Ablation on a 250-word image prompt: only the first two sections do the work
Relevant_Square4919 · reddit · 2026-11-20
The author ablated a 250-word, five-section miniature-diorama image prompt across seven variants on the same CC0 photo with a fixed seed. Findings: a one-line prompt yields only a clay texture; removing [Task and composition] kills the diorama entirely; removing [Materials and depth] degrades it to a plain architectural model; the remaining three sections barely matter. Verified with a second seed on locally-run SenseNova-U1.5-8B-MoT (8-step, CFG 1). The template is open-source in the SenseNova-Skills repo.
More from Multimodal
- From 20 photos to a walkable 3D apartment: World Labs Atlas + Fable 5.2 pipeline demoed — XRarchitect · 2026-11-21
- Gaussian splat top-down scan shows how Fable plans object placement in pixel apartment — XRarchitect · 2026-11-21
- Pixel replica of a real apartment built from Atlas gaussian splats with Fable 5.2 — XRarchitect · 2026-11-21
- Fable 5.2 autonomously drafts detailed architectural plan of John Hancock building in Aircad — iskander · 2026-11-21
- One dev orchestrates Fable, Opus and Astra to build a visually striking survival game — lyraxana · 2026-11-21
- CapCut plugin connects ChatGPT to auto-cut gameplay highlights, then human picks the payoff — iamfakhrealam · 2026-11-21