A 'block first, generate second' AI video tool seeks feedback on its 3D pre-vis workflow
KeyCod3923 · reddit · 2026-09-08
The author describes their in-progress AI video tool built around a "block first, generate second" approach: an agent on dimension.so lays out a 3D pre-vis scene — object blocking, character motion, camera — which is then used as a video-to-video reference for Seedance 2.0 Mini instead of prompting from scratch.
Unlike a companion post, the author here says adding GPT-6/Astra to the agent's reasoning layer did not produce a noticeable improvement in multi-step task handling or composition consistency. They're asking the community what works in the pre-vis stage and what to improve before generation.
Related event: Block First, Generate Second: 3D Pre-vis Workflow Tames AI Video(2 posts)→
More from coding & agent
- Using opencode + local Qwen to build Windows apps: how to let the LLM test its own code — mototuneup · 2026-09-08
- Same spec, two AI VTuber apps: Claude Fable 5.1 vs GPT-6 Astra builds, both on MiniMax H3 — Fresh_Chemistry7021 · 2026-09-08
- Arduino AI coding assistant user burns through monthly tokens by day 7 — unixterminal · 2026-09-08
- Modular MiniMax H3 workflow enables seamless ultra-long video with segment regeneration — wjc_5 · 2026-09-08
- What broke running an agent fleet 24/7 — it wasn't prompt quality — ParrotIntegrated · 2026-09-08
- Microsoft shipped Friday; by Tuesday a stranger had reverse-engineered it with official tokens — unixterminal · 2026-09-08