Blender AI-Replication

Drive a Blender scene from a target image via LLM-orchestrated MCP commands and a VLM critic loop. The pattern is: target image → first-version model → per-round screenshot + VLM critic + diff prompt → next round, then render → export glTF. Companion skill to blender-design, which handles non-iterative Blender work without an LLM critic.

This skill is the bridge between three otherwise independent worlds: a Blender MCP transport (PartMe / Codex / ahujasid), an LLM driver (豆包 / Claude / OpenAI / Doubao Seed 2.1 Pro), and a VLM critic (豆包 vision / OpenAI vision / any multimodal model). The skill ships the orchestration glue — the user wires the three endpoints.

For a full first-version command sequence of a 《清明上河图》 micro-scene, see references/qingming-sequence.md. For the orchestrator, see scripts/dream_loop.py.

什么时候使用（When to Use）

Use this skill when: