acceptodds
Under review as a conference paper at ICLR 2027

BlenderLore: Learning 3D Coding from Internet Tutorial Videos

Abstract

Large language models can create 3D assets through code, but still struggle with complex geometry, rigging, and simulation. Blender tutorial videos contain expert workflows that can help address these challenges, but these workflows are difficult to extract and reuse. We introduce BlenderLore, a dataset and pipeline for acquiring and reusing procedural knowledge from tutorial videos for 3D coding. We convert videos into illustrated tutorials, reconstruct assets in Blender Python, and extract reusable procedural knowledge. Adapting and combining methods across tutorials expands the collection to about 22K project instances associated with over 3K videos. We package the reconstruction workflow and procedural knowledge library as BlenderLore Skill for use in existing coding agents. We also introduce \BlenderLore-Bench, a 100-task benchmark covering eight Blender capabilities. Our experiments show that BlenderLore Skill improves the scores of leading large language models and reduces their runtime. Expanding the procedural knowledge library further improves overall performance and time efficiency.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
3D Code Generation, LLM Agents, Datasets and Benchmarks

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_fsn6OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents BlenderLore, which converts Blender tutorial videos into tested, reusable procedures packaged as an agent skill. This pipeline extracts instructions, reconstructs executable assets, and distills methods for adaptation and composition. Main contributions include approximately 22,000 project instances, a knowledge library, and a 100-task benchmark. Across three coding models, the skill improves average rubric scores and reduces task time.

Critical strengths
  1. The paper addresses a timely and important problem given the growing interest in agents for 3D content creation and the scarcity of high-quality procedural 3D code data. Blender tutorial videos are a valuable source of expert workflows, including intermediate operations and implementation details that are often absent from collections of finished assets. Turning these demonstrations into executable reconstructions and reusable procedural knowledge is a meaningful contribution. The substantial reconstruction and expert-validation effort is also a strength.

  2. The extracted procedures provide a practical way to improve existing coding agents without updating their parameters. The reported improvements across the three evaluated models, together with reduced task runtime, suggest that the library provides useful guidance for new 3D tasks. The gains in challenging capabilities such as rigging and simulation are particularly encouraging.

Critical weaknesses
  1. The current experiments demonstrate the usefulness of the complete BlenderLore Skill package, but do not sufficiently isolate the contributions of its individual components. Comparisons against workflow-only guidance, direct retrieval of tutorial text or code examples, and knowledge extracted without executable reconstruction would help establish the value of the proposed construction pipeline. Evaluation beyond the authors’ benchmark and additional agent configurations would also strengthen the evidence for generalization.

  2. Although the collection effort is substantial, the source collection remains relatively modest given the breadth of Blender capabilities the paper aims to cover. Moreover, the approximately 22K project instances include adaptations and combinations of existing methods, while the number of distinct reusable skills is not clearly reported. More detailed statistics on skill diversity, coverage, and redundancy would help establish the library’s effective scale.

Questions for authors
  1. Can you report construction costs, reconstruction success rates, and the amount of human intervention needed to create this dataset?

  2. It seems that the dataset mainly contain atomic skills level demonstrations like Luminous silhouette, Emissive star details, etc. Those skills cover only a basic operation of a complete artist creation process. How well can the model combine those skills and maybe make an asset from beginning?

  3. Do you think the 3d ability of frontier llm just emerge or is trained using some demonstration data like those you made?(Just curious, do not affect my final score)

4 comments
AuthorsOpenReview ↗

Reply to Reviewer fsn6, Part 1

Dear Reviewer fsn6

We sincerely appreciate your insightful feedback and your recognition of our work. Below, we address each concern:

W1: Lack of component ablation.

To directly address your concern, we added an ablation study that isolates the contribution of each component.

Setup. BlenderLore Skill has two parts: generic workflow guidance, refined with human experts, and a procedural knowledge library that the agent retrieves from on demand. We separate the two. Each configuration adds exactly one retrieval resource to the generic workflow guidance: (i) tutorial transcripts, in the form of our illustrated tutorials, where each step transcribed from the raw tutorial video records the narration, key frames, on-screen text, parameters, and intermediate result; (ii) executable code examples, namely the build.py scripts of the 8,918 expert-accepted projects; (iii) unverified procedural knowledge: an early version of our pipeline extracted 21,344 runnable entries from all 3,475 videos in one pass, without reconstruction or expert review; (iv) the procedural knowledge library (BlenderLore Skill). These configurations use Codex with GPT-6 Astra on all 100 tasks, with the same judge as the main experiments. Every resource is mounted the same way, as Markdown files with a manifest that the agent searches, and the agent is told the data structure. Following your suggestion, we also evaluate VIGA [1], with GPT-6 Astra as well, as a representative 3D code agent.

