cs.AIApr 7, 2026
SaveSCMAPR: Self-Correcting Multi-Agent Prompt Refinement for Complex-Scenario Text-to-Video Generation
Organizations: HiThink Research · East China Normal University · Guangzhou University · Tsinghua University
Abstract
Text-to-Video (T2V) generation has benefited from recent advances in diffusion models, yet current systems still struggle under complex scenarios, which are generally exacerbated by the ambiguity and underspecification of text prompts. In this work, we formulate complex-scenario prompt refinement as a stage-wise multi-agent refinement process and propose SCMAPR, i.e., a scenario-aware and Self-Correcting Multi-Agent Prompt Refinement framework for T2V prompting. SCMAPR coordinates specialized agents to (i) route each prompt to a taxonomy-grounded scenario for strategy selection, (ii) synthesize scenario-aware rewriting policies and perform policy-conditioned refinement, and (iii) conduct structured semantic verification that triggers conditional revision when violations are detected. To clarify what constitutes complex scenarios in T2V prompting, provide representative examples, and enable rigorous evaluation under such challenging conditions, we further introduce T2V-Complexity, which is a complex-scenario T2V benchmark consisting exclusively of complex-scenario prompts. Extensive experiments on 3 existing benchmarks and our T2V-Complexity benchmark demonstrate that SCMAPR consistently improves text-video alignment and overall generation quality under complex scenarios, achieving up to 2.67% and 3.28 gains in average score on VBench and EvalCrafter, and up to 0.028 improvement on T2V-CompBench over 3 State-Of-The-Art baselines. The codes of SCMAPR are publicly available at https://github.com/HiThink-Research/SCMAPR.
Figures & tables
Figure 1: Self-Correcting Multi-Agent Prompt Refinement Framework (SCMAPR). SCMAPR organizes prompt refinement as a stage-wise multi-agent collaboration involving six specialized agents. The framework proceeds through five functional stages: (I) Scenario Routing , where Scenario Router assigns a scenario tag to the input prompt. (II) Policy Synthesis , where a Policy Generator generates a scenario-conditioned rewriting policy. (III) Policy-Conditioned Refinement , where a Prompt Refiner rewrites the prompt. (IV) Semantic Verification , where Atomizer and Validator collaboratively verify semantic fidelity through atomic extraction and entailment judgment. (V) Conditional Revision , where verification feedback conditionally triggers targeted revision, enabling self-correcting refinement.
Figure 2: Illustration of the Semantic Verification Stage in SCMAPR. Given a user input and the corresponding refined prompt, semantic verification is performed in four steps. (1) Atomic Extraction decomposes the user input into atom elements. (2) Chunking segments the refined prompt into semantically coherent evidence units. (3) Atom-Chunk Matching retrieves the most relevant evidence chunk for each atom. (4) Entailment Validation assesses atom-level semantic relations between atoms and evidence chunks. Through this design, semantic missing and contradictions in the refined prompt can be detected and subsequently used to trigger downstream revision.
| Method | Average Score | Aesthetic Quality | Background Consistency | Imaging Quality | Motion Smoothness | Subject Consistency | Temporal Flickering |
|---|---|---|---|---|---|---|---|
| LaVie | 81.89 | 53.52 | 95.01 | 62.83 | 95.15 | 90.96 | 93.85 |
| LaVie + Open-Sora | 81.95 | 53.67 | 94.82 | 62.93 | 95.28 | 91.24 | 93.78 |
| LaVie + RAPO | 82.80 | 54.38 | 95.76 | 63.06 | 96.03 | 93.21 | 94.35 |
| LaVie + SCMAPR | 84.56 | 56.53 | 97.43 | 63.87 | 97.18 | 95.95 | 96.39 |
| Wan | 86.19 | 63.75 | 95.41 | 69.78 | 97.62 | 95.12 | 95.43 |
| Wan + Open-Sora | 86.27 | 63.82 | 95.22 | 70.05 | 97.76 | 95.28 | 95.52 |
Table 1: Quantitative results (%) on VBench with LaVie and Wan as T2V backbones. Bold values denote the best performance for each metric under each backbone.
