Spaces:
Running on Zero
Running on Zero
Commit History
-
archive every API generation to a private dataset at generation time (SA3_ARCHIVE_REPO/TOKEN); off when unset, failures only log Running
cf4824b
-
rebuild
65e086a -
api: generate_inpaint mask_soft_s — the kept context fades out of the conditioning before the mask edge instead of stopping at a wall
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>72d8940 -
api: /generate_inpaint — keep audio outside a mask range, regenerate inside (continuation for hisia chunks)
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>d8ac0be
-
POC tabs: audio→prompt (3 lenses) and audio→audio+prompt
Two new demo tabs, placed ahead of the raw playground tabs (renamed "Playground · text" / "Playground · audio"). 1 · Audio → prompt — one upload fans out to three concurrent audio-brief /audio_to_prompt calls (lens=match/loose/free, llm+fast) for Faithful / Fresh arrangement / Exploratory prompts, a measured readout (bpm, key, LUFS, true peak, sections, stems, top tags) and a copy-paste markdown block for Ableton MCP. Analysis is CPU/network and sits outside @spaces.GPU so a 20-40s LLM wait can't burn GPU seconds; .submit()+poll, never .predict(). Per-prompt "Generate 10s preview" buttons hit the local _generate under a GPU wrapper, so visitors spend their own quota. 2 · Audio → audio + prompt — "Generate pair" runs A2A at noise 0.4 and a pure T2A of the same prompt sequentially (ZeroGPU serialises anyway), then "Generate 3 more" exposes two plain-language knobs: Match to source % → init_noise_level = 1 - pct/100 clamped [0.05, 1.0], and Prompt creativity Less/Default/More → cfg_scal…11e52c4 -
a2a: decode init audio via soundfile — torchaudio.load needs torchcodec, absent on ZeroGPU
d6581bf -
README: short_description within HF 60-char limit
0df673c -
abv1 sa3 engine: SA3 medium + small-music, UI + headless API
0560c42