ConfigurationScore ↑Time (min) ↓Tokens In / Out (k) ↓
VIGA [1]86.5112.99153.43 / 12.28
Native Codex (no skill)85.8016.97862.19 / 13.80
Generic workflow guidance only87.9612.10978.8 / 14.69
+ Tutorial transcripts (illustrated)88.8910.951350.2 / 15.56
+ Executable code examples (build.py)89.418.72838.6 / 13.27
+ Unverified procedural knowledge (direct extraction)88.6019.702191.6 / 16.65
+ Procedural knowledge library (BlenderLore Skill)90.9410.581165.57 / 15.55

Findings.

  • VIGA scores higher and runs faster than native Codex, but is weaker than Codex with the expert-refined generic workflow guidance on both. It uses far fewer tokens, because it performs far fewer visual checks.
  • Most of BlenderLore Skill's time savings come from the generic workflow guidance. It alone cuts the runtime of native Codex from 16.97 to 12.10 minutes and raises the score by 2.16 points.
  • Executable code examples help more than tutorial transcripts. Over generic workflow guidance alone, the illustrated transcripts of the raw tutorials raise the score by 0.93 points, and the code examples by 1.45 points. Among the Codex configurations, the code examples are also the fastest and cheapest, because a complete script lets the agent converge quickly.
  • Unverified procedural knowledge does not outperform the high-quality illustrated tutorial transcripts (88.60 vs. 88.89). The unverified set may cover more methods, but it moves the cost of filtering and repair from construction to inference: it takes 19.70 minutes per task, against 10.58 with the library.
  • The procedural knowledge library scores highest, 1.53 points above the code examples. A script is tied to a single asset, whereas a library entry states the general conditions under which a method applies, so it generalizes better. The cost is efficiency: the agent converges more slowly than with code examples (10.58 vs. 8.72 minutes).

We have added this ablation study and its analysis to Sec. 6.4\color{blue}{\text{Sec. 6.4}} of the revised PDF.

AuthorsOpenReview ↗

Reply to Reviewer fsn6, Part 2

W2: Statistics on the distinct reusable skills: scale, coverage, and redundancy.

We thank you for this suggestion. We added descriptive statistics of the procedural knowledge library.

1. Size and coverage. Our procedural knowledge library retains 9,411 distinct entries, covering all eight capabilities of the benchmark:

GEOPROCSURFSCNRIGANMSIMPIPETotal
2,4301,2252,3471,5944404561757449,411

These statistics are now in Sec. 3.1\color{blue}{\text{Sec. 3.1}} of the revised PDF.

2. Redundancy. The loop of reconstruction and expert review yields 13,733 entries. We then run several rounds of heuristic deduplication and merging of similar entries with Codex (GPT-6 Astra), which leaves 9,411. To further address your concern, we embedded each entry (title, Applies when, and Rule, concatenated) with Qwen3-Embedding-0.6B and ranked all entry pairs by cosine similarity. Four human experts examined the 100 most similar pairs and found no redundant pair.

Q1: Construction costs, reconstruction success rates, and the amount of human intervention.

We thank you for this question. Building the dataset cost about USD 40K in API usage and about USD 9K in expert review. We processed 3,475 videos with 9,504 assets; experts accepted 8,918 assets (93.8%) from 3,003 videos (86.4%). Experts helped refine the workflow early on, which is a one-time effort. With the workflow fixed, our data pipeline is highly automated, and experts only accept or reject its outputs. On our annotation platform, an expert compares the result images extracted from the raw tutorial with an interactive 3D view of the asset built by the pipeline, and accepts or rejects in batches. This takes about 30 seconds per asset, about 1% of the total time to produce an asset.

Q2: The dataset seems to contain mainly atomic skills; can the model combine them to make an asset from the beginning?

The unit of our dataset is a complete asset, not an atomic skill. Each of the 8,918 reconstructed assets has a build.py that builds it from an empty scene to the final result. Entries such as Luminous silhouette are the reusable methods we extract from these complete builds.

Agents do combine them into complete assets. Every benchmark task asks for a complete asset built from scratch. We automatically inspected the 100 evaluation trajectories of GPT-6 Astra using Claude Code with Claude Opus 5.5. Per task, retrieval hits 27.1 entries on average, the agent reads 11.4 of them in full, and applies 6.8 in its final solution. These 6.8 entries come from 2.7 different raw tutorials on average, and 87% of the trajectories apply entries from two or more tutorials.

