ContextMaster: Interactive Multi-Shot Video Creation via Fixed-Budget Sparse Context Routing

• Across T2MV, R2MV, V2MV, and compositional X2MV, ContextMaster improves task fulfillment and cross-shot consistency over task-specific baselines, reaching 16 FPS on a single GPU.Source: Figure 1 · p. 2, §1 · p. 3
• IMVC formalizes stateful multi-turn generation, reference incorporation, and editing over evolving shared visual history, using role-aware rotary coordinates to distinguish heterogeneous context streams and preserve shot alignment.Source: §1 · pp. 1–3
• Cacheable clean-context states and query-dependent block-sparse reads bound denoising access as history grows; ConstraintSink reserves explicit references and aligned sources while remaining capacity retrieves relevant history.Source: §1 · pp. 2–3
• Privileged context distillation transfers dense full-context behavior to a sparse few-step student through consistency distillation, then refines deployment-matched rollouts with distribution matching.The refinement recovers perceptual detail and reduces errors accumulated in generated history.Source: §1 · pp. 2–3

Unified and interactive video creation

• Holistic methods plan multi-shot sequences, streaming methods generate from historical memory, and unified editors generally separate generation, reference conditioning, and editing; IMVC composes them statefully.The cited holistic, streaming, and unified approaches are described in the passage.Source: §2 · p. 3

Heterogeneous and growing visual context

• Prior work distinguishes shot identities and manages growing history with sampling, windows, sinks, retrieval, or memory; ContextMaster shares a fixed read budget across mandatory and dynamic context.Role-aware coordinates cover reference, history, source, and target streams; clean context is cached across denoising steps and history is retained as routing candidates.Source: §2 · p. 3, Figure 2 · p. 4

Sparse and few-step video diffusion

• Unlike sparse-attention and few-step diffusion methods, ContextMaster distills from a teacher reading complete clean context, then distribution-matches fixed-budget sparse rollouts.The two stages transfer teacher denoising behavior through consistency distillation and refine sparse few-step rollouts through distribution matching.Source: §2 · p. 3

3.1 IMVC Formulation and Role-Aware Context

• The role-aware interface separates finalized shot history from request-local references and sources, while encoding each visual role with distinct temporal semantics.Accepted outputs enter history and are re-encoded next round; rejected candidates can instead serve as sources for further editing. Role and frame coordinates distinguish inputs, with source and target frames aligned during editing.Source: §3.1 · pp. 4–5

3.2 Cacheable Fixed-Budget Context

• Routing splits residual capacity between source and history, preventing strong local source similarity from suppressing long-range history without cross-role score calibration.When a role is absent or has too few candidates, unused capacity transfers to the other role; mandatory and routed blocks share the same budget.Source: §3.2 · pp. 5–6
• Each target query reads at most B context blocks, with reference and aligned source blocks kept visible before adaptive retrieval.Clean context states are cached once per interaction round and reused during sampling, making latency-critical context reads independent of history length.Source: §3.2 · pp. 5–6

3.3 Privileged Context Distillation

• Privileged consistency distillation trains the sparse student to match clean predictions across adjacent noise levels after a full-context teacher transition.The frozen dense teacher uses classifier-free guidance, while the student evaluations are conditional, absorbing guidance without an unconditional inference pass.Source: §3.3 · p. 6
• Distribution matching refinement restores fine appearance and editing detail using full-context real scores and budgeted fake scores on student samples.Training samples deployment exit steps and autoregressively generated histories, so refinement accounts for few-step errors and errors accumulated in the student’s own history.Source: §3.3 · pp. 6–7

3.4 Interactive Inference

• An optional external Director plans shots, retrieves references, and localizes source segments, then executes the specifications through the same video model.Its planning and tool calls do not alter the model’s latency-critical denoising path.Source: §3.4 · p. 7

