Adde dnoise offset

jobs/process/TrainSliderProcess.py

@@ -17,7 +17,7 @@ from diffusers import StableDiffusionPipeline
 
 from jobs.process import BaseTrainProcess
 from toolkit.metadata import get_meta_for_safetensors
-from toolkit.train_tools import get_torch_dtype
+from toolkit.train_tools import get_torch_dtype, apply_noise_offset
 import gc
 
 import torch
@@ -38,6 +38,9 @@ def flush():
     gc.collect()
 
 
+UNET_IN_CHANNELS = 4  # Stable Diffusion の in_channels は 4 で固定。XLも同じ。
+VAE_SCALE_FACTOR = 8  # 2 ** (len(vae.config.block_out_channels) - 1) = 8
+
 class StableDiffusion:
     def __init__(self, vae, tokenizer, text_encoder, unet, noise_scheduler):
         self.vae = vae
@@ -94,6 +97,7 @@ class TrainConfig:
         self.xformers = kwargs.get('xformers', False)
         self.train_unet = kwargs.get('train_unet', True)
         self.train_text_encoder = kwargs.get('train_text_encoder', True)
+        self.noise_offset = kwargs.get('noise_offset', 0.0)
 
 
 class ModelConfig:
@@ -506,13 +510,19 @@ class TrainSliderProcess(BaseTrainProcess):
                     1, self.train_config.max_denoising_steps, (1,)
                 ).item()
 
-                latents = train_util.get_initial_latents(
-                    noise_scheduler,
-                    self.train_config.batch_size,
-                    height,
-                    width,
-                    1
-                ).to(self.device_torch, dtype=dtype)
+                # get noise
+                noise = torch.randn(
+                    (
+                        self.train_config.batch_size,
+                        UNET_IN_CHANNELS,
+                        height // VAE_SCALE_FACTOR,
+                        width // VAE_SCALE_FACTOR,
+                    ),
+                    device="cpu",
+                )
+                noise = apply_noise_offset(noise, self.train_config.noise_offset)
+                latents = noise * noise_scheduler.init_noise_sigma
+                latents = latents.to(self.device_torch, dtype=dtype)
 
                 with self.network:
                     assert self.network.is_active
@@ -673,7 +683,7 @@ class TrainSliderProcess(BaseTrainProcess):
             # end of step
             self.step_num = step
 
-        self.sample(self.step_num)
+        self.sample(self.step_num + 1)
         print("")
         self.save()
 

toolkit/train_tools.py

@@ -369,3 +369,11 @@ def sample_images(
     if cuda_rng_state is not None:
         torch.cuda.set_rng_state(cuda_rng_state)
     vae.to(org_vae_device)
+
+
+# https://www.crosslabs.org//blog/diffusion-with-offset-noise
+def apply_noise_offset(noise, noise_offset):
+    if noise_offset is None or noise_offset < 0.0000001:
+        return noise
+    noise = noise + noise_offset * torch.randn((noise.shape[0], noise.shape[1], 1, 1), device=noise.device)
+    return noise