Q3: Does the 3D ability of frontier LLMs just emerge, or is it trained with demonstration data?

Our observations suggest that frontier LLMs such as GPT-6 Astra and the recent Claude Fable 5.2 have been trained on at least some demonstration data. On the assets of some tutorial videos they perform far better than their general level would predict: given only a reference image, without the tutorial, they reproduce the asset almost exactly.

We sincerely thank you for your exceptionally careful and constructive review. We hope that our rebuttal has addressed your concerns, and we would be happy to answer any further questions.

[1] VIGA: Vision-as-Inverse-Graphics Agent via Interleaved Multimodal Reasoning.

Thank you for your additional experiments and clarification. My concerns have been addressed and i will keep my positive rating.

AuthorsOpenReview ↗

Appreciation for the Reviewer’s Feedback

We again thank you for the time you have devoted to improving our manuscript. We will carefully incorporate the above suggestions into the next revision.

Best regards,

Authors

Reviewer_CP8sOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents BlenderLore, a framework and dataset designed to improve 3D code generation using LLMs by mining procedural knowledge from online Blender tutorial videos. The authors convert video tutorials into illustrated step-by-step guides, reconstruct the assets via executable Blender Python scripts, and curate an external, reusable procedural knowledge library. By composing and adapting techniques across tutorials, the dataset expands to ~22K project instances. To evaluate this approach, the paper introduces BlenderLore-Bench, a 100-task benchmark spanning eight core Blender capabilities. Experiments show that plugging this knowledge library (as a lightweight skill) into frontier LLM coding agents consistently boosts generation scores and noticeably speeds up runtime.

Critical strengths

Extracting structured procedural knowledge from real-world multimodal video tutorials bridges a gap in 3D coding. Moving beyond final 3D meshes to capturing intermediate creative workflows makes sense.

The strategy of combining techniques across disparate tutorials (e.g., merging static modeling, character rigging, and procedural shaders) to expand the dataset from ~8.9K to ~22K projects is effective and well-executed.

Equipping agents with the curated procedural skill library leads to solid performance jumps (e.g., +14.4 to +14.8 points for mid-tier models) while noticeably reducing agent reasoning time.

Critical weaknesses

Dataset construction and benchmark rubrics heavily rely on experienced 3D artists for repair, review, and task curation. This raises questions about how well the pipeline can scale purely automatically to new software versions or larger corpora.

The procedural knowledge representation is tightly coupled to Blender's Python API and node systems. It remains unclear how easily this framework could transfer to other DCC tools or CAD software in a zero-shot or few-shot manner (like trained on blender, tested on others)

Questions for authors

See weaknesses.

3 comments
AuthorsOpenReview ↗

Reply to Reviewer CP8s

Dear Reviewer CP8s

We sincerely appreciate your valuable feedback and your positive assessment of our work. Below, we address each concern:

W1: Scalability of the dataset and the benchmark.

Experts helped refine the workflow early on, which is a one-time effort. With the workflow fixed, our data pipeline is highly automated; experts only accept or reject its outputs. On our annotation platform, an expert compares the result images extracted from the raw tutorial with an interactive 3D view of the asset built by the pipeline, and accepts or rejects in batches. This takes about 30 seconds per asset, about 1% of the total time to produce an asset. Expert review is essential to the quality of the reconstructed data and the extracted knowledge, and at this cost it does not limit scale.

Every reconstructed asset has an executable build.py. For a new Blender version, re-running the scripts identifies the failing ones automatically, and only those re-enter the loop. Part of our early pipeline used Blender 4.1; migrating to 5.1.2 required changes to only about 11% of the code. For the benchmark, curating the 100 tasks and their rubrics is a one-time cost, and scoring is automatic.

W2: Transferability to other DCC tools and CAD software.

Our pipeline does not depend on a specific modeling software. BlenderLore Skill consists of generic workflow guidance, refined with human experts, and procedural knowledge bound to the 3D software. The pipeline applies to any software with a scripting interface and renderable results.

We have run the pipeline on other software at a small scale. In fact, among the 3,003 illustrated tutorials, 21 are Maya tutorials, 11 are 3ds Max tutorials, and 7 are CAD tutorials, each with one asset successfully reconstructed by the pipeline. Most of our data uses Blender because it has the strongest open-source ecosystem and the most public tutorials.

We thank you again for your time and valuable feedback. We hope our responses directly address your concerns, and we are happy to provide further clarification or additional experiments if needed.

Thank you for the reply! Most of my concerns are well-addressed so I am happy to maintain my positive rating.

AuthorsOpenReview ↗

Appreciation for the Reviewer’s Feedback

Thank you once again for your kind and constructive feedback. We truly appreciate your support throughout the review process.

