Spaces:
Running
Running
| # S-Lab License 1.0 | |
| # Copyright 2023 S-Lab | |
| # Redistribution and use for non-commercial purpose in source and binary forms, with or without modification, are permitted provided that the following conditions are met: | |
| # 1. Redistributions of source code must retain the above copyright notice, this list of conditions and the following disclaimer. | |
| # 2. Redistributions in binary form must reproduce the above copyright notice, this list of conditions and the following disclaimer in the documentation and/or other materials provided with the distribution. | |
| # 3. Neither the name of the copyright holder nor the names of its contributors may be used to endorse or promote products derived from this software without specific prior written permission. | |
| # THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. | |
| # 4. In the event that redistribution and/or use for commercial purpose in source or binary forms, with or without modification is required, please contact the contributor(s) of the work. | |
| # Code has been modified from https://github.com/TianxingWu/FreeInit | |
| import torch | |
| import torch.fft as fft | |
| import math | |
| class FreeInitFilter: | |
| GAUSSIAN = "gaussian" | |
| IDEAL = "ideal" | |
| BOX = "box" | |
| BUTTERWORTH = "butterworth" | |
| LIST = [GAUSSIAN, BUTTERWORTH, IDEAL, BOX] | |
| def freq_mix_3d(x, noise, LPF): | |
| """ | |
| Noise reinitialization. | |
| Args: | |
| x: diffused latent | |
| noise: randomly sampled noise | |
| LPF: low pass filter | |
| """ | |
| # FFT | |
| x_freq = fft.fftn(x, dim=(-4, -2, -1)) | |
| x_freq = fft.fftshift(x_freq, dim=(-4, -2, -1)) | |
| noise_freq = fft.fftn(noise, dim=(-4, -2, -1)) | |
| noise_freq = fft.fftshift(noise_freq, dim=(-4, -2, -1)) | |
| # frequency mix | |
| HPF = 1 - LPF | |
| x_freq_low = x_freq * LPF | |
| noise_freq_high = noise_freq * HPF | |
| x_freq_mixed = x_freq_low + noise_freq_high # mix in freq domain | |
| # IFFT | |
| x_freq_mixed = fft.ifftshift(x_freq_mixed, dim=(-4, -2, -1)) | |
| x_mixed = fft.ifftn(x_freq_mixed, dim=(-4, -2, -1)).real | |
| return x_mixed | |
| def get_freq_filter(shape, device, filter_type, n, d_s, d_t): | |
| """ | |
| Form the frequency filter for noise reinitialization. | |
| Args: | |
| shape: shape of latent (T, C, H, W) | |
| filter_type: type of the freq filter | |
| n: (only for butterworth) order of the filter, larger n ~ ideal, smaller n ~ gaussian | |
| d_s: normalized stop frequency for spatial dimensions (0.0-1.0) | |
| d_t: normalized stop frequency for temporal dimension (0.0-1.0) | |
| """ | |
| if filter_type == FreeInitFilter.GAUSSIAN: | |
| return gaussian_low_pass_filter(shape=shape, d_s=d_s, d_t=d_t).to(device) | |
| elif filter_type == FreeInitFilter.IDEAL: | |
| return ideal_low_pass_filter(shape=shape, d_s=d_s, d_t=d_t).to(device) | |
| elif filter_type == FreeInitFilter.BOX: | |
| return box_low_pass_filter(shape=shape, d_s=d_s, d_t=d_t).to(device) | |
| elif filter_type == FreeInitFilter.BUTTERWORTH: | |
| return butterworth_low_pass_filter(shape=shape, n=n, d_s=d_s, d_t=d_t).to(device) | |
| else: | |
| raise NotImplementedError | |
| def gaussian_low_pass_filter(shape, d_s=0.25, d_t=0.25): | |
| """ | |
| Compute the gaussian low pass filter mask. | |
| Args: | |
| shape: shape of the filter (volume) | |
| d_s: normalized stop frequency for spatial dimensions (0.0-1.0) | |
| d_t: normalized stop frequency for temporal dimension (0.0-1.0) | |
| """ | |
| T, H, W = shape[-4], shape[-2], shape[-1] | |
| mask = torch.zeros(shape) | |
| if d_s==0 or d_t==0: | |
| return mask | |
| for t in range(T): | |
| for h in range(H): | |
| for w in range(W): | |
| d_square = (((d_s/d_t)*(2*t/T-1))**2 + (2*h/H-1)**2 + (2*w/W-1)**2) | |
| mask[t, ..., h,w] = math.exp(-1/(2*d_s**2) * d_square) | |
| return mask | |
| def butterworth_low_pass_filter(shape, n=4, d_s=0.25, d_t=0.25): | |
| """ | |
| Compute the butterworth low pass filter mask. | |
| Args: | |
| shape: shape of the filter (volume) | |
| n: order of the filter, larger n ~ ideal, smaller n ~ gaussian | |
| d_s: normalized stop frequency for spatial dimensions (0.0-1.0) | |
| d_t: normalized stop frequency for temporal dimension (0.0-1.0) | |
| """ | |
| T, H, W = shape[-4], shape[-2], shape[-1] | |
| mask = torch.zeros(shape) | |
| if d_s==0 or d_t==0: | |
| return mask | |
| for t in range(T): | |
| for h in range(H): | |
| for w in range(W): | |
| d_square = (((d_s/d_t)*(2*t/T-1))**2 + (2*h/H-1)**2 + (2*w/W-1)**2) | |
| mask[t, ..., h,w] = 1 / (1 + (d_square / d_s**2)**n) | |
| return mask | |
| def ideal_low_pass_filter(shape, d_s=0.25, d_t=0.25): | |
| """ | |
| Compute the ideal low pass filter mask. | |
| Args: | |
| shape: shape of the filter (volume) | |
| d_s: normalized stop frequency for spatial dimensions (0.0-1.0) | |
| d_t: normalized stop frequency for temporal dimension (0.0-1.0) | |
| """ | |
| T, H, W = shape[-4], shape[-2], shape[-1] | |
| mask = torch.zeros(shape) | |
| if d_s==0 or d_t==0: | |
| return mask | |
| for t in range(T): | |
| for h in range(H): | |
| for w in range(W): | |
| d_square = (((d_s/d_t)*(2*t/T-1))**2 + (2*h/H-1)**2 + (2*w/W-1)**2) | |
| mask[t, ...,h,w] = 1 if d_square <= d_s*2 else 0 | |
| return mask | |
| def box_low_pass_filter(shape, d_s=0.25, d_t=0.25): | |
| """ | |
| Compute the ideal low pass filter mask (approximated version). | |
| Args: | |
| shape: shape of the filter (volume) | |
| d_s: normalized stop frequency for spatial dimensions (0.0-1.0) | |
| d_t: normalized stop frequency for temporal dimension (0.0-1.0) | |
| """ | |
| T, H, W = shape[-4], shape[-2], shape[-1] | |
| mask = torch.zeros(shape) | |
| if d_s==0 or d_t==0: | |
| return mask | |
| threshold_s = round(int(H // 2) * d_s) | |
| threshold_t = round(T // 2 * d_t) | |
| cframe, crow, ccol = T // 2, H // 2, W //2 | |
| mask[cframe - threshold_t:cframe + threshold_t, ..., crow - threshold_s:crow + threshold_s, ccol - threshold_s:ccol + threshold_s] = 1.0 | |
| return mask | |