Marigold V2 recipe: 4-bit quantized DiT + rank-128 QLoRA, fine-tuned on one 32GB GPU
AntonObukhov1 · x · 2026-09-09
Marigold V2 keeps unconventional choices that work. V1 pushed depth maps through an image VAE — no business working, but it did. V2 pushes ground-truth depth through DINOv3 and aligns DiT features to it (iREPA-depth). Recipe: take a large pretrained DiT (Qwen-Image-Edit-2509), quantize to 4-bit, add a rank-128 QLoRA, fine-tune on a single 32GB consumer GPU — sensible depth after a few hours, done in a few days, no 80GB cards or multi-node setups. Single-step inference, no OOM at 2K resolution.
Related event: Marigold V2: A Single-GPU DiT Depth Estimator Hits SIGGRAPH Asia 2026(11 posts)→
More from Infra
- Two CVSS 9.9 Wasmtime flaws allow sandbox escape, one ARM64-only since 2020 — basedjensen · 2026-10-27
- AI Infra Meme: While Countries Need TW-Scale Power for Datacenters, People Argue About Lightbulbs — mgostIH · 2026-10-27
- Legora runs 2,000 parallel agent evals on Modal Sandboxes for legal AI — blelbach · 2026-10-27
- Cloud release velocity scores: AWS 95, Azure 90, Google Cloud 82 — DavidLinthicum · 2026-10-27
- Selling a 5090 to scale local inference: one business weighs 5 hardware paths — Anarchaotic · 2026-10-27
- AWS tutorial: fleet-wide edge CV model lifecycle with SageMaker SDK v3, ~$0.07 per run — AWS ML Blog · 2026-10-27