Best regards,

Authors

Reviewer_hEvQOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

BlenderLore tranforms Blender tutorial videos into a skill package that can be used by coding agent without training. A benchmark named BlenderLore-Bench is also introduced. Experiments show that BlenderLore improves performances of tested models.

The problem is interesting and practical. The dataset is valuable and provide useful resources for future research. The benchmark evaluates scene structure and functional behavior rather than relying solely on rendered-image similarity, and the reported judge–human agreement provides useful validation.

Critical strengths
  1. The problem is interesting and practical. Tutorial videos contain valuable expertise that is difficult to recover from finished assets alone, particularly for 3D assets creation.
  2. The dataset is valuable and provide useful resources for future research.
  3. The benchmark evaluates scene structure and functional behavior rather than relying solely on rendered-image similarity, and the reported judge–human agreement provides useful validation.
Critical weaknesses
  1. The improvements of BlenderLore is not fully convincing. The main comparison is between naive agents and agents given the BlenderLore knowledge package. However, there are no controls using other resources, such as documentation, raw tutorial transcripts, code examples, or generic workflow guidance. This incomplete comparison is somehow makes the effectivness of BlenderLore less convincing.

  2. The methodological difference between BlenderLore and previous works, such as Resource2Skill, is not clear.

  3. The paper reports approximately 22K project instances, but provides limited quantitative statisitices of the resulting reusable knowledge library such as its size, redundancy, coverage, extraction accuracy, and retrieval effectiveness. The cross-tutorial composition examples are appealing, yet the benchmark does not separately measure adaptation and composition or show when successful solutions require combining methods from multiple tutorials.

Questions for authors
  1. How does BlenderLore compare with equally accessible Blender documentation, tutorial transcripts, and executable code examples under the same agent configuration?
  2. How many distinct knowledge entries are retained, and how are their correctness, redundancy, and applicability assessed?
5 comments
AuthorsOpenReview ↗

Reply to Reviewer hEvQ, Part 1

Dear Reviewer hEvQ

We sincerely appreciate your detailed feedback. Below, we address each concern:

W1/Q1: Lack of component ablation. There are no controls using other resources, such as documentation, tutorial transcripts, code examples, or generic workflow guidance.

We thank you for this insightful suggestion. We added an ablation study that isolates the contribution of each component.

Setup. BlenderLore Skill has two parts: generic workflow guidance, refined with human experts, and a procedural knowledge library that the agent retrieves from on demand. We separate the two. Each configuration adds exactly one retrieval resource to the generic workflow guidance: (i) Blender documentation (the official Manual and Python API reference); (ii) tutorial transcripts, in the form of our illustrated tutorials, where each step transcribed from the raw tutorial video records the narration, key frames, on-screen text, parameters, and intermediate result, a stronger baseline than audio transcripts; (iii) executable code examples, namely the build.py scripts of the 8,918 expert-accepted projects; (iv) the procedural knowledge library (BlenderLore Skill). All configurations use Codex with GPT-6 Astra on all 100 tasks, with the same judge as the main experiments. Every resource is mounted the same way, as Markdown files with a manifest that the agent searches, and the agent is told the data structure.

ConfigurationScore ↑Time (min) ↓Tokens In / Out (k) ↓
Native Codex (no skill)85.8016.97862.19 / 13.80
Generic workflow guidance only87.9612.10978.8 / 14.69
+ Blender documentation87.7511.301065.6 / 15.05
+ Tutorial transcripts (illustrated)88.8910.951350.2 / 15.56
+ Executable code examples (build.py)89.418.72838.6 / 13.27
+ Procedural knowledge library (BlenderLore Skill)90.9410.581165.57 / 15.55

Findings.

  • Most of the time savings come from the generic workflow guidance. It alone cuts runtime from 16.97 to 12.10 minutes and raises the score by 2.16 points.
  • Blender documentation does not help; tutorial transcripts do. Compared with generic workflow guidance alone, the illustrated transcripts of the raw tutorials raise the score by 0.93 points and save 1.15 minutes per task.
  • Executable code examples help more. They raise the score by 1.45 points over generic workflow guidance alone. This is also the fastest and cheapest configuration, because a complete script lets the agent converge quickly.
  • The procedural knowledge library scores highest, 1.53 points above the code examples. A script is tied to a single asset, whereas a library entry states the general conditions under which a method applies, so it generalizes better. The cost is efficiency: the agent converges more slowly than with code examples (10.58 vs. 8.72 minutes).

We have added the above results and analysis to Sec. 6.4\color{blue}{\text{Sec. 6.4}} of the revised PDF.

