BlenderWorld

BlenderWorld is the first comprehensive benchmark dedicated to evaluating frontier models on realistic 3D creation tasks. BlenderWorld comprises 223 tasks spanning architecture, products, natural environments, and scientific visualization, driven by text, image, video, and structured specification inputs. To rigorously assess these complex deliverables, a unified evaluation protocol is established: The submitted Blender projects are rendered from multiple viewpoints and evaluated against 2,651 task-specific binary rubrics alongside holistic visual quality scores. For more details, please refer to our Project: BlenderWorld.

BlenderWorld-Harbor

The benchmark comprises 223 tasks and 2,651 rubric criteria. Each instance includes a Base64-encoded image preview and corresponding JSON data.

The package contains 78 text, 130 image-and-text, and 15 video-and-text tasks. Use Harbor 0.23.0, Python 3.12+ and a GPU-capable Linux sandbox. The commands below use the Modal backend. 中文使用说明

Task categories

The 223 tasks span 9 categories and 22 subcategories. Use taxonomy.csv for the per-task mapping, taxonomy.json for the hierarchy, counts and figure labels, or classification_summary.md for a readable table. Each task also includes category, subcategory, category_short, subcategory_short and taxonomy_version in task.toml metadata and its evaluator dataset record.

View tasks and embedded images

Extract the ZIP and open index.html in a browser. Search by task ID or query, or filter by modality. Each task's environment/inputs/instruction.html displays its query, Base64-embedded reference images and declared attachments. Videos play from the bundled files using relative paths.

All 130 input images are available in environment/inputs/reference_images.base64.json. Each image record includes path, media_type, bytes, sha256, width, height and a complete data_uri. Tasks without input images have an empty images array. The one verifier-only image is in bw_0182/tests/rubric_reference_images.base64.json.

import base64, hashlib, json

with open("bw_0006/environment/inputs/reference_images.base64.json") as f:
    image = json.load(f)["images"][0]
raw = base64.b64decode(image["data_uri"].split(",", 1)[1], validate=True)
assert hashlib.sha256(raw).hexdigest() == image["sha256"]

Base64 decoding returns the full image file bytes. base64_manifest.json lists all image locations and fingerprints. Standard image files are also available for Harbor agents and the evaluator. The agent's instruction.md contains a short text prompt and file paths; encoded image payloads are in the HTML/JSON files.

Run with Harbor

python -m pip install "harbor[modal]==0.23.0"
cd BlenderWorld-harbor-base64
sha256sum -c SHA256SUMS

Configure Modal credentials with modal token new, or use your existing Modal credentials. You can select another GPU-capable Harbor backend with --env. Harbor 0.23.0's built-in Docker backend does not allocate GPUs from task.toml; use a backend that supports the declared GPU requirement.

Configure the generation agent's provider credentials. For the verifier, set the environment variables documented in verifier.env.example:

Variable	Purpose
BW_AGENT_JUDGE_ENDPOINT	Anthropic-compatible endpoint for rendering/scene-inspection agents
BW_AGENT_JUDGE_API_KEY	Key for that endpoint
BW_VISION_JUDGE_ENDPOINT	Full Chat Completions endpoint for the vision judge
BW_VISION_JUDGE_API_KEY	Key for the vision endpoint
BW_RENDER_MODEL	Optional rendering-agent model alias
BW_AGENT_JUDGE_MODEL	Optional scene-inspection model alias
BW_VISION_JUDGE_MODEL	Optional vision-judge model alias

Use model aliases supported by your endpoints. Then run a task or the full dataset:

harbor run -p ./bw_0006 --env modal -a codex --ak version=0.147.0 -m 'provider/model'
harbor run -p . --env modal -a codex --ak version=0.147.0 -m 'provider/model' --n-concurrent 1

For Claude Code, use -a claude-code --ak version=2.1.220. Set -p to a task directory or the package directory. dataset.toml is the publication manifest.

Inputs, outputs and scoring

The agent reads /workspace/inputs/ and writes its editable project to /workspace/exports/final.blend. Keep requested images, videos, model exports and texture dependencies inside /workspace/exports/ with relative paths.

Harbor transfers the exports to a separate verifier container. The verifier packages a standard bw_XXXX.zip submission and uses the BlenderWorld evaluator. /logs/verifier/reward.json contains reward, rubric_score and dimension_score on a 0–1 scale. Multiply mean reward by 100 for the overall score. Detailed scores and evidence are in metrics.json and run/.

To inspect a collected submission using installed BlenderWorld tools:

bw evaluate --dataset ./bw_0006/tests/dataset \
  --submission /path/to/bw_0006.zip --output /path/to/new-evaluation --dry-run

Configure the runtime and judge endpoints, then omit --dry-run to render and score.

Layout and runtime

BlenderWorld-harbor-base64/
  index.html
  README.md
  使用说明.md
  dataset.toml
  base64_manifest.json
  export_manifest.json
  validation.json
  verifier.env.example
  SHA256SUMS
  bw_0001/
    instruction.md
    task.toml
    environment/
      Dockerfile
      inputs/
        query.txt
        input.json
        instruction.html
        reference_images.base64.json
        references/                # when present
        attachments/               # when present
    tests/
      Dockerfile
      test.sh
      blenderworld/
      dataset/
  ... bw_0223/

Runtime: Blender 4.2.10, Node.js 22.17.0, Codex CLI 0.147.0, Claude Code 2.1.220, MCP for Blender 2.1.3. The generation environment and the separate verifier environment each request 8 CPUs, 16 GiB RAM, 100 GiB storage and 1 GPU. The package does not restrict the GPU model. Use hardware compatible with Blender Cycles; configure CUDA or OptiX when rendering on NVIDIA GPUs. The generation timeout is 8 hours (28,800 seconds) per task. The verifier has its own 8-hour timeout. These are separate phase budgets; environment build timeout is 1 hour. The GPU requirement is specified in both [environment] and [verifier.environment] in every task.toml. Use a generation agent; no oracle solution is bundled. The MIT LICENSE covers the adapter/evaluator code. Use SHA256SUMS to check package integrity and validation.json to inspect the available validation results.