Fine-tuned Apple Depth Pro 2 weights produce far sharper 3D normal maps
flngr · x · 2026-11-22
A community experiment fine-tunes Apple's Depth Pro 2 (DepthPro-S, 97MB) weights to produce detailed 3D normal maps the original never output. Benchmark images show the original model with slope correlation near 0 and MAE of 4–11 degrees across brick, fabric and pavement textures, while the fine-tuned exp005/exp007 checkpoints reach slope r of 0.51–0.72 with MAE around 3.6–4.4 degrees on structured materials — a clear detail-fidelity gain.
More from Multimodal
- AI short film: a 30-year knight-vs-sky-whale revenge tale made with Seedream 5.5 and Seedance 2.5 — aziz4ai · 2026-11-22
- Open-source Claude Code plugin cuts complete music videos from a single song, now at v1.7 — AIandDesign · 2026-11-22
- CapCut Web Ultra offers 7 days of unlimited AI image gen with Seedream and GPT Image, Nov 20–26 — Shruti_0810 · 2026-11-22
- Mflux Runs MOSS Image Gen on 8GB Macs: 82s for 768x768 With 4-Step Distill LoRA — bio_risk · 2026-11-22
- AI video demo: a princess leads her wounded husband's army unnoticed — azed_ai · 2026-11-22
- Kotoba CEO: Audio Tokenization Is the Core Research Problem in Real-Time Speech AI — lmthang · 2026-11-22