michaelgold/blenderbench-direct-results

BlenderBench Direct reproduction artifacts

This dataset preserves artifacts and provenance for an independent, community-run reproduction of the public BlenderBench task set. It is not an official Blender Foundation product, official BlenderBench submission, or leaderboard result.

Experiment

• Dataset: DietCoke4671/BlenderBench revision 203e4d325e9438ca55b29bdfc4f6a90842d74e68
• Attribution: DietCoke4671 and contributors, CC BY 4.0
• Generation model: gpt-6-astra
• Codex CLI: 0.154.0
• Runner implementation: bpy-dev/blender-mcp commit 4a9f5cc6e616080185252bb919fa3d1e730760b7, with the portable publication package linked after its review PR opens
• Scope: 27 tasks, one sequential 10-round trajectory per task, 270 rendered/scored rounds
• Generation isolation: CLIP scores and evaluator output were not fed back to generation

Metric

The reported CLIP value is image-to-image cosine similarity from openai/clip-vit-base-patch32, revision 3d74acf9a28c67741b2f4f2ea7635f0aaf6f0268. N-CLIP is 1 - similarity. This is not literal task accuracy. Best-round selection minimizes N-CLIP; paired photometric loss is taken from the same round, with the earliest round winning ties.

The evaluator source is pinned to Fugtemypt123/VIGA-release revision 69cb8ef0651bf68124815682df4a0f8e8b57c141, SHA-256 9ee08818aa858dccfff3efbb36cb9a98e8887f14e6168ad1a23731351cdeb3b5.

Contents

• manifests/artifact-manifest.json: task/round hashes and byte sizes for blends, renders, targets, scores, process records, and video
• manifests/clip-snapshot-manifest.json: exact offline CLIP snapshot allowlist, sizes, and SHA-256 digests
• results/results-summary.json: compact aggregate and task-level results
• results/reference-results.json: results plus historical generation/scoring provenance and limitations
• video/blenderbench-complete-compilation.mp4: 27-task, 270-round compilation
• renders/: rendered PNG checkpoints (uploaded in resumable batches)
• checkpoints/: generated .blend checkpoints (uploaded in resumable batches; total approximately 161 GiB)
• evidence/: sanitized process records where publication review confirms no credentials, goal/solution code, or machine-specific secrets

Every uploaded object is checked against artifact-manifest.json. Missing directories indicate an upload batch that has not yet completed; the manifest remains the inventory of the completed run.

Limitations

• One trajectory per task; rounds within a task are sequential revisions, not independent samples.
• The public task set may be present in model training data; there is no hidden evaluation set here.
• The generation model, inference budget, and direct-runner protocol differ from VIGA, so numeric comparisons are not controlled architecture comparisons.
• BlenderBench exposes no hosted submission endpoint or official public leaderboard for this run.
• No VLM judge was used by design.
• Local runtime preflight and byte manifests are provenance checks for an honest operator, not remote, hardware, or hostile-host attestation.

Reproduction package

The portable runner, frozen protocol, scoring lock, evaluator retrieval, tests, and documentation are published through bpy-dev/blender-mcp. Large artifacts live here rather than in normal Git history.