add immiscible diffusion

Author: lucidrains

README.md

@@ -355,3 +355,14 @@ You could consider adding a suitable metric to the training loop yourself after
     url     = {https://api.semanticscholar.org/CorpusID:265659032}
 }
 ```
+
+```bibtex
+@article{Li2024ImmiscibleDA,
+    title   = {Immiscible Diffusion: Accelerating Diffusion Training with Noise Assignment},
+    author  = {Yiheng Li and Heyang Jiang and Akio Kodaira and Masayoshi Tomizuka and Kurt Keutzer and Chenfeng Xu},
+    journal = {ArXiv},
+    year    = {2024},
+    volume  = {abs/2406.12303},
+    url     = {https://api.semanticscholar.org/CorpusID:270562607}
+}
+```

denoising_diffusion_pytorch/denoising_diffusion_pytorch.py

@@ -20,6 +20,8 @@
 from einops import rearrange, reduce, repeat
 from einops.layers.torch import Rearrange
 
+from scipy.optimize import linear_sum_assignment
+
 from PIL import Image
 from tqdm.auto import tqdm
 from ema_pytorch import EMA
@@ -488,7 +490,8 @@ def __init__(
         auto_normalize = True,
         offset_noise_strength = 0.,  # https://www.crosslabs.org/blog/diffusion-with-offset-noise
         min_snr_loss_weight = False, # https://arxiv.org/abs/2303.09556
-        min_snr_gamma = 5
+        min_snr_gamma = 5,
+        immiscible = False
     ):
         super().__init__()
         assert not (type(self) == GaussianDiffusion and model.channels != model.out_dim)
@@ -564,6 +567,10 @@ def __init__(
         register_buffer('posterior_mean_coef1', betas * torch.sqrt(alphas_cumprod_prev) / (1. - alphas_cumprod))
         register_buffer('posterior_mean_coef2', (1. - alphas_cumprod_prev) * torch.sqrt(alphas) / (1. - alphas_cumprod))
 
+        # immiscible diffusion
+
+        self.immiscible = immiscible
+
         # offset noise strength - in blogpost, they claimed 0.1 was ideal
 
         self.offset_noise_strength = offset_noise_strength
@@ -759,10 +766,20 @@ def interpolate(self, x1, x2, t = None, lam = 0.5):
 
         return img
 
+    def noise_assignment(self, x_start, noise):
+        x_start, noise = tuple(rearrange(t, 'b ... -> b (...)') for t in (x_start, noise))
+        dist = torch.cdist(x_start, noise)
+        _, assign = linear_sum_assignment(dist.cpu())
+        return torch.from_numpy(assign).to(dist.device)
+
     @autocast(enabled = False)
     def q_sample(self, x_start, t, noise = None):
         noise = default(noise, lambda: torch.randn_like(x_start))
 
+        if self.immiscible:
+            assign = self.noise_assignment(x_start, noise)
+            noise = noise[assign]
+
         return (
             extract(self.sqrt_alphas_cumprod, t, x_start.shape) * x_start +
             extract(self.sqrt_one_minus_alphas_cumprod, t, x_start.shape) * noise

denoising_diffusion_pytorch/version.py

@@ -1 +1 @@
-__version__ = '2.0.12'
+__version__ = '2.0.15'

setup.py

@@ -23,6 +23,7 @@
     'numpy',
     'pillow',
     'pytorch-fid',
+    'scipy',
     'torch',
     'torchvision',
     'tqdm'