iSrSSKJr SSKrSSKJr S SjrS SjrSSjrSS jr SSS jjr SSS jjr SSS jjr g)uWavelet-domain DCW primitives — the actual ``dcw_low``/``high``/``double``/``pix`` math. Verbatim from the paper's reference implementation (``AMAP-ML/DCW/FlowMatchEulerDiscreteScheduler.py``) but adapted to the 1-D temporal layout of ACE-Step latents (``[B, T, C]``): * The reference operates on 2-D image latents with ``DWT2DForward``; ours uses ``DWT1DForward`` along the temporal axis after transposing ``[B, T, C] → [B, C, T]``. * ``pytorch_wavelets`` zero-pads odd ``T`` to the next even before the filter bank, so the IDWT output is one sample longer than the input. We trim ``x_new[:, :, :out_T]`` so ACE-Step's odd-duration latents round-trip cleanly. ) annotationsN) WAVELET_CACHEcBURSS5R5$)z?Rearrange ACE-Step latents from ``[B, T, C]`` to ``[B, C, T]``.r transpose contiguousxs [/mnt/workspace/acestep.cpp/tests/../../ACE-Step-1.5/acestep/models/common/dcw_primitives.py _btc_to_bctr ;;q!  ' ' ))cBURSS5R5$)zInverse of :func:`_btc_to_bct`.rrrr s r _bct_to_btcrrrc$US:XaU$XX- --$)uPixel/latent-space differential correction (no wavelet transform). Matches the ``dcw_pix`` baseline in the DCW reference code — corrects directly in latent space. Useful as an ablation and as the fallback on platforms without ``pytorch_wavelets``. )r yscalers r dcw_pixr!s }  rcV[R"URURU5nUcgUupE[ UR [ R55n[ UR [ R55nU"U5upU"U5upXXXPRS4$)uURun DWT on both latents. Returns ``(xl, xh, yl, yh, iwt, out_T)`` or ``None`` if the optional ``pytorch_wavelets`` dependency is missing. ``out_T`` is the original time length of ``x`` — we slice the IDWT output back to this length because ``pytorch_wavelets`` pads odd-T inputs up to the next even value before running the filter bank. Nr) rgetdevicedtypertotorchfloat32shape) r rwaveletmodulesdwtiwtx_bcty_bctxlxhylyhs r _dwt_pairr+-s!''7;GHC U]]+ ,E U]]+ ,E ZFB ZFB 23 **rcUS:XaU$[XU5nUcU$UupVpxpXRXW- --nU "XV45n [U SS2SS2SU 245RURS9$)uUApply differential correction to the low-frequency sub-band only. Implements Eq. 18 / 20 of the DCW paper: * ``xL, xH = DWT(x)``; ``yL, yH = DWT(y)`` * ``xL ← xL + scaler · (xL − yL)`` * ``x_new = IDWT(xL, xH)`` Args: x: Current latent ``x_next`` after the sampler step, shape ``[B, T, C]``. y: Predicted clean sample ``denoised = x − v · t``, shape ``[B, T, C]``. scaler: Correction strength. ``0`` short-circuits to identity. wavelet: PyWavelets basis, e.g. ``"haar"`` / ``"db4"`` / ``"sym8"``. Returns: Corrected latent with the same shape and dtype as ``x``. rNr)r+rrr) r rrr!pairr'r(r)_yhr$out_Tx_news r dcw_lowr2Bsx(} Q7 #D |"&BBS  B ME uQ6E6\* + . .QWW . ==rcUS:XaU$[XU5nUcU$UupVpxp[XhSS9V V s/sHupXX- --PM n n n U "X]45n[USS2SS2SU 245RURS9$s sn n f)zBApply differential correction to the high-frequency sub-band only.rNTstrictr-r+ziprrr)r rrr!r.r'r(_ylr*r$r0xhiyhixh_newr1s r dcw_highr<as} Q7 #D |"&BCS9A"QU>V W>V(#c39-->V W ME uQ6E6\* + . .QWW . ==XsB)r torch.TensorreturnrA)r rArrArfloatrBrA)r rArrAr!str)haar) r rArrArrCr!rDrBrA) r rArrAr>rCr?rCr!rDrBrA) __doc__ __future__rr dcw_loaderrrrrr+r2r<r@rrr rIs # %* *  +,EK>>$>.3>>A>>@EK > >$ >.3 >>A > >( >>>> >  >  >r