SAVRN's Take
Nothing on this page needs a rack. Parameter counts run from 4M for NVIDIA's segformer-b0 to 503M for Microsoft's beit-large-finetuned-ade-640-640, and that largest checkpoint asks for 1.2 GB at 16-bit. BiRefNet at 221M parameters wants 0.5 GB, CLIPSeg 0.4 GB, and segformer-b0 rounds to zero. All seven priced entries among the top downloads land on the same Index floor, 1x MI300X at $1.85 an hour, so what you are sizing is images per hour and batch depth, not which card.
The work divides three ways. BRIA AI's RMBG-1.4 and RMBG-2.0 cut a foreground from its background, and Peng Zheng's BiRefNet does high-resolution dichotomous segmentation. Meta's Mask2Former checkpoints and NVIDIA's SegFormer b0 and b2 handle semantic segmentation, trained on Cityscapes or ADE20k. CIDAS's clipseg-rd64-refined takes a text or image prompt inside a 77-token window for zero-shot and one-shot work, and it leads the category at 1,126,497 downloads a month, ahead of BiRefNet at 976,013 and RMBG-2.0 at 639,266.
Licenses are where to slow down. The label other covers 28 of the 67, MIT 24, Apache 2.0 11, 3 state nothing, and 1 is AGPL-3.0. BiRefNet and RMBG-2.0 share one architecture and the same 220,700,242 parameters, but BiRefNet is MIT and open access while RMBG-2.0 is gated and licensed as other, so the MIT path is shorter. AI at Meta's 15 entries and NVIDIA's 9 lead the publisher list, and their Mask2Former and SegFormer checkpoints carry other too. Read the terms behind other before you commit, and match the fine-tuning set, Cityscapes or ADE20k, to the scenes you serve.
SAVRN Research, 2026-09-18