4.1 Setup

• Evaluation covers 100 T2MV prompts, 50 five-shot R2MV cases, 50 V2MV editing cases, and 20 X2MV interactive trajectories assessed in a user study.All methods receive the same shot-level instructions and required reference or source inputs.Source: §4.1 · p. 7
• Table 1 compares T2MV, R2MV, and V2MV on text alignment, aesthetic quality, motion smoothness, dynamic degree, and task fulfillment; reported speed is 16.74 FPS.Task fulfillment is rated by Gemini 2.5 Pro on a 1–5 scale; speed averages five-shot runs across the three tasks.Source: Table 1 · p. 8
• The deployed student starts from Wan2.1-T2V-1.3B, generates 832 × 480 clips with 81 frames per shot, and uses four denoising steps and a 6 FE active-read budget.Accepted shots are uniformly sampled at 1 FPS to build append clean-context memory; the same sparse operator and context budget are used for distillation, refinement, and inference.Source: Implementation details · p. 8
• Inference uses block-wise zeroth-order Taylor routed sparse attention in TileLang, and training uses an internal dataset of 1M multi-shot videos plus public reference and editing datasets.Training and data-construction details are provided in the supplementary material.Source: Implementation details · p. 8

Implementation details.

• Evaluation measures video quality with TA, AQ, MS, and DD; consistency with DINO-based subject similarity, CLIP-based background similarity, and Inter-Shot; task fulfillment with Gemini 2.5 Pro; and efficiency on a single H200 GPU under matched resolution and frame count.Inter-Shot groups shots containing the same character and compares their ViCLIP features; task fulfillment is rated on a 1–5 scale.Source: Implementation details · p. 8

4.2 Comparison

• On R2MV, ContextMaster’s largest gains over Phantom are in Inter-Shot and TF, and it reaches 16.74 FPS; Phantom retains higher DD and slightly higher intra-shot subject consistency.The qualitative example shows more stable reference appearance across close-up and wider views, preserving the elderly potter through the described vessel-making progression.Source: §4.2 · pp. 8–9
• On V2MV, ContextMaster leads or ties on eight of nine reported metrics, while VideoCoF retains a small AQ advantage.It applies the requested appearance edit across full-body, close-up, and wide shots while preserving source scene structure.Source: §4.2 · p. 9

4.3 User Study

• In the user study, the model obtains the highest IF and CC scores against the strongest baseline across the three primitive tasks; on X2MV, scores remain favorable across mixed creation trajectories.Participants rated VQ, IF, TC, and CC on a 1–5 scale; Table 2 reports mean ratings.Source: §4.3 · p. 9, Table 2 · p. 9

4.4 Ablation Studies

• With the 6-FE budget fixed, role separation benefits all task modes, while mandatory reference or aligned-source routing is especially important for reference-guided generation and editing.Without ConstraintSink, similarity-based Top-k routing leaves visible source remnants; the full model removes the subject and reconstructs the chair and background.Source: Table 3 · p. 9, §4.4 · p. 10, Figure 4 · p. 10
• On T2MV, PCD primarily improves Inter-Shot and TF, DMD contributes more to appearance quality, and their sequential combination performs best on every metric.The qualitative ablation shows PCD followed by DMD preserving recurring identity while retaining sharper visual detail.Source: Table 3 · p. 9, §4.4 · p. 10, Figure 4 · p. 10

5 Conclusion

• ContextMaster formalizes interactive multi-shot creation as stateful composition of generation, reference conditioning, editing, and continuation over persistent visual history.Role-aware rotary coordinates distinguish heterogeneous visual roles.Source: §5 · p. 10
• 16 FPS; ContextMaster improves task fulfillment and cross-shot consistency across three primitive tasks and compositional trajectories.The conclusion reports these gains alongside the model’s inference speed.Source: §5 · p. 10
• Asymmetric clean-context prefill and fixed-budget sparse routing bound each target query’s active reads, while ConstraintSink preserves reference and source correspondences within that budget.These mechanisms keep relevant constraints visible without increasing the active read budget.Source: §5 · p. 10
• Privileged context distillation transfers full-context teacher behavior through consistency distillation, then refines deployment-matched rollouts with distribution matching.The two stages train the sparse few-step model.Source: §5 · p. 10

