, medium, or close-up. Stage two adapts IC-LoRA, a method that generates several images at once from one concatenated prompt, to a FLUX diffusion transformer via LoRA fine-tuning. The decisive mechanism is a 16-pixel checkerboard border inserted between frames: it functions as an explicit spatial cue that lets the model separate frames and learn the intended number of shots, and it is what the paper's automated frame-count measurement detects. The pipeline is trained on a new dataset of roughly 8.5k multi-shot movie scenes with shot descriptions, shot-size labels, settings, and character appearances extracted by a vision-language model.","core_discovery":"The central discovery of the paper is that the two-stage CineVerse pipeline makes multi-shot keyframe generation controllable. Given only a scene description, the LLM planning stage produces a shot plan whose format the authors call a movie script, with a setting, per-character appearance descriptions, and per-shot descriptions that each name a shot size. The generation stage takes this plan as one concatenated prompt with special tokens separating frames, and a FLUX-based diffusion model fine-tuned with LoRA on the CineVerse dataset renders all keyframes together. Two technical choices carry most of the weight: fixed-height frames with a 16-pixel checkerboard border between shots, which gives the model an explicit spatial cue for frame separation and correct shot counts, and removal of the cropping step that truncated first and last frames in the base approach. The paper reports that CineVerse outperforms all five baselines in text alignment and is preferred by human raters, GPT-4, and a vision-language model across scene alignment, consistency, and continuity, with a frame-count accuracy of 95.45% versus 34.84% for IC-LoRA on three-shot scenes.","pith_inferences":["Because the structured plan separates shot decisions from pixel generation, the same two-stage recipe should transfer to other structured visual narratives—comics, animatics, or multi-panel diagrams—where the number and type of panels are specified beforehand; the paper only evaluates movie keyframes.","The explicit border cue suggests a simpler design principle for multi-frame models: when a model must output a fixed number of images, give it spatial markers for the boundaries rather than forcing it to infer the count from prompt length alone; this could improve other in-context generation methods.","The dataset pipeline—using a vision-language model to turn real movie frames into structured shot descriptions—could be reused to create training data for longer sequences or full scenes, since the same annotations are exactly what a generation model needs; the paper does not test this reuse."],"forward_implications":["A user can go from one line of text to a storyboard whose shot count matches the plan, eliminating the need to hand-prompt each frame.","Directors and editors can inspect and modify the shot plan—shot sizes, character descriptions, setting—before committing to image generation.","Multi-character scenes with wide/medium/close-up grammar become tractable, since the model is trained on real movie keyframes with labeled shot sizes.","The reported shot-count accuracy grows harder at longer sequences but remains high at short ones (95.45% at three shots, 42.24% at ten), suggesting the approach is most reliable for typical scene lengths.","Human preference and GPT-4/vision-language-model ratings favor CineVerse over 1P1S, ConsiStory, StoryDiff, VideoStudio, and IC-LoRA across textual alignment, consistency, and continuity."],"supporting_citations":[{"why":"Supplies the in-context LoRA generation approach that CineVerse adapts for multi-shot keyframe synthesis.","marker":"[12]"},{"why":"Provides the underlying storyboard scenes and bounding-box annotations on which the CineVerse dataset is built.","marker":"[31]"},{"why":"Adds global plot annotations used to resolve pronouns and clarify scene descriptions.","marker":"[13]"},{"why":"The vision-language model used to extract shot descriptions, settings, and character appearances for the dataset.","marker":"[15]"},{"why":"The text-to-image diffusion backbone whose learned prior allows consistent multi-frame generation.","marker":"[14]"},{"why":"The low-rank adaptation method used to fine-tune the generator on the CineVerse dataset.","marker":"[11]"},{"why":"The LLM that performs in-context scene planning from scene descriptions.","marker":"[6]"},{"why":"Supplies the off-the-shelf shot-size classifier used to label keyframes as wide, medium, or close-up.","marker":"[1]"}],"fun_headline_variants":["Scene description to consistent movie storyboard in one go","LLM plans shots, diffusion renders them with consistent look","CineVerse: 95.45% frame-count accuracy for multi-shot keyframes","Two-stage pipeline turns one scene description into consistent shots"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The claim that CineVerse reliably outputs the planned number of shots rests on a fair way of counting frames in both methods, and the counting rule applied to the comparison method is a pixel-difference heuristic that the paper never validates against ground truth.","fun_headline_variants_meta":{"raw":{"variants":["Scene description to consistent movie storyboard in one go","LLM plans shots, diffusion renders them with consistent look","CineVerse: 95.45% frame-count accuracy for multi-shot keyframes","Two-stage pipeline turns one scene description into consistent shots"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000325,"raw_usage":{"total_tokens":1815,"prompt_tokens":930,"completion_tokens":885,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":546,"completion_tokens_details":{"reasoning_tokens":814}},"tokens_in":546,"tokens_out":885,"duration_ms":8320,"temperature":1.0,"reasoning_tokens":814,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T05:41:09.394477+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a blinded frame-count benchmark on the same test scenes with a single, ground-truth-validated boundary detector for both methods; if three-shot accuracy for IC-LoRA rises toward CineVerse's 95.45% or CineVerse falls toward IC-LoRA's 34.84%, the reported superiority is largely a measurement artifact. A second check: have human raters count the number of shots in each generated sequence without seeing borders, and compare those counts.","supporting_citations":[{"cited_title":"Movienet: A holistic dataset for movie under- standing","cited_arxiv_id":null,"evidence_quote":"Adds global plot annotations used to resolve pronouns and clarify scene descriptions."},{"cited_title":"Lora: Low-rank adaptation of large language models","cited_arxiv_id":null,"evidence_quote":"The low-rank adaptation method used to fine-tune the generator on the CineVerse dataset."},{"cited_title":"The anatomy of video editing: A dataset and benchmark suite for ai-assisted video editing","cited_arxiv_id":null,"evidence_quote":"Supplies the off-the-shelf shot-size classifier used to label keyframes as wide, medium, or close-up."}],"review_version":1}