arxiv: 2605.00632 · v1 · submitted 2026-05-01 · 💻 cs.CV · cs.AI· cs.GR· cs.HC· cs.LG

Recognition: unknown

BlenderRAG: High-Fidelity 3D Object Generation via Retrieval-Augmented Code Synthesis

Massimo Rondelli , Francesco Pivi , Maurizio Gabbrielli

Authors on Pith no claims yet

Pith reviewed 2026-05-09 20:12 UTC · model grok-4.3

classification 💻 cs.CV cs.AIcs.GRcs.HCcs.LG

keywords 3D object generationretrieval-augmented generationBlender Python scriptinglarge language modelscode synthesissemantic alignmentmultimodal datasettext-to-3D

0 comments

The pith

Retrieving similar examples lets LLMs generate executable Blender code for 3D objects much more reliably.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper presents a retrieval-augmented system that pulls semantically matching examples from a fixed dataset of 500 expert-validated text-code-image triples spanning 50 object categories. Direct prompting of large language models for Blender Python scripts produces frequent syntax failures and geometrically wrong objects, but inserting retrieved examples during generation raises the fraction of code that compiles from 40.8 percent to 70 percent. The same change lifts measured semantic match to the input description from 0.41 to 0.77 on a normalized CLIP similarity score. The improvement holds across four different LLMs and requires no weight updates or extra hardware. A reader cares because the result turns an unreliable text-to-3D pipeline into one that works on ordinary computers using only an existing model plus a modest curated store.

Core claim

BlenderRAG shows that semantic retrieval from a curated multimodal collection of 500 examples across 50 categories supplies the missing context that lets current large language models output Blender scripts whose syntax succeeds at compile time and whose rendered geometry aligns with the original natural-language request.

What carries the argument

Retrieval-augmented generation that selects and inserts the most similar text-code-image examples from the 500-example store before the LLM produces the final Blender script.

If this is right

Blender scripts produced this way run without syntax errors at more than twice the baseline rate.
The generated objects match the semantic intent of the text prompt at substantially higher measured similarity.
The gains appear for every one of the four LLMs tested and need no model retraining.
Deployment uses only standard hardware and the existing language model plus the fixed dataset.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The same retrieval pattern could be tried in other domains where high-quality code examples are limited but expensive to obtain.
Growing the dataset beyond 50 categories would test whether performance scales to truly open-ended object descriptions.
Pairing the retrieval step with a subsequent self-correction loop might push success rates still higher while staying training-free.

Load-bearing premise

The 500 examples across 50 categories are representative enough that retrieved code snippets will consistently help the model rather than introduce new errors on unseen descriptions.

What would settle it

Apply the system to object descriptions drawn from categories absent from the original 50 and check whether compilation success and CLIP alignment fall back to the 40.8 percent and 0.41 levels seen without retrieval.

Figures

Figures reproduced from arXiv: 2605.00632 by Francesco Pivi, Massimo Rondelli, Maurizio Gabbrielli.

**Figure 2.** Figure 2: BlenderRAG architecture: user queries (text/image) are embedded and matched against the Qdrant vector database. Retrieved [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗

**Figure 4.** Figure 4: Code length distribution for outdoor objects. Red [PITH_FULL_IMAGE:figures/full_fig_p002_4.png] view at source ↗

**Figure 5.** Figure 5: BlenderRAG add-on interface integrated in Blender, show [PITH_FULL_IMAGE:figures/full_fig_p003_5.png] view at source ↗

read the original abstract

Automatic generation of executable Blender code from natural language remains challenging, with state-of-the-art LLMs producing frequent syntactic errors and geometrically inconsistent objects. We present BlenderRAG, a retrieval-augmented generation system that operates on a curated multimodal dataset of 500 expert-validated examples (text, code, image) across 50 object categories. By retrieving semantically similar examples during generation, BlenderRAG improves compilation success rates from 40.8% to 70.0% and semantic normalized alignment from 0.41 to 0.77 (CLIP similarity) across four state-of-the-art LLMs, without requiring fine-tuning or specialized hardware, making it immediately accessible for deployment. The dataset and code will be available at https://github.com/MaxRondelli/BlenderRAG.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

