Marigold V2 hands-on: running batch depth/normals on a dedicated GPU
Admirable-Cod-Lara · reddit · 2026-10-01
The author has been testing Marigold V2 since it landed in ComfyUI 0.36, building a small setup around the official demo covering VRAM requirements and running it on a dedicated GPU for larger batch workloads.
The most interesting finding is the significant VRAM jump between 1024px and 2048px workloads. Details are posted in the model's Hugging Face discussion, and the author asks what hardware others use — 3090/4090, A-series, or cloud GPUs.
More from Infra
- Ex-OpenAI policy VP: compute is now split into ~3 tiers and the gap is widening — Miles_Brundage · 2026-10-01
- DRAM supply shows no line of sight to catching up with demand, says analyst — BenBajarin · 2026-10-01
- Micron: humanoid robots may need memory comparable to autonomous vehicles, driving demand by 2030 — McDonaghMatthew · 2026-10-01
- antirez praises 192GB Framework Desktop, plugs external GPU for heterogeneous inference — antirez · 2026-10-01
- RTX 5090 + Intel Arc B70 for local LLMs: halved throughput vs bigger context — indiealexh · 2026-10-01
- Google explores space data centers: keeping AI chips from overheating in orbit — GraceToSentience · 2026-10-01