A Interactive Director

• Given a concise request, the Interactive Director expands it into an ordered multi-shot script and prepares conditions for each operation.It is a tool-augmented VLM agent inspired by Aurora (Yu et al., 2026).Source: A Interactive Director · p. 14
• The Director can retrieve missing reference images through the Serper API, identify editing requests, and invoke OmniShotCut (Wang et al., 2026a) to segment source video and localize shots for modification.These tools prepare references and source segments for ContextMaster operations.Source: A Interactive Director · p. 14
• The Director dispatches instructions, references, and source segments through ContextMaster’s generation, reference, and editing interfaces, then commits accepted outputs to shared visual history for subsequent planning.The updated project state supports the Director’s next operations.Source: A Interactive Director · p. 14

B.1 Training Configuration

• The dense full-context teacher uses AdamW at 1 × 10−5 with 200 warmup steps, 0.01 weight decay, mixed precision, and initializes both PCD student and EMA target.Source: §B.1 · p. 14
• PCD uses a 2 × 10−6 learning rate, 0.99 EMA decay, 50-level teacher-noise discretization, and the deployment four-step schedule.The teacher uses classifier-free guidance scale 6.0, which is absorbed into the student for single-branch inference.Source: §B.1 · p. 14
• DMD initializes its generator from the PCD student; its frozen real-score model reads dense full context with classifier-free guidance, while the trainable fake-score model shares the generator’s fixed-budget context.Source: §B.1 · p. 14
• DMD uses five fake-score updates per generator update, and all experiments are conducted on 64 NVIDIA H200 GPUs.Source: §B.1 · p. 14

B.2 Training Data

• Teacher training starts with public HuMo (Chen et al., 2025), Kiwi-Edit (Lin et al., 2026), and Ditto-1M (Bai et al., 2025) data for single-shot reference-conditioned generation and video editing.Source: §B.2 · p. 15
• An internal collection of approximately one million multi-shot videos supplies T2MV, R2MV, and V2MV examples; R2MV uses ID-similar identity-matched references, while V2MV uses OmniShotCut (Wang et al., 2026a) segmentation and Kling 3.0-generated edit targets.Joint training samples T2MV, R2MV, and V2MV at a 4:3:3 ratio; V2MV targets cover addition, removal, and replacement.Source: §B.2 · p. 15

C Additional Ablation Studies

• 6 FE is selected as the quality–efficiency operating point: increasing the budget from 2 to 6 FE substantially improves AQ, Inter-Shot consistency, and TF, while 8 FE yields only marginal gains and reduces throughput by 0.56 FPS.All configurations use the same checkpoint and evaluation protocol on five-shot T2MV sequences.Source: C Additional Ablation Studies · p. 15

D.1 User Study

• Figure 6 shows the interface used to collect human ratings across four evaluation criteria.Source: §D.1 · p. 15

D.2 Gemini Task-Fulfillment Evaluation

• Figure 7 provides the task-specific prompts used by Gemini 2.5 Pro to evaluate task fulfillment.Source: §D.2 · p. 16

E Limitations

• Approximately 0.4 FPS per additional shot is lost because the context branch performs bidirectional prefill over the full accumulated history to construct its KV cache.Future work will explore persistent per-shot caches and compact shot summaries.Source: E Limitations · p. 16

• Figure 8 presents additional qualitative results of ContextMaster on T2MV.Source: F Additional Qualitative Results · p. 16, Figure 8 · p. 18
• Figure 9 presents additional qualitative results of ContextMaster on V2MV.Source: F Additional Qualitative Results · p. 16, Figure 9 · p. 19
• Figure 10 presents additional qualitative results of ContextMaster on R2MV and X2MV.Source: F Additional Qualitative Results · p. 16, Figure 10 · p. 20