Current multi-shot video generators render only limited durations per pass, requiring long movies to be stitched from local segments across which narrative progression and visual identity drift. To synthesize long-horizon movies with Coherence across Time, we present CaT-Movie, a multi-shot generation system consisting of a global story planner and a local window renderer. The planner expands user requests via a large language model to preserve, infer, and complete story details, while deterministic code maintains the resulting storyboard across shots to compile self-contained shot prompts and index reference frames from generated history. To execute this guidance, the renderer applies reference-routed sparse self-attention and shot-wise cross-attention to route identities from the indexed frames while isolating per-shot textual conditions. Given the scarcity of high-quality open-source multi-shot data, we construct a dataset of over 350K film shots annotated with semantic state transitions, entity grounding, and multi-granularity requests. We further build CaT-MBench, a benchmark on held-out films with the same annotations. It gives every baseline its native format and counts cross-shot consistency only when a cut opens a new camera setup, preventing repeated framing from passing as consistency. On CaT-MBench, CaT-Movie attains the highest average score and the most consistent characters among all methods. The project page is available at https://jinplu.github.io/CaT-Movie/.
Figures & tables
Figure 1: Cross-window drift, copied framing, and cut-gated evaluation. Each film strip is one generation pass. Joint generation switches camera setup at the cut from shot 3 to 4 but changes the man in the next pass, from shot 4 to shot 6, whereas memory-augmented generation keeps the man yet repeats the framing of shot 3 across planned cuts. Scores average character and scene consistency over all 11 shots, and the cut gate of CaT-MBench fails consistency checks across planned cuts that keep the camera setup.
Figure 2: Overview of CaT-Movie. A planner writes one storyboard, from which a prompt compiler maintains cross-shot state and derives per-shot text ck , the frames each shot recalls, and an identity line cref for each recalled frame. The renderer denoises each window jointly with reference-routed sparse self-attention and shot-wise cross-attention, placing recalled frames before the reading shot’s first latent frame.
Figure 3: Overview of the dataset and benchmark. One film pool and one schema feed the training data, (a) curation and annotation and (b) entity grounding, and CaT-MBench, (c) stories and questions and (d) cut-gated judging. The center chart gives the composition of both.
Transition Control ↑
Inter-shot Consistency ↑
State Continuity ↑
Text ↑
Method
Video base
Cut F1
Execution
Character
Object
Scene
Establish
Persist
Alignment
Avg. ↑
Per-shot generation
Wan2.1, per shot
Wan2.1-T2V-1.3B
0.998
0.994
0.219
0.169
0.310
0.274
0.694
0.181
0.474
Wan2.2, per shot
Wan2.2-TI2V-5B
0.999
0.996
0.227
0.275
0.310
0.387
0.701
0.174
0.497
StoryMem
Wan2.2-I2V-A14B
0.928
0.719
0.631
0.453
0.687
0.484
0.545
0.179
0.527
StoryDiffusion
Wan2.2-I2V-A14B
0.905
0.825
0.300
0.265
0.245
0.597
0.535
0.171
0.468
Table 1: Quantitative comparison on 92 test stories. CaT-Movie renders its planner’s storyboard from the shot-list request, which states less than the reference storyboard the baselines render. Avg. averages the four aspects, with best and second-best marked.
Figure 4: Inter-shot consistency against transition control on the 92 test stories. The dotted line marks the baseline Pareto frontier, and the blue region lies beyond every baseline.
Figure 5: Qualitative comparison on a test story of 11 shots and 28.4 s , which HoloCine generates in three windows. Each of four shots is shown at its middle frame. Icons mark characters that CaT-MBench judges inconsistent, enlarged at the bottom right of the shot, and crossed scissors mark planned cuts where the video keeps the camera setup.
Setting
Exec.
Char.
Obj.
Scene
Persist
Avg.
CaT-Movie
0.958
0.581
0.585
0.566
0.615
0.571
Trained without
identity lines
0.942
0.582
0.607
0.557
0.653
0.561
routing
0.923 †
0.593
0.513
0.595
0.609
0.564
references
0.966
0.289 †
0.342 †
0.474
0.696
0.521 †
Reference selection
Table 2: Ablations on 30 development stories over three seeds. CaT-Movie renders the reference storyboard, the first group retrains from the base model, and the last replaces it with the shot-list request compiled directly or our planner’s storyboard. Avg. also counts Cut F1, Establish, and text alignment, and † marks a significant difference from CaT-Movie.