| Method | Average | Motion Quality | Text-Video Alignment | Visual Quality | Temporal Consistency |
|---|---|---|---|---|---|
| LaVie | 62.12 | 53.19 | 69.60 | 64.81 | 60.87 |
| LaVie + Open-Sora | 62.78 | 53.07 | 71.38 | 65.26 | 61.41 |
| LaVie + RAPO | 63.91 | 53.34 | 74.38 | 66.62 | 61.29 |
| LaVie + SCMAPR | 65.18 | 53.87 | 75.42 | 68.84 | 62.56 |
| Wan | 63.46 | 53.79 | 73.05 | 65.17 | 61.85 |
| Wan + Open-Sora | 63.95 | 53.93 | 73.84 | 65.60 | 62.43 |
Table 2: Quantitative comparisons on EvalCrafter. SCMAPR consistently achieves better results, demonstrating a clear lead on this benchmark.
| Method | Average | Consistent Attribute | Dynamic Attribute | Action Binding | Motion Binding |
|---|---|---|---|---|---|
| LaVie | 0.388 | 0.620 | 0.232 | 0.483 | 0.215 |
| LaVie + Open-Sora | 0.361 | 0.532 | 0.214 | 0.470 | 0.226 |
| LaVie + RAPO | 0.460 | 0.692 | 0.267 | 0.635 | 0.243 |
| LaVie + SCMAPR | 0.476 | 0.704 | 0.273 | 0.640 | 0.285 |
| Wan | 0.454 | 0.694 | 0.263 | 0.591 | 0.269 |
| Wan + Open-Sora | 0.446 | 0.672 | 0.258 | 0.583 | 0.274 |
Table 3: Quantitative comparisons on T2V-CompBench. SCMAPR achieves the highest average score.
| Method | Average Score | Aesthetic Quality | Background Consistency | Imaging Quality | Motion Smoothness | Subject Consistency | Temporal Flickering |
|---|---|---|---|---|---|---|---|
| Wan | 82.95 | 56.95 | 94.83 | 63.80 | 96.27 | 91.28 | 94.59 |
| Wan + SCMAPR | 85.69 | 59.32 | 96.24 | 67.87 | 98.87 | 93.94 | 97.92 |
Table 4: Quantitative results (%) on T2V-Complexity using VBench metrics with Wan as the T2V backbone.
| T2V Benchmark | Prompts | CS Prompts | CS Proportion | CS Categories |
|---|---|---|---|---|
| VBench | 946 | 311 | 32.9% | 9 |
| EvalCrafter | 700 | 422 | 60.3% | 9 |
| T2V-CompBench | 1400 | 972 | 69.4% | 10 |
| T2V-Complexity | 1000 | 1000 | 100.0% | 10 |
Table 5: Statistics of complex-scenario prompts across different T2V benchmarks. Here, CS is the abbreviation for complex-scenario.
| Method | Average Score |
|---|---|
| SCMAPR | 88.21% |
| w/o Scenario Routing | 86.49% |
| w/o Policy Generation | 87.75% |
| w/o Verification & Self-Correction | 87.63% |
Table 6: Ablation studies of different components in SCMAPR on VBench.
Figure 3: Comparisons of videos generated using Wan ( Wang et al., 2025a ) conditioned on user input and refined prompts from SCMAPR.
Appendix figures & tables16 assets
Supplementary material from the paper’s appendix.
Appendix
| Taxonomy Category | Core Difficulty | Non-overlap Rationale |
|---|---|---|
| Abstract Descriptions | Mapping non-visual, metaphorical, or symbolic language to coherent visual realizations. | Not tied to concrete visual structures; focuses on metaphorical or symbolic interpretation. |
| Complex Spatial Relations | Satisfying explicit geometric constraints (e.g., left, right, between or center) with stable layout, depth perception and occlusion reasoning. | Independent of entity count, interaction, or physics; concerns layout, depth, and occlusion. |
| Multi-Element Scenes | Preserving all salient elements, entity counts, and scene completeness under high visual density without omissions or unintended merging. | Requires stable configurations among multiple entities; does not involve contact or force dynamics. |
| Fine-Grained Appearance | Sustaining high-frequency details such as textures, text, and identity-specific features across frames. | Addresses textures, identity, and small-scale details; orthogonal to spatial, temporal, and physical reasoning. |
| Temporal Consistency | Avoiding temporal drift, flickering, or motion discontinuities in cross-frame evolution. | Focuses on continuity of motion and appearance across frames; unrelated to layout or object count. |
| Stylistic Hybrids | Enforcing coherent visual style under mixed or evolving artistic domains without degradation. | Fully decoupled from spatial, physical, or interaction constraints; concerns aesthetic specification. |
Table A.1: Summary of the ten complex-scenario taxonomy categories, including each category’s core difficulty for text-to-video generation and the rationale used to ensure non-overlapping primary-label annotation.
