in ddim, noise should be predicted after x0 is maybe clipped, thanks to @lukovnikov for pointing this out in another repository

README.md

@@ -1126,6 +1126,7 @@ For detailed information on training the diffusion prior, please refer to the [d
 - [x] add inpainting ability using resampler from repaint paper https://arxiv.org/abs/2201.09865
 - [x] add the final combination of upsample feature maps, used in unet squared, seems to have an effect in local experiments
 - [ ] consider elucidated dalle2 https://arxiv.org/abs/2206.00364
+- [ ] add simple outpainting, text-guided 2x size the image for starters
 - [ ] interface out the vqgan-vae so a pretrained one can be pulled off the shelf to validate latent diffusion + DALL-E2
 
 ## Citations

dalle2_pytorch/dalle2_pytorch.py

@@ -1070,7 +1070,7 @@ class DiffusionPriorNetwork(nn.Module):
 
         null_text_embeds = self.null_text_embeds.to(text_embed.dtype)
 
-        text_embeds = torch.where(
+        text_embed = torch.where(
             text_keep_mask,
             text_embed,
             null_text_embeds
@@ -1259,7 +1259,7 @@ class DiffusionPrior(nn.Module):
     def p_sample_loop_ddim(self, shape, text_cond, *, timesteps, eta = 1., cond_scale = 1.):
         batch, device, alphas, total_timesteps = shape[0], self.device, self.noise_scheduler.alphas_cumprod_prev, self.noise_scheduler.num_timesteps
 
-        times = torch.linspace(0., total_timesteps, steps = timesteps + 2)[:-1]
+        times = torch.linspace(-1., total_timesteps, steps = timesteps + 1)[:-1]
 
         times = list(reversed(times.int().tolist()))
         time_pairs = list(zip(times[:-1], times[1:]))
@@ -1281,12 +1281,14 @@ class DiffusionPrior(nn.Module):
 
             pred = self.net.forward_with_cond_scale(image_embed, time_cond, self_cond = self_cond, cond_scale = cond_scale, **text_cond)
 
+            # derive x0
+
             if self.predict_x_start:
                 x_start = pred
-                pred_noise = self.noise_scheduler.predict_noise_from_start(image_embed, t = time_cond, x0 = pred)
             else:
-                x_start = self.noise_scheduler.predict_start_from_noise(image_embed, t = time_cond, noise = pred)
-                pred_noise = pred
+                x_start = self.noise_scheduler.predict_start_from_noise(image_embed, t = time_cond, noise = pred_noise)
+
+            # clip x0 before maybe predicting noise
 
             if not self.predict_x_start:
                 x_start.clamp_(-1., 1.)
@@ -1294,6 +1296,17 @@ class DiffusionPrior(nn.Module):
             if self.predict_x_start and self.sampling_clamp_l2norm:
                 x_start = self.l2norm_clamp_embed(x_start)
 
+            # predict noise
+
+            if self.predict_x_start:
+                pred_noise = self.noise_scheduler.predict_noise_from_start(image_embed, t = time_cond, x0 = x_start)
+            else:
+                pred_noise = pred
+
+            if time_next < 0:
+                image_embed = x_start
+                continue
+
             c1 = eta * ((1 - alpha / alpha_next) * (1 - alpha_next) / (1 - alpha)).sqrt()
             c2 = ((1 - alpha_next) - torch.square(c1)).sqrt()
             noise = torch.randn_like(image_embed) if time_next > 0 else 0.
@@ -2845,12 +2858,13 @@ class Decoder(nn.Module):
         inpaint_mask = None,
         inpaint_resample_times = 5
     ):
-        batch, device, total_timesteps, alphas, eta = shape[0], self.device, noise_scheduler.num_timesteps, noise_scheduler.alphas_cumprod_prev, self.ddim_sampling_eta
+        batch, device, total_timesteps, alphas, eta = shape[0], self.device, noise_scheduler.num_timesteps, noise_scheduler.alphas_cumprod, self.ddim_sampling_eta
 
         times = torch.linspace(0., total_timesteps, steps = timesteps + 2)[:-1]
 
         times = list(reversed(times.int().tolist()))
         time_pairs = list(zip(times[:-1], times[1:]))
+        time_pairs = list(filter(lambda t: t[0] > t[1], time_pairs))
 
         is_inpaint = exists(inpaint_image)
         resample_times = inpaint_resample_times if is_inpaint else 1
@@ -2892,16 +2906,25 @@ class Decoder(nn.Module):
 
                 pred, _ = self.parse_unet_output(learned_variance, unet_output)
 
+                # predict x0
+
                 if predict_x_start:
                     x_start = pred
-                    pred_noise = noise_scheduler.predict_noise_from_start(img, t = time_cond, x0 = pred)
                 else:
                     x_start = noise_scheduler.predict_start_from_noise(img, t = time_cond, noise = pred)
-                    pred_noise = pred
+
+                # maybe clip x0
 
                 if clip_denoised:
                     x_start = self.dynamic_threshold(x_start)
 
+                # predict noise
+
+                if predict_x_start:
+                    pred_noise = noise_scheduler.predict_noise_from_start(img, t = time_cond, x0 = pred)
+                else:
+                    pred_noise = pred
+
                 c1 = eta * ((1 - alpha / alpha_next) * (1 - alpha_next) / (1 - alpha)).sqrt()
                 c2 = ((1 - alpha_next) - torch.square(c1)).sqrt()
                 noise = torch.randn_like(img) if not is_last_timestep else 0.

dalle2_pytorch/version.py

@@ -1 +1 @@
-__version__ = '1.10.1'
+__version__ = '1.10.5'