Figure 6: (Left) Character checks passed by time from first appearance to return , pooled over the test stories, with 95% intervals. (Right) Consistency by window distance on development stories, where Recent is recent-shots recall and W/o refs trains without references.
Figure 7: Planner text and rendering. (Left) Opening states and held items of the reference storyboard stated by the request alone and with our planner, for synopsis, paragraph, and shot-list requests marked Syn, Para, and Shot. (Right) State questions of the storyboard rows of Tab. 2 whose change the per-shot text states and the video shows, states but the video omits, or leaves unstated.
Appendix figures & tables14 assets
Supplementary material from the paper’s appendix.
Appendix
Item
Setting
Training
Model
Wan2.2-TI2V-5B; all 5.0B transformer parameters trained, text encoder and VAE frozen
Objective
Rectified flow velocity target of Eq. (3), with the MSE averaged over target latents and Wan’s per-timestep weight; timesteps uniform with shift 5; reference latents excluded
Video
1280×704 at 24 fps
Windows
Complete source spans of at most 8 s (128 frames at 16 fps); a longer shot fills a window alone and shots over 10 s are excluded; one, two, three, and four or more shots drawn at 15, 25, 30, and 30%, half of the one-shot draws being shots of at least 5 s with an action or a within-shot change; windows beyond the measured memory envelope of 8 GPUs (69 latent frames, 4,209 attended frame pairs) are excluded
References
1 to 10 frames from other usable shots of the same scene block, shots sharing an entity with the window first, one random frame per shot from the 16 fps annotation grid, in shot order; each zoom-cropped to a random 65–100% of its area at the original aspect ratio, the crop covering all registered characters in the frame and skipped when they do not fit (about 47% cropped); temporal RoPE positions five latent frames apart, ending five before the reading shot’s first latent frame; whole group dropped with probability 0.2; noise σ∼0.2U[0,1) , applied as (1−σ)x+σϵ ( Fig. S1 ); each frame carries an identity line listing the registered characters annotated as visible in it and the location of its source shot
Appendix
Table S1 : Renderer configuration. Frame counts are planned at 16 fps and rendered at 24 fps.
Figure S1 : Reference noise. (Left) Power gap between generated and real first latent frames, where Low, Mid, and High pool the lowest two, middle four, and highest two radial frequency bands and the gray band is the sampling spread of real frames. (Right) DINOv2 [ 32 , 24 ] cosine between decoded noised and clean real references, median with 10–90% range, against the median of two adjacent shots. Hatching marks the training range of σ .
Setting
Exec.
Char.
Obj.
Scene
Persist
Avg.
CaT-Movie
0.958
0.581
0.585
0.566
0.615
0.571
Window cap
7 s / window
0.969
0.576
0.594
0.601
0.625
0.573
6 s / window
0.960
0.571
0.620
0.580
0.701
0.583
Packing
even packing
0.958
0.578
0.538
0.549
0.635
0.570
Appendix
Table S2 : Window settings on 30 development stories over three generation seeds, under the renderer ablation of Table 2.
Figure S2 : State changes in the text of the planner ablation: the share of state questions whose change the per-shot text states in the changing shot, when the character returns, and over all its shots.
Figure S3 : Planner on one development story , seed 44. Each film strip is one generation pass of the same renderer, which reads the storyboard of its row. Shot 3 is shown at its end and the other shots at their middle frame. Under each row is the shot-3 text the renderer read, with the opening and closing states in bold; Object 3 is the white cloth, and the marks give the judge’s answer on whether Character 2 holds it at the end of shot 3.
Entries
Code
Model
Qwen3.8-27B, thinking at medium effort
Held item (454)
59.0
57.0
Outfit (1,730)
85.9
82.7 †
Outfit after change (256)
55.1
52.7
change recorded
52.7
53.9
correct if recorded
65.9
65.2
Appendix
Table S5 : Cross-shot state stated correctly in later shots (%), change sample, two seeds, for each planner model. † marks a significant difference from code. The indented rows split the outfit entries after a change by whether the plan records that change.
Table S7 : Multi-shot training data in detail. Size counts clips, episodes for DramaBoard, and blocks for ours. Present per shot lists the registered entities in each shot, and Camera gives each shot’s size, angle, or movement. Per shot gives each entity’s posture, clothing, or held object, and before/after the initial and final values of a within-shot change. ∗ marks artifacts listed on the project page.
