GPT-6 Astra's 3D modeling falls far short of the demos in hands-on Blender testing
Next_Technology6361 · reddit · 2026-09-10
A hands-on tester finds GPT-6 Astra + Blender MCP underwhelming at actual shape modeling: given reference photos of a vintage fire truck, Tripo3D produced a convincing rough model, but Astra's attempts to "improve" or rebuild geometry degraded proportions into generic rounded boxes and primitive extrusions — the same happened with human heads, including a neck-fix that turned horror-movie. Astra excels instead at MCP setup, scripting, materials, lighting, rigging, scene assembly and mechanical geometry. His best workflow: Tripo3D as the shape generator, Astra as the Blender operator/technical artist. He suspects the flashy launch demos lean on pre-generated assets and heavily scaffolded tasks, and notes even its prompt-translation skills needed a complex cinematic prompt to get close to expectations.
More from Multimodal
- Open-Source Ref2V Workflow Auto-Transcribes Media and Formats Prompts for H3 Video Generation — bstr3k · 2026-09-10
- fal launches Multi-angle for H3 Max: any camera angle from one image in under 3 seconds — gorkem · 2026-09-10
- Valeo fits scaling laws for video diffusion using 5,500 hours of driving footage — abursuc · 2026-09-10
- Zero-cost fix for Minimax H3 plastic skin: drop contrast before VAE decode — Cequejedisestvrai · 2026-09-10
- 15-second single-pass AI bike chase: three craft lessons from a MiniMax H3 video — LudovicCreator · 2026-09-10
- Seven-episode AI-generated Dragon Ball fan series drops episode 7 teaser — Stunning_Act_1530 · 2026-09-10