BlenderRAG gets solid lifts in Blender code compilation and alignment from a small curated retrieval set, but the write-up skips too many setup details to judge how far the gains travel.

read the letter

BlenderRAG applies retrieval-augmented generation to turn natural language into executable Blender Python scripts. The main new element is a 500-example multimodal dataset of text, code, and images across 50 object categories, all marked as expert-validated. Feeding the LLM the closest retrieved examples raises compilation success from 40.8% to 70% and normalized CLIP alignment from 0.41 to 0.77, and the same pattern appears across four different LLMs. No fine-tuning or extra hardware is required, and the authors plan to release both the dataset and the code. That combination of a public artifact and a zero-training recipe is the practical hook here.

Referee Report

4 major / 2 minor

Summary. The manuscript proposes BlenderRAG, a retrieval-augmented generation system for synthesizing executable Blender Python code from natural language prompts. It operates on a curated multimodal dataset of 500 expert-validated examples (text, code, image) across 50 object categories. By retrieving semantically similar examples at inference time, the approach claims to raise compilation success rates from 40.8% to 70.0% and semantic normalized alignment (CLIP similarity) from 0.41 to 0.77 across four state-of-the-art LLMs, without fine-tuning or specialized hardware. The dataset and code are promised to be released publicly.

Significance. If the reported gains prove robust, BlenderRAG would demonstrate a practical, low-resource route to higher-fidelity text-to-3D code synthesis by substituting retrieval for model adaptation. The explicit commitment to release the 500-example dataset and accompanying code is a clear strength that supports reproducibility and community follow-up. The work targets a concrete pain point in procedural 3D modeling and could be immediately usable by practitioners.

major comments (4)

Abstract and §4 (Experimental Evaluation): The baseline figures (40.8% compilation success, 0.41 CLIP alignment) are presented without any description of the prompting strategy, system prompts, temperature settings, or exact LLM configurations used in the non-RAG condition, preventing isolation of the retrieval contribution.
§3.1 (Dataset) and §4.1 (Setup): The construction, expert-validation protocol, category coverage, and train/test split procedure for the 500-example dataset are described only at high level; no information is given on how held-out prompts were chosen or how leakage was prevented, directly affecting the representativeness assumption underlying the retrieval gains.
§4.2 (Results): The performance deltas are reported as aggregate numbers across four LLMs with no per-model tables, standard deviations across runs, or statistical significance tests, so it is impossible to judge whether the jump from 40.8% to 70.0% is reliable or sensitive to random seeds.
Method section: The retrieval implementation (embedding model, similarity metric, top-k value, and any filtering for syntactic validity) is not specified, making the 70% success rate unreproducible from the given description.

minor comments (2)

Abstract: The phrase 'semantic normalized alignment' is used without a precise definition or formula for the normalization step applied to CLIP similarity.
Related Work: Additional citations to recent RAG-for-code papers and existing text-to-Blender efforts would better situate the contribution.

Simulated Author's Rebuttal

4 responses · 0 unresolved

We thank the referee for the constructive and detailed feedback. We have revised the manuscript to address all major comments by expanding descriptions, adding missing details, and including additional analyses to improve clarity and reproducibility.

read point-by-point responses

Referee: [—] Abstract and §4 (Experimental Evaluation): The baseline figures (40.8% compilation success, 0.41 CLIP alignment) are presented without any description of the prompting strategy, system prompts, temperature settings, or exact LLM configurations used in the non-RAG condition, preventing isolation of the retrieval contribution.

