Qwen3.5 learned to make valid Blender rooms, but adding detail was hard, a user reports

A user says they spent Astra tokens trying to teach Qwen3.5 (9B) to create low-poly 3D rooms in Blender. Their initial experiments got the model to make valid rooms, they report, but adding detail proved very hard. They say stricter rewards offered less benefit, and neither GRPO nor image guidance helped. The user shared the experiment setup and suggested trying different rewards, a larger Qwen model, or fine-tuning on 150–200 paired Blender renders and code examples before another attempt.