Spaces:
Running
Running
Commit History
27 days ago
-
SANA-Sprint is a model switch, at any step count (flux-klein.js 0.5.1)
The badge next to the step picker is now the switch: SANA-Sprint LoRA or the stock 4B weights, whatever the step count. The engine is created with lora: 'web_weights_lora_sprint' and each generate() passes lora: true|false (0.5.1's per-call override), so flipping is instant — the base never reloads. Default stays SANA-Sprint at 2 steps. The headless hook takes { sprint, fastDecode }.5be91a0 -
Add SANA-Sprint few-step model: 1-2 steps run the sprint LoRA, 3-4 stock
The engine (flux-klein.js) already applies a runtime lora-i8 set selected by step count; this wires the SANA-Sprint LoRA (rank-256, from radames/flux2-klein-edge-web web_weights_lora_sprint) to the 1-2 step presets and shows a model badge in the prompt bar. Default is now 2 steps (sprint). A dev-only redirect (VITE_SPRINT_LORA_BASE + scripts/serve-lora.mjs) can point just the LoRA at a local copy for testing.c12f0b0
-
fast decode on by default
d14b575 -
fast decode: tiny VAE decoder toggle via flux-klein.js ^0.4.0
Surfaces the fal FLUX.2-Tiny-AutoEncoder decoder (decoder:'tiny', per-generation) as a 'fast decode' checkbox in the prompt bar — ~20x faster decode, slightly softer textures. Engine dep is the npm release.8ff29b4
-
bump
b775f29
-
initial commit
9c78030