cs.CVNov 5, 2026

ManimCode: Learning Visual Self-Correction for Educational Animation

Authors: Zhengyang Liang, Lizi Liao

Organizations: Singapore Management University

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Manim supports precise, programmatic video generation, but a successfully rendered educational animation can still have visual defects, missing explanations, or incorrect content. Rendering does not directly identify the code changes needed to correct these visual failures, so feedback-driven revision is needed. Existing agentic frameworks use iterative rendering and revision but typically rely on large models, often proprietary. Existing Manim datasets primarily contain final programs and often lack quality verification and the trajectories of feedback and revisions needed to learn this refinement process. We introduce ManimData and ManimCode, a data and training framework for feedback-driven educational animation generation. We construct ManimData through a render-guided pipeline that iteratively generates, executes, reviews, and revises Manim programs with retrieved API documentation. Its 24,034 render-verified, quality-filtered records preserve intermediate programs, execution errors, visual critiques, and corrective revisions. We use role-conditioned supervised fine-tuning on these trajectories to train ManimCode, a single 4B model, to write, review, and revise animation programs. To improve correction of errors that differ from those in the collected trajectories, we then train the Reviser with reinforcement learning on self-generated states, using rendering outcomes and reference-free quality feedback. ManimCode achieves 96.6% and 97.0% render success on ManimData-test and ManimBench, compared with 52.6% and 82.0% for the untrained 4B model under the same inference setup, while also improving video quality scores from 2.48 to 2.94 and from 3.00 to 3.44. Integrating ManimCode into two long-form educational video generation systems increases completed videos from 19/117 to 107/117 and from 53/239 to 239/239, alongside improvements in educational effectiveness and visual quality.

Figures & tables

DatasetSizePurposeTarget VideoRenderQualityRevision
VerifiedFilteredTrajectories
ManimBench ( Rammuni Silva et al., 2026a )416Train/EvalScene-level Animation✓✗✗
JanusCode-800K (Manim) ( Sun et al., 2026 )8,316TrainScene-level Animation✗✗✗
HF Community (7 datasets)20,976TrainScene-level Animation✗✗✗
MMMC (Code2Video) ( Chen et al., 2025b )117EvalLong-form Explanationn/an/a✗
TheoremExplainBench ( Ku et al., 2025 )239EvalLong-form Explanationn/an/a✗
Table 1: Comparison with existing Manim datasets and benchmarks. n/a marks a property that does not apply to the evaluation-only long-form benchmarks.
Figure 1: ManimData construction pipeline. (1) Request collection: Wikipedia STEM requests, community programs, and the API knowledge base. (2) Render-guided generation: Qwen3.8-27B writes or revises the program from retrieved API cards, and each version is rendered, checked for layout defects, and reviewed on eight frames; an error or a critique (a crash and an overlapping label) starts the next revision. (3) Curation: an input whose best version reaches a mean review score of at least 3.75 is admitted with its code, video, and trajectory.
Figure 2: ManimData statistics. Top: domain shares of the 24,034 records. Bottom: per-record distributions of final program length, video length, and number of renders.
Figure 3: Overview of ManimCode. (1) Final programs and intermediate steps of ManimData trajectories supervise five tasks of one Qwen3.5-4B model. (2) Reinforcement learning further trains the Reviser on repair and critique states that the supervised model produces itself: the Reviser samples eight revisions per state, Qwen3.8-27B scores each rendered revision (request coverage for repair, two order-swapped comparisons with the draft for critique), and revisions that score above the rest of their group become more likely (Sec. 4.2 ). The two states shown are real training states; their revisions and scores are illustrative. (3) At inference, our model runs the render–review–revise loop. Qwen3.8-27B is used only to build the data and to compute training rewards.
  ManimData-testManimBench
ModelSizeRS ↑\uparrowVQ ↑\uparrowRS ↑\uparrowVS ↑\uparrowVQ ↑\uparrow
Large general-purpose models
GPT-4o–60.22.8779.071.03.34
GLM-5.3 FlashMoE57.23.1883.073.73.51
DeepSeek-V4 FlashMoE62.43.0581.072.93.33
Small general-purpose models
Table 2: Scene-level animation generation on ManimData-test and ManimBench. VQ is averaged over successfully rendered videos. † 71 of the 100 ManimBench requests appear in JanusCoder’s training data.
Metric (higher is better)Native+ Ours
Code2Video on MMMC
Videos completed19/117107/117
TeachQuiz gain ( ×\times 100)−- 2.627.2
Aesthetics (0–100)35.558.7
TheoremExplainAgent on TEB
Videos completed53/239239/239
Table 3: Scene-module substitution in long-form systems. Both runs of a system share one lecture plan written by the untrained Qwen3.5-4B. Native writes and repairs the scenes with the same untrained model, and + Ours generates them with our system.
Figure 7
Figure 5: One revision step of our system on ManimData-test. Our Reviewer flags that the title overlaps the table header; the Reviser moves the table down and places the current-state label below it, and the revised video is kept as the final output. Dashed boxes mark the defect (red) and the repaired region (green). Under each frame are our Reviewer’s mean score and Gemini 3.8 Flash score. More examples are in Appendix C.5 .
Appendix figures & tables7 assets

Supplementary material from the paper’s appendix.

