m3ultra
|
58a2a8177d
|
operator: qwen_layered_local — image -> N editable RGBA layers (Qwen-Image-Layered 20B, MLX)
Vendored mflux PR#302 fork (monster/Qwen-Image-Layered-MRP-MLX), tuned
defaults 20 steps/640/baked-q8 (243s m3ultra, 366s m1ultra). Baked-model
resolution via QWEN_LAYERED_MODEL env or per-box paths, HF+q8 fallback.
nodes.json (gitignored, machine-local) updated on m3 to route to m1 too.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
|
2026-07-19 22:06:05 +10:00 |
|