laurentvv/agents-kit/blender-pipeline-automation

Pipeline Automation

The execution model is in blender-core (run_in_blender.py finds Blender and prints only the JSON results).

Batch one-liners

blender -b -P scripts/batch_process.py -- --in "assets/**/*.fbx" --op convert --to glb --out export/
blender -b -P scripts/batch_process.py -- --in library/ --op report                      # inventory JSON lines
blender -b -P scripts/batch_process.py -- --in "src/*.blend" --op script \
    --script ../blender-asset-qa/scripts/qa_check.py --kwargs '{"profile":"game"}'       # QA a whole library
blender -b -P scripts/batch_process.py -- --in src/ --op script --script ../blender-game-optimization/scripts/make_lods.py \
    --kwargs '{"rename":true}' --save-to processed/                                      # LODs for every file

Each file is streamed as one JSON line with ok, timing and results. Failures don't stop the batch, and the summary lists them. Parallel: start N processes with --shard 0/N … --shard N-1/N (Blender uses one core for Python, so processes scale with CPU/GPU count).

Headless CLI essentials

Environment variables: BLENDER_USER_SCRIPTS, BLENDER_USER_EXTENSIONS, BLENDER_SYSTEM_PYTHON. Use OCIO for custom color configs.

bpy as a Python module

pip install bpy==5.2.* (Python 3.13, matching the Blender release) gives import bpy in plain Python, which is handy for pytest/CI/web services. It has no UI and no GPU module by default, and it's large (~300 MB). Version-lock it with the Blender you target. The bpy wheel and the CLI behave the same for data API work.

Library patterns

• Append/link: bpy.data.libraries.load(path, link=...) (see blender-core/references/bpy-patterns.md). 5.2: data_from.libraries exposes nested library paths when loading.
• Write a subset: bpy.data.libraries.write(path, {datablocks}, fake_user=True, path_remap="RELATIVE_ALL") builds asset/library files from a session.
• Find files: bpy.data.file_path_foreach(...) (5.2 flags: EXPAND_TOKENS for UDIMs, EXPAND_SEQUENCES, EXPAND_CACHES), and bpy.utils.blend_paths(absolute=True) lists every external path.
• Remap/relocate: bpy.ops.file.find_missing_files(directory=...), make_paths_relative, pack_all. Libraries: lib.filepath = new; lib.reload().
• Iterate every ID: bpy.data.all_ids (5.2), and bpy.data.orphans_purge(do_recursive=True) for cleanup.

Robustness rules

• One task per process where possible. Long Blender sessions leak (images, undo). batch_process.py reuses the process but resets with read_factory_settings(use_empty=True) / open_mainfile.
• Never write into the source folder by default. Use separate --out/--save-to, and write with copy=True.
• Make scripts idempotent: re-running replaces their own outputs (named prefixes, exist_ok).
• Log JSON, not prose: one object per file/job, so agents and CI can parse it cheaply.
• Pin the Blender version (5.2 LTS) for a project. Record bpy.app.version_string + hash (bpy.app.build_hash) in the outputs.
• Paths: bpy.path.abspath("//x") is relative to the opened .blend. In headless batches use absolute paths, or set them after open_mainfile.
• GPU in containers: Cycles CUDA/OptiX needs the NVIDIA container runtime. Otherwise use CPU. EEVEE needs a GPU/EGL even headless.

CI recipe (GitHub Actions)

Download the Linux tarball (https://download.blender.org/release/Blender5.2/blender-5.2.3-linux-x64.tar.xz), cache it, then blender -b --factory-startup --python-exit-code 1 -P tests/run_all.py. This repo's .github/workflows/ci.yml is a working example.