mlx-community/Ming-Image-0.1-Design-bf16

Text-to-Image • Updated about 13 hours ago

Note Design text-to-image with native RGBA alpha. bf16 reference precision (49.8 GB). Peak 49.9 GB (the MLLM conditioning stage): under MLXEngine it needs a 96 GB Mac — the 8-bit and 4-bit tiers cover 48 GB and 36 GB.

mlx-community/Ming-Image-0.1-Design-Layer-bf16

Image-to-Image • Updated about 13 hours ago • 1

Note Flat design → editable straight-alpha RGBA layers (front-most first) + composite. bf16 reference precision (49.8 GB). Peak 50.4 GB: under MLXEngine it needs a 96 GB Mac — see the 8-bit and 4-bit tiers.

mlx-community/Ming-Image-0.1-Design-8bit

Text-to-Image • Updated about 14 hours ago

Note 8-bit tier (MLLM, connector, DiT; 28.3 GB). Matches bf16 on every gate: text 118/122 strings, zones 19/20, native alpha 5/6; 36.5 dB from the fp32-truth render. Peak 27.8 GB — admitted on a 48 GB Mac.

mlx-community/Ming-Image-0.1-Design-Layer-8bit

Image-to-Image • Updated about 13 hours ago

Note 8-bit tier (28.3 GB). Decompositions within 0.1 dB of bf16 on every gated signage still. Peak 28.4 GB — admitted on a 48 GB Mac.

mlx-community/Ming-Image-0.1-Design-4bit

Text-to-Image • Updated about 13 hours ago

Note 4-bit MLLM, 8-bit connector + DiT (20.5 GB). Peak 19.7 GB — admitted on a 36 GB Mac. Small measured cost: text 116/122 strings (10/16 boards), native alpha less reliable — use 8-bit or bf16 for transparent assets.

mlx-community/Ming-Image-0.1-Design-Layer-4bit

Image-to-Image • Updated about 13 hours ago

Note 4-bit MLLM, 8-bit connector + DiT (20.5 GB). Decompositions within 0.2 dB of bf16. Peak 20.6 GB — admitted on a 36 GB Mac.