Dimension
Question
Character
Is Character 1 at shot 3 the same person as at shot 1?
Object
Is Object 5 at shot 10 the same object as at shot 8?
Scene
Are shot 2 and shot 1 in the same place (Location 1)?
Establish
End of shot 3: Is Object 3 open? Passes only with “Does Character 5 open Object 3?” in shot 3.
Persist
Shot 4: Is Object 3 open?
Appendix
Table S9 : Example questions from two test stories, one per scored dimension.
Dimension
Items
Judge κ
Bal. acc.
Annot. κ
Execution
115
0.724
0.906
0.903
Character
87
0.609
0.809
0.845
Object
91
0.641
0.836
0.719
Scene
88
0.592
0.796
0.761
Establish
87
0.770
0.885
0.810
Persist
89
0.775
0.888
0.946
Appendix
Table S11 : Agreement between the judge and reviewed human labels. Bal. acc. averages the judge’s accuracy on items the labels pass and on those they fail, and Annot. κ compares each annotator’s blind labels with the reviewed labels, pooled over 532 annotator–item pairs.
Inter-shot Consistency
Judge
ViCLIP
Method
Execution
Gated checks
With gate
No gate
Character
Same character
All pairs
Source films
–
–
–
–
–
0.665
0.598
Per-shot generation
Wan2.1, per shot
0.994
0.000
0.233
0.233
0.219
0.658
0.583
Wan2.2, per shot
0.996
0.000
0.271
0.271
0.227
0.654
0.578
StoryMem
0.719
0.157
0.590
0.765
0.631
0.749
0.674
Appendix
Table S12 : What the gate changes, and ViCLIP against the judge on the 92 test stories. Gated checks is the share of consistency checks behind a failed gate. ViCLIP pairs shots showing the same character or all shots.
Table S13 : Multi-shot benchmarks in detail. Columns follow Tab. S7 without grounding, which no benchmark here labels. Size counts stories, episodes for EntityBench, clips for CineBench, and scripts for CAST-Bench. A dash marks a script with no video, and ∗ marks artifacts listed on the project page.
Method
Configuration
Input conversion
Wan2.1-T2V-1.3B
832×480 , 16 fps, UniPC 50 steps, CFG 6, shift 8
One shot per call with its compiled caption, no prompt extension; shots concatenated.
Wan2.2-TI2V-5B
1280×704 , 24 fps, UniPC 50 steps, CFG 5, shift 5
As above.
MultiShotMaster
832×480 , 50 steps, CFG 5, shift 5; up to 5 shots and 19.3 s (308 frames) per segment; saved at 16 fps
Story: global caption and Now: shot caption, with Subject N tags for characters, objects, and places.
HoloCine
832×480 , 50 steps, CFG 5, shift 5; up to 15.1 s (241 frames) and 512 tokens per segment; saved at 16 fps
Global caption, per-shot captions joined by [shot cut] , [characterN] tags, and cut indices rounded to 4n+1 as the model requires.
StoryMem
832×480 , 16 fps, UniPC 40 steps, CFG 3.5, shift 12 for the first shot and 4 after; memory of up to 10 keyframes (first 3 fixed, newest 7)
Each shot rendered at its planned length from its caption; every shot boundary updates the memory.
SDXL draws each shot’s keyframe from its opening state with consistent characters; the I2V model animates it with the full caption, including the end-state clause.
Appendix
Table S14 : Baseline configurations and input conversion. Frame counts planned at 16 fps are kept, except for official Wan2.2, which rounds 1.5× the planned count to the nearest 4n+1 .
Story
Shots
Seconds
cinedance-009077-b0004-shot01
12
36.3
cinedance-010142-b0004-shot01
6
19.4
cinedance-010691-b0000-shot01
7
39.2
cinedance-010942-b0000-shot03
11
28.4
cinedance-014083-b0000-shot01
12
21.8
cinedance-014481-b0000-shot01
7
26.7
Appendix
Table S16 : Stories in the videos on the project page.
configs/train/renderer_wan22_final_fixed_role_rows.py , renderer_wan22_final_window_shared.py , and renderer_wan22_final_without_references.py : the final configuration without identity lines, routing, or references
Planner ( Appendix B )
tools/run/planner_form_plans.py plan --prompt current --sampling thinking-medium , served by run/vllm_serve.py --config configs/serving/vllm_planner_qwen38_27b.yaml
Appendix
Table S17 : Code entry points. Paths are relative to the repository root.