Spaces:
Running on Zero
Running on Zero
Commit History
4 days ago
-
Drop the comment on ssr_mode Running
Verified signature6f3a894 -
Shorten the comment on ssr_mode
Verified signature3d8a890 -
Launch without server-side rendering
Verified signatureGradio's launch check of the page through the Node proxy can time out while the model loads, which crashes the app at startup.306b9ee
11 days ago
-
Add domain tags for the Home tab filters of HARP
Verified signaturee619a16
14 days ago
-
Return an AIFF input as WAV
Verified signatureHARP accepts .aif, the extension Logic Pro gives AIFF, but soundfile cannot write a file named with it, so such an input failed when its result was written, inside the GPU call. Gradio also converts an AIFF output to 16-bit WAV, since browsers cannot play AIFF, so keeping that container gained nothing. Lossless inputs now keep WAV and FLAC alone, and AIFF comes back as WAV at the bit depth this Space writes. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>6e74851
15 days ago
-
Refuse an empty file before the GPU is reserved
Verified signatureAn audio file of no frames passed the checks in process_fn, since soundfile reads it without complaint and the limits checked are maximums, so it reached the GPU call and failed there, spending the caller's ZeroGPU allowance. It is refused with a message first. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>4ceb571 -
Move the model card ahead of the model loading
Verified signatureFollows the layout of docs/DEPLOYMENT.md, which puts the card after the imports and before the loading. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>c31ecc7 -
Keep a lossless container
Verified signatureReturns a FLAC or AIFF input in its own container. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>4c6f562 -
Drop the length cap on the audio, keeping the reservation sized to it
Verified signatureThe model works through it in segments or chunks, so its length costs time rather than memory. ZeroGPU already refuses a request that the caller's tier or remaining allowance cannot cover, before reserving anything, and tells a free account how to get more. A cap set so that a free account could afford every request therefore only turned away callers who could. The only request refused here now is one whose reservation would outlast PyHARP's 15 minute timeout, which would stop it anyway.2e68f3b
16 days ago
-
Reserve GPU time by chunk, as measured, and declare the Python version the Space runs.
Verified signatureed17409 -
Move to ZeroGPU: the Gradio SDK held at 6.17.3, torch 2.8.0, the pinned checkpoint built once in the worker, and requests checked and sized before the GPU.
Verified signaturef626c4a
19 days ago
-
Label the output with what it holds, which HARP shows as its name.
Verified signatureac5cb0b
26 days ago
-
Tidy the launch call and drop stray files
Verified signature40b9ecb -
Roll pyharp and tidy the card description
Verified signaturef90a300 -
Prune tags that predate the taxonomy
Verified signaturea6b25d0 -
Declare the model's own sample rate and channel count
Verified signature3ce8224
27 days ago
-
Restore the input's sample rate and channel count
Verified signature10ef29b -
Describe what each control does to the sound
Verified signaturef4061db -
Credit the original authors and link the repository and paper
Verified signaturee99434d -
Fix the launch keywords and the taxonomy import
Verified signature3d8ce40 -
Bring the Space up to the pyharp example standard (taxonomy tags, __main__ guard, show_error, v0.4.0)
Verified signature295b938
-
Monkeypatch lowpass_filtering_prepare_inference boundary bug
audiosr==0.0.7's lowpass_filtering_prepare_inference caps cutoff_freq at exactly 24000 Hz (both via normal detection saturating there for broadband audio, and via its 'give up' fallback for quiet/narrowband audio), while sampling_rate for this step is hardcoded to 48000 upstream regardless of actual input. nyquist == 24000 exactly, so Wn == cutoff_freq/nyquist can equal exactly 1.0, which scipy's iirfilter rejects (requires 0 < Wn < 1 strictly). This affected all 13 model_validation test cases identically, regardless of input sample rate, DDIM steps, guidance scale, or multiband settings — not just a low-sample-rate edge case. Patches the pipeline module's bound reference (not audiosr.utils's copy, which pipeline.py already imported by name at its own load time) to clamp cutoff_freq to 98% of nyquist as a safety margin.997cf76
-
Resample low-sample-rate input to 48kHz before inference
AudioSR's internal cutoff-frequency detection (lowpass_filtering_prepare_inference) hardcodes a 24kHz-Nyquist assumption regardless of actual input sample rate, which can produce an invalid scipy filter Wn and crash on input below 48kHz. Confirmed via model_validation (teamup-tech/audiosr, 2026-08-06) — every test case with sample_rate=8000 failed identically with: ValueError: Digital filter critical frequencies must be 0 < Wn < 12e960ad
-
add multiband ensemble control, clean up debug prints
9a89e2a
-
add server name
c053253 -
add startup debug prints
8f7580c -
add startup debug prints
55c1452 -
add startup debug prints
0378e6f -
add startup debug prints
9c8e5e1 -
add startup debug prints
d28ce35 -
lazy load model on first request
53ce9f0
-
add flush=True to print statements
c3d0f7a -
use super_resolution_long_audio from repo
e89501c
-
remove latent_t_per_second hardcode, use library default
51984cc
-
add guidance_scale and seed controls, fix author, fix ddim steps range
e13f695
-
initial commit
64638ca