Authors: We agree that the baseline prompting details are necessary to isolate the retrieval contribution. In the revised manuscript, we have added a dedicated paragraph in §4 describing the exact system prompts, temperature settings (0.7 for all models), and LLM configurations (including model versions and API parameters) used for the non-RAG baselines. revision: yes
Referee: [—] §3.1 (Dataset) and §4.1 (Setup): The construction, expert-validation protocol, category coverage, and train/test split procedure for the 500-example dataset are described only at high level; no information is given on how held-out prompts were chosen or how leakage was prevented, directly affecting the representativeness assumption underlying the retrieval gains.

Authors: We acknowledge that the dataset details require expansion. The revised §3.1 now includes the full expert-validation protocol (three independent Blender experts with majority agreement), breakdown of the 50 categories, the 80/20 train/test split procedure, selection of held-out prompts, and leakage prevention via embedding-based semantic dissimilarity checks (threshold of 0.85 cosine similarity) between train and test sets. revision: yes
Referee: [—] §4.2 (Results): The performance deltas are reported as aggregate numbers across four LLMs with no per-model tables, standard deviations across runs, or statistical significance tests, so it is impossible to judge whether the jump from 40.8% to 70.0% is reliable or sensitive to random seeds.

Authors: We agree that aggregate reporting alone is insufficient. We have added a per-model results table in §4.2, along with means and standard deviations computed over five independent runs with different random seeds for each LLM. We also report p-values from paired t-tests demonstrating statistical significance (p < 0.01) of the observed improvements. revision: yes
Referee: [—] Method section: The retrieval implementation (embedding model, similarity metric, top-k value, and any filtering for syntactic validity) is not specified, making the 70% success rate unreproducible from the given description.

Authors: We thank the referee for highlighting this reproducibility gap. The Method section has been updated to fully specify the retrieval pipeline: CLIP ViT-B/32 as the embedding model, cosine similarity metric, top-k=5, and a syntactic validity filter using Python's AST module to exclude invalid code before in-context augmentation. revision: yes

Circularity Check

0 steps flagged

No significant circularity

full rationale

The paper describes an empirical RAG system for Blender code generation and reports measured performance gains (compilation success 40.8% → 70.0%, CLIP alignment 0.41 → 0.77) on held-out prompts using a fixed curated dataset of 500 examples. No derivation chain, equations, fitted parameters renamed as predictions, or self-citation load-bearing steps are present; the central claims are direct experimental deltas that do not reduce to the inputs by construction.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on the quality and coverage of the 500-example dataset and the assumption that semantic retrieval will surface useful code examples; no free parameters are fitted to data and no new entities are postulated.

axioms (1)

domain assumption Semantic similarity search over a curated set of text-code-image triples reliably identifies examples that improve downstream code generation quality
This assumption underpins the entire retrieval step and is not independently verified in the abstract.

pith-pipeline@v0.9.0 · 5451 in / 1331 out tokens · 65127 ms · 2026-05-09T20:12:14.010954+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

14 extracted references · 5 canonical work pages · 4 internal anchors

[1]

Claude opus 4.1 system card addendum

[Anthropic, 2025] Anthropic. Claude opus 4.1 system card addendum. Technical report, Anthropic,

2025
[2]

Program Synthesis with Large Language Models

[Austinet al., 2021 ] Jacob Austin, Augustus Odena, Maxwell Nye, Maarten Bosma, Henryk Michalewski, David Dohan, Ellen Jiang, Carrie Cai, Michael Terry, Quoc Le, et al. Program synthesis with large language models.arXiv preprint arXiv:2108.07732,

work page internal anchor Pith review Pith/arXiv arXiv 2021
[3]

Improving language models by retrieving from trillions of tokens

[Borgeaudet al., 2022 ] Sebastian Borgeaud, Arthur Men- sch, Jordan Hoffmann, Trevor Cai, Eliza Rutherford, Katie Millican, George Bm Van Den Driessche, Jean-Baptiste Lespiau, Bogdan Damoc, Aidan Clark, et al. Improving language models by retrieving from trillions of tokens. InInternational conference on machine learning, pages 2206–2240. PMLR,

