Instructions to use KoshiMazaki/akuspace-ltx25 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- LTX.io
How to use KoshiMazaki/akuspace-ltx25 with LTX.io:
# Install the LTX-2 pipelines git clone https://github.com/Lightricks/LTX-2.git cd LTX-2 uv sync --frozen
# Download the weights from this repo, plus the Gemma text encoder hf download KoshiMazaki/akuspace-ltx25 --local-dir models/akuspace-ltx25 hf download google/gemma-3-12b-it-qat-q4_0-unquantized --local-dir models/gemma-3-12b
# Text/image-to-video with the LoRA on the HQ two-stage base pipeline uv run python -m ltx_pipelines.ti2vid_two_stages_hq \ --checkpoint-path path/to/checkpoint.safetensors \ --distilled-lora path/to/distilled_lora.safetensors 0.8 \ --spatial-upsampler-path path/to/spatial_upsampler.safetensors \ --gemma-root models/gemma-3-12b \ --lora models/akuspace-ltx25/<weights>.safetensors 1.0 \ --prompt "your prompt here" \ --output-path output.mp4 # For image-to-video, add: --image path/to/image.jpg 0 0.8 - Reverb
How to use KoshiMazaki/akuspace-ltx25 with Reverb:
# No code snippets available yet for this library. # To use this model, check the repository files and the library's documentation. # Want to help? PRs adding snippets are welcome at: # https://github.com/huggingface/huggingface.js
- Notebooks
- Google Colab
- Kaggle
File size: 2,971 Bytes
e088a4d | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 | """Envelope cross-correlation between a dry reference and its a2a render.
Reproduces the timing measurement published in the AKUSPACE model card, so new
examples can be added to that table on the same basis rather than asserted.
Method: amplitude envelope of each signal in 2 ms RMS windows, mean-removed and
normalised, cross-correlated over +/-250 ms. Reports the offset of peak
correlation and the peak value.
Reading the result: peak r measures how much the envelope CHANGED, and reverb
changes it by design — most on transient-dense material, where the tail fills
the gaps between hits. A low r on percussion is the effect working, not drift.
The offset is the timing claim; r is not a quality score.
"""
import argparse
import subprocess
import sys
import wave
from pathlib import Path
import numpy as np
WIN_MS = 2.0
MAX_LAG_MS = 250.0
def load_mono(path: Path, sr: int = 48000) -> np.ndarray:
"""Decode anything ffmpeg reads into mono float at sr."""
out = subprocess.run(
["ffmpeg", "-v", "error", "-i", str(path), "-ac", "1", "-ar", str(sr),
"-f", "wav", "-c:a", "pcm_s16le", "-"],
capture_output=True, check=True).stdout
import io
with wave.open(io.BytesIO(out), "rb") as w:
raw = w.readframes(w.getnframes())
return np.frombuffer(raw, dtype="<i2").astype(np.float64) / 32768.0
def envelope(x: np.ndarray, sr: int) -> np.ndarray:
n = max(1, int(sr * WIN_MS / 1000.0))
trimmed = x[: len(x) - len(x) % n]
return np.sqrt((trimmed.reshape(-1, n) ** 2).mean(axis=1) + 1e-12)
def xcorr(a: np.ndarray, b: np.ndarray, max_lag: int) -> tuple[int, float]:
n = min(len(a), len(b))
a, b = a[:n], b[:n]
a = (a - a.mean()) / (a.std() or 1e-12)
b = (b - b.mean()) / (b.std() or 1e-12)
best_lag, best_r = 0, -2.0
for lag in range(-max_lag, max_lag + 1):
if lag < 0:
x, y = a[-lag:], b[: n + lag]
elif lag > 0:
x, y = a[: n - lag], b[lag:]
else:
x, y = a, b
if len(x) < 10:
continue
r = float((x * y).mean())
if r > best_r:
best_r, best_lag = r, lag
return best_lag, best_r
def main() -> int:
p = argparse.ArgumentParser(description=__doc__, formatter_class=argparse.RawDescriptionHelpFormatter)
p.add_argument("--dry", required=True)
p.add_argument("--wet", nargs="+", required=True)
p.add_argument("--sr", type=int, default=48000)
a = p.parse_args()
dry = envelope(load_mono(Path(a.dry), a.sr), a.sr)
max_lag = int(MAX_LAG_MS / WIN_MS)
print(f"{'example':38s} {'offset':>9s} {'peak r':>8s}")
for w in a.wet:
wet = envelope(load_mono(Path(w), a.sr), a.sr)
lag, r = xcorr(dry, wet, max_lag)
# positive lag = wet later than dry; report with the card's sign convention
print(f"{Path(w).stem:38s} {lag * WIN_MS:+8.0f}ms {r:8.2f}")
return 0
if __name__ == "__main__":
sys.exit(main())
|