Appendix

Table 9
SettingValue
RendererManimCE v0.19.1 with eight domain plugins, 480p
Budget10 renders per input
WriterQwen3.8-27B; writes the full program from retrieved cards and feedback
Retrieval2–4 queries per step; Qwen3-Embedding-8B search, Qwen3-Reranker-8B reranking; top 10 cards
Layout checksObjects outside the frame, overlapping text, labels too small to read, at animation pause points
ReviewerQwen3.8-27B; request, code, and 8 uniformly sampled frames; four dimensions scored 1–5
Appendix
Table 7: Construction loop settings. Qwen3.8-27B plays every role. The Writer row also applies to the Reviser, which rewrites the full program from the error or the critique. Acceptance decides when a candidate replaces the best candidate c∗c^{*} , and Admission decides whether a record enters ManimData.
SupervisionInput contextTarget outputExamples
WritingRequest; retrieved API cardsFinal program selected during construction24,034
Execution repairRequest; failed code; execution error; API cardsRenderable repaired program14,587
Quality revisionRequest; code; scores and critique; layout diagnostics; API cardsHigher-rated revised program32,683
ReviewingRequest; code; sampled video framesScores on the four review dimensions and a critique46,965
Retrieval queriesRequest; error or critique when presentTwo to four API search queries187,183
Total  305,452
Appendix
Table 8: Role-conditioned supervision. Input context, target output, and number of examples for each supervision type extracted from the 24,034 ManimData trajectories.
Supervised fine-tuningRevision RL
Base modelQwen3.5-4BInitializationsupervised model
LoRArank 32, α=64\alpha=64 , dropout 0.05LoRArank 32, α=64\alpha=64
Target modulesall linear layersAdvantageleave-one-out (RLOO)
Learning rate1×10−41\times 10^{-4} , cosineLearning rate1×10−41\times 10^{-4}
Warm-up10 stepsBatch16 states ×\times G=8G=8 samples
Batch size16 sequencesPolicy updates2 per step (8 states each)
Appendix
Table 9: Training hyperparameters. Both stages train LoRA adapters [ Hu et al., 2021 ] ; the RL adapter starts from the merged supervised model.
Gemini VQ gapPairsGemini’s higherAbout the sameGemini’s lowerAgreement (%)
0 (equal scores)14–6––
(0,0.25)(0,0.25)19451029 [7, 50]
[0.25,0.5)[0.25,0.5)26155671 [52, 90]
[0.5,1)[0.5,1)31223679 [61, 93]
≥1\geq 1852183 [50, 100]
All9846212367 [55, 78]
Appendix
Table 10: Human choices by VQ gap. The gap is the absolute difference between the VQ scores that the judge (Gemini) gave the two videos of a pair (98 pairs, one annotator). Among pairs in which both the annotator and the judge chose a side, agreement is the share in which the annotator picked the video the judge scored higher (95% bootstrap interval in brackets). In the 14 pairs with equal judge scores, the annotator answered about the same in 6 and chose a side in 8; in the All row, those 8 count only in the Pairs column.
OpponentBenchmarkPairsHuman W / T / LHuman win rateGemini W / T / L
GPT-4oManimData-test3118 / 4 / 90.65 [0.48, 0.81]17 / 7 / 7
GPT-4oManimBench177 / 7 / 30.62 [0.44, 0.79]9 / 3 / 5
Qwen3.8-27BManimData-test3611 / 2 / 230.33 [0.19, 0.49]14 / 3 / 19
Qwen3.8-27BManimBench145 / 8 / 10.64 [0.50, 0.79]6 / 1 / 7
Appendix
Table 11: Human and judge (Gemini) verdicts on the same pairs. Wins, ties, and losses of our system against each opponent; for the judge, a tie means equal VQ scores. The human win rate counts a tie as half (95% bootstrap interval in brackets).
Figure 6: More revision steps of our system on ManimData-test. Each row is one revision step; the best-candidate rule decides which version the loop returns. From the top: a crash repaired from the renderer’s error message; a wrong sum corrected in the explanation text; an overlap removed by moving the result below the derivation; and two consecutive revisions of the same request. In the last two rows, moving the transaction list down clears the progress bar but pushes the list past the bottom edge, and moving it back up restores the overlap; our Reviewer scores the second revision lower, so the loop keeps the first. Under each frame are our Reviewer’s mean score and the score that the judge, Gemini 3.8 Flash, gives the re-rendered video (1–5).
More in this topic

Video Generation

Video generation studies models that create and edit moving images from text, images, audio, and structured controls.