2022
[4]

Evaluating Large Language Models Trained on Code

[Chen, 2021] Mark Chen. Evaluating large language models trained on code.arXiv preprint arXiv:2107.03374,

work page internal anchor Pith review Pith/arXiv arXiv 2021
[5]

Blenderllm: Training large language models for computer-aided design with self-improvement.arXiv preprint arXiv:2412.14203, 2024

[Duet al., 2024 ] Yuhao Du, Shunian Chen, Wenbo Zan, Peizhao Li, Mingxuan Wang, Dingjie Song, Bo Li, Yan Hu, and Benyou Wang. Blenderllm: Training large language models for computer-aided design with self- improvement.arXiv preprint arXiv:2412.14203,

work page arXiv 2024
[6]

Gemini 3 flash: Benchmarks and global availability

[Google DeepMind, 2026] Google DeepMind. Gemini 3 flash: Benchmarks and global availability. https://blog. google/products/gemini/gemini-3-flash/,

2026
[7]

[Guuet al., 2020 ] Kelvin Guu, Kenton Lee, Zora Tung, Panupong Pasupat, and Mingwei Chang

Accessed: 2026-02-08. [Guuet al., 2020 ] Kelvin Guu, Kenton Lee, Zora Tung, Panupong Pasupat, and Mingwei Chang. Retrieval aug- mented language model pre-training. InInternational con- ference on machine learning, pages 3929–3938. PMLR,

2026
[8]

Retrieval-augmented generation for knowledge-intensive nlp tasks.Advances in neural infor- mation processing systems, 33:9459–9474,

[Lewiset al., 2020 ] Patrick Lewis, Ethan Perez, Aleksan- dra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich K ¨uttler, Mike Lewis, Wen-tau Yih, Tim Rockt¨aschel, et al. Retrieval-augmented generation for knowledge-intensive nlp tasks.Advances in neural infor- mation processing systems, 33:9459–9474,

2020
[9]

StarCoder: may the source be with you!

[Liet al., 2023 ] Raymond Li, Loubna Ben Allal, Yangtian Zi, Niklas Muennighoff, Denis Kocetkov, Chenghao Mou, Marc Marone, Christopher Akiki, Jia Li, Jenny Chim, et al. Starcoder: may the source be with you!arXiv preprint arXiv:2305.06161,

work page internal anchor Pith review arXiv 2023
[10]

Introducing mistral

[Mistral AI, 2025] Mistral AI. Introducing mistral

2025
[11]

[OpenAI, 2025] OpenAI

Accessed: 2026-02-08. [OpenAI, 2025] OpenAI. GPT-5 system card. Technical re- port, OpenAI,

2026
[12]

Code Llama: Open Foundation Models for Code

[Roziereet al., 2023 ] Baptiste Roziere, Jonas Gehring, Fabian Gloeckle, Sten Sootla, Itai Gat, Xiaoqing Ellen Tan, Yossi Adi, Jingyu Liu, Romain Sauvestre, Tal Remez, et al. Code llama: Open foundation models for code. arXiv preprint arXiv:2308.12950,

work page internal anchor Pith review arXiv 2023
[13]

3d-gpt: Procedural 3d modeling with large language models

[Sunet al., 2025 ] Chunyi Sun, Junlin Han, Weijian Deng, Xinlong Wang, Zishan Qin, and Stephen Gould. 3d-gpt: Procedural 3d modeling with large language models. In 2025 International Conference on 3D Vision (3DV), pages 1253–1263. IEEE,

2025
[14]

Scenecraft: Layout-guided 3d scene generation.Advances in Neural Information Processing Systems, 37:82060–82084, 2024

[Yanget al., 2024 ] Xiuyu Yang, Yunze Man, Junkun Chen, and Yu-Xiong Wang. Scenecraft: Layout-guided 3d scene generation.Advances in Neural Information Processing Systems, 37:82060–82084, 2024

2024