AuthorsOpenReview ↗

Reply to Reviewer hEvQ, Part 2

W2: The methodological difference between BlenderLore and previous works, such as Resource2Skill (R2S), is not clear.

To directly address your concern: BlenderLore verifies each method in a closed loop of executable reconstruction and expert review before it enters the library. Specifically, BlenderLore repairs or reconstructs each tutorial's asset from an illustrated tutorial transcribed from the raw video. Only after experts confirm that the implementation reproduces the tutorial's result do we extract procedural knowledge from the illustrated tutorial and its build.py. By contrast, R2S generates a skill entry from each human resource with one LLM call and accepts it once it passes structural checks and its code runs.

To show the benefit of the closed loop, we ran a comparison. An early version of our pipeline followed a similar direct-extraction approach: it extracted 21,344 runnable procedural knowledge entries from all 3,475 videos in one pass, without reconstruction or expert review. The current BlenderLore Skill holds 9,411 entries from 3,003 videos, after expert review, deduplication, and merging of similar entries. The setup is the same as in W1.

ConfigurationScore ↑Time (min) ↓Tokens In / Out (k) ↓
+ Tutorial transcripts (illustrated)88.8910.951350.2 / 15.56
+ Unverified procedural knowledge (direct extraction)88.6019.702191.6 / 16.65
+ Procedural knowledge library (BlenderLore Skill)90.9410.581165.57 / 15.55

Directly extracted procedural knowledge does not outperform the high-quality illustrated tutorial transcripts (88.60 vs. 88.89). The smaller, verified library scores higher (90.94) and is also more efficient. The unverified set may cover more methods, but it moves the cost of filtering and repair from construction to inference: it takes 19.70 minutes per task, against 10.58 with the library.

We thank you for pointing us to this work. We appreciate its contribution to this area, and we now cite it in the Introduction lines 44-47\color{blue}{\text{lines 44-47}} and state how BlenderLore differs.

W3/Q2: Limited statistics on the knowledge library (size, redundancy, coverage, extraction accuracy, retrieval effectiveness); adaptation and composition are not measured separately.

We thank you for these valuable suggestions. We respond to each concern in turn.

1. Size and coverage. Our procedural knowledge library retains 9,411 distinct entries, covering all eight capabilities of the benchmark:

GEOPROCSURFSCNRIGANMSIMPIPETotal
2,4301,2252,3471,5944404561757449,411

These statistics are now in Sec. 3.1\color{blue}{\text{Sec. 3.1}} of the revised PDF.

2. Redundancy. The loop of reconstruction and expert review yields 13,733 entries. We then run several rounds of heuristic deduplication and merging of similar entries with Codex (GPT-6 Astra), which leaves 9,411. To further address your concern, we embedded each entry (title, Applies when, and Rule, concatenated) with Qwen3-Embedding-0.6B and ranked all entry pairs by cosine similarity. Four human experts examined the 100 most similar pairs and found no redundant pair.

3. Correctness. We sampled 100 entries, proportionally across the eight capabilities. Four human experts compared each entry with its source tutorial and labeled the extraction as accurate, imprecise, or wrong: 92 are accurate, 7 are imprecise, and 1 is wrong.

4. Retrieval effectiveness. We automatically inspected the 100 evaluation trajectories of GPT-6 Astra using Claude Code with Claude Opus 5.5. Per task, retrieval hits 27.1 entries on average, the agent reads 11.4 of them in full, and applies 6.8 in its final solution.

5. Adaptation and composition. No benchmark task comes from a source tutorial of the library: the benchmark is designed to test how the procedural knowledge library adapts to entirely new tasks. In addition, the same trajectory analysis shows that the 6.8 entries applied per task come from 2.7 different raw tutorials on average, and 87% of the trajectories apply entries from two or more tutorials.

Automatic extraction can still introduce a small number of errors, but they do not prevent the library from delivering the overall gains reported above. Reducing these errors with further expert-in-the-loop review is our next step.

We sincerely thank you again for these constructive comments, which have greatly improved the precision and rigor of our work. We remain available to address any further questions.

AuthorsOpenReview ↗

Looking forward to further discussion

Dear Reviewer hEvQ,

We sincerely thank you for your thorough and insightful review. We have carefully addressed your concerns and provided additional experimental results. We sincerely hope these results directly respond to the issues you raised.

With the discussion period nearing its end, we wanted to kindly ask whether you have any remaining questions or whether our responses have sufficiently addressed your concerns. We would be grateful for any further feedback.

Thank you again for your time and valuable contribution to the review process.

Best regards,

Authors

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.