| Operation | Duration (s) | 1 GPU (%) | 2 GPUs (%) | 4 GPUs (%) | 8 GPUs (%) |
|---|---|---|---|---|---|
| Scenario Routing | 2.2 | 0.44% | 0.85% | 1.42% | 2.11% |
| Policy Generation | 5.9 | 1.19% | 2.28% | 3.82% | 5.65% |
| Prompt Refinement | 2.8 | 0.57% | 1.08% | 1.81% | 2.68% |
| Entailment Verification | 14.7 | 2.97% | 5.68% | 9.51% | 14.07% |
| Content Revision | 1.9 | 0.38% | 0.74% | 1.23% | 1.82% |
| Video Generation | — | 94.44% | 89.36% | 82.20% | 73.68% |
Table A.2: Time cost distribution of each operation under different numbers of GPUs for T2V generation.
| Operation | Duration (s) |
|---|---|
| Video Generation (1 GPU) | 467 |
| Video Generation (2 GPUs) | 231 |
| Video Generation (4 GPUs) | 127 |
| Video Generation (8 GPUs) | 77 |
Table A.3: Time cost of video generation under different numbers of GPUs.
Figure A.1: Examples of hallucination elimination after prompt refinement by SCMAPR.
Figure A.2: Examples of hallucination elimination after prompt refinement by SCMAPR. The hallucinated information is highlighted in red, and its elimination or correction is marked in green.
Figure A.3: Distribution of complex scenarios across VBench.
Figure A.4: Distribution of complex scenarios across EvalCrafter.
Figure A.5: Distribution of complex scenarios across CompBench.
Figure A.6: Distribution of prompt lengths on VBench and EvalCrafter (in words).
Figure A.7: Distribution of prompt lengths on T2V-CompBench and T2V-Complexity (in words).
| VBench | Quantity | Proportion |
|---|---|---|
| rewrite once | 818 | 86.47% |
| rewrite twice | 118 | 12.47% |
| rewrite 3 times | 10 | 1.06% |
| Total | 946 | 100.00% |
Table A.4: Statistics of rewrite rounds on VBench.
| T2V-Complexity | Quantity | Proportion |
|---|---|---|
| rewrite once | 823 | 82.30% |
| rewrite twice | 151 | 15.10% |
| rewrite 3 times | 24 | 2.40% |
| rewrite 4 times | 2 | 0.20% |
| Total | 1000 | 100.00% |
Table A.5: Statistics of rewrite rounds on T2V-Complexity.
Figure A.8: An end-to-end case study of SCMAPR with self-correction. Given a user input, the framework performs scenario routing, policy generation, policy-conditioned prompt refinement, atom-level verification, and targeted revision. Entailment Validator labels each atom-evidence pair and conditionally triggers targeted revision, producing a verified refined prompt for downstream video generation.
Figure A.9: Qualitative comparison of generated videos under an abstract user input (user-provided prompt) and refined prompt. Sampled frames generated from the user-provided prompt (top) and the prompt refined by SCMAPR (bottom) are contrasted. The original prompt yields a visually plausible but semantically drifting dark-water scene, while the refined prompt concretizes the intended metaphor and produces a consistent depiction of a flower bud emerging through cracked dark soil with a warm glow, better matching the target concept.
Figure A.10: Qualitative comparison of generated videos under a user input (user-provided prompt) and refined prompt from T2V-Complexity benchmark.
Figure A.11: Qualitative comparison of generated videos under a user input (user-provided prompt) and refined prompt from T2V-Complexity benchmark.