See all 376 papers

Explore similar work

CardsList
May 15, 2026

See Before You Code: Learning Visual Priors for Spatially Aware Educational Animation Generation

Large language models can generate executable code for educational animations, but the resulting renders often exhibit visual defects, including element overlap, misalignment, and broken animation continuity. These defects cannot be reliably detected from the code alone and become apparent only after execution. We formalize this problem as render-feedback-aware constrained code generation: given a natural language specification, the model must generate executable code whose rendered output satisfies structured quality criteria that can be evaluated only after rendering. To address this problem, we introduce OmniManim, a render-feedback-aware educational animation generation framework built around a shared scene state, explicit visual planning, structured post-render diagnostics, and localized repair. Within OmniManim, the Vision Agent is a task-specific visual planning module: it predicts sparse keyframe layouts with coarse-to-fine bounding-box denoising and optimizes an interpolation-aware objective to reduce intermediate-frame failures induced by downstream animation interpolation. We further construct two datasets, ManimLayout-1K and EduRequire-500, and provide a reproducible evaluation protocol covering executability, instructional quality, visual quality, and efficiency. On EduRequire-500, OmniManim improves measured render quality over both single-model baselines and existing multi-agent frameworks. Systematic ablation studies further verify that explicit visual planning, especially its coarse spatial prior, bounding-box refinement, and interpolation-aware optimization, is central to these gains.

Yuejia Li, Ke He, Junheng Li +5

Wuhan University · University of Chinese Academy of Sciences

Execution-Guided Code GenerationCode GenerationGenerative AI in EducationVideo Generation
AskPDF
Apr 20, 2026

Training and Agentic Inference Strategies for LLM-based Manim Animation Generation

Generating programmatic animation using libraries such as Manim presents unique challenges for Large Language Models (LLMs), requiring spatial reasoning, temporal sequencing, and familiarity with domain-specific APIs that are underrepresented in general pre-training data. A systematic study of how training and inference strategies interact in this setting is lacking in current research. This study introduces ManimTrainer, a training pipeline that combines Supervised Fine-tuning (SFT) with Reinforcement Learning (RL) based Group Relative Policy Optimisation (GRPO) using a unified reward signal that fuses code and visual assessment signals, and ManimAgent, an inference pipeline featuring Renderer-in-the-loop (RITL) and API documentation-augmented RITL (RITL-DOC) strategies. Using these techniques, this study presents the first unified training and inference study for text-to-code-to-video transformation with Manim. It evaluates 17 open-source sub-30B LLMs across nine combinations of training and inference strategies using ManimBench. Results show that SFT generally improves code quality, while GRPO enhances visual outputs and increases the models' responsiveness to extrinsic signals during self-correction at inference time. The Qwen 3 Coder 30B model with GRPO and RITL-DOC achieved the highest overall performance, with a 94% Render Success Rate (RSR) and 85.7% Visual Similarity (VS) to reference videos, surpassing the baseline GPT-4.1 model by +3 percentage points in VS. Additionally, the analysis shows that the correlation between code and visual metrics strengthens with SFT and GRPO but weakens with inference-time enhancements, highlighting the complementary roles of training and agentic inference strategies in Manim animation generation.

Ravidu Suien Rammuni Silva, Ahmad Lotfi, Isibor Kennedy Ihianle +2

Department of Computer Science, Nottingham Trent University, Nottingham, United Kingdom · Department of Physics and Mathematics, Nottingham Trent University, Nottingham, United Kingdom

RL for Code GenerationT2V GenerationLLM Agent Training
AskPDF
Jun 29, 2026

ManimAgent: Self-Evolving Multimodal Agents for Visual Education

Multi-round reflection lets agents built on large language models recover from failures within a single task, but each task remains an isolated episode: lessons learned across many reflection rounds on one task are discarded before the next begins. We study this gap on a code-generation task: from a scientific paper section, the agent writes Python in the open-source Manim library to render a mathematical animation. We present ManimAgent, a self-evolving multimodal agent that carries reflection experience across tasks through a dual-channel Episodic Memory Bank grown entirely from its own task stream, with no weight updates and no human seeds. After each animation converges, a vision-language model scores the rendered keyframes; the resulting signals populate a positive channel M+ that stores success rationales as soft Reference Examples, and a negative channel M- that stores validated failure patterns as hard Known Pitfalls. On a fixed-probe evaluation against no-memory, matched-budget retrieval-augmented generation, and shuffled-memory baselines, blind human Pass@1 rises and reflection rounds fall as memory size grows.

Wenjia Jiang, Zongyuan Cai, Yuanhang Shao +7

University of Alberta · Southeast University · Virginia Tech +2

Continual Learning for LLM AgentsLLM Agent Self-ImprovementAgentic Code GenerationMultimodal AgentsMemory-Augmented Language Model AgentsSelf-Reflection in Language Model Agents
AskPDF