tiled encoding

mochi_preview/t2v_synth_mochi.py

@@ -323,6 +323,5 @@ class T2VSynthMochiModel:
                 comfy_pbar.update(1)
        
         self.dit.to(self.offload_device)
-        #samples = unnormalize_latents(z.float(), self.vae_mean, self.vae_std)
         logging.info(f"samples shape: {z.shape}")
         return z

mochi_preview/vae/model.py

@@ -899,6 +899,11 @@ class Encoder(nn.Module):
         assert logvar.shape == means.shape
         assert means.size(1) == self.latent_dim
 
+        noise = torch.randn(means.shape, device=means.device, dtype=means.dtype, generator=None)
+
+        # Just Gaussian sample with no scaling of variance.
+        return noise * torch.exp(logvar * 0.5) + means
+
         return LatentDistribution(means, logvar)
 
 

nodes.py

@@ -43,7 +43,8 @@ def linear_quadratic_schedule(num_steps, threshold_noise, linear_steps=None):
     sigma_schedule = linear_sigma_schedule + quadratic_sigma_schedule + [1.0]
     sigma_schedule = [1.0 - x for x in sigma_schedule]
     return sigma_schedule
-   
+
+#region ModelLoading
 class DownloadAndLoadMochiModel:
     @classmethod
     def INPUT_TYPES(s):
@@ -358,7 +359,8 @@ class MochiVAEEncoderLoader:
             encoder = torch.compile(encoder, fullgraph=torch_compile_args["fullgraph"], mode=torch_compile_args["mode"], dynamic=False, backend=torch_compile_args["backend"])
     
         return (encoder,)
-    
+#endregion
+
 class MochiTextEncode:
     @classmethod
     def INPUT_TYPES(s):
@@ -412,7 +414,7 @@ class MochiTextEncode:
         }
         return (t5_embeds, clip,)
     
-
+#region Sampler
 class MochiSampler:
     @classmethod
     def INPUT_TYPES(s):
@@ -427,7 +429,6 @@ class MochiSampler:
                 "steps": ("INT", {"default": 50, "min": 2}),
                 "cfg": ("FLOAT", {"default": 4.5, "min": 0.0, "max": 30.0, "step": 0.01}),
                 "seed": ("INT", {"default": 0, "min": 0, "max": 0xffffffffffffffff}),
-                #"batch_cfg": ("BOOLEAN", {"default": False, "tooltip": "Enable batched cfg"}),
             },
             "optional": {
                 "cfg_schedule": ("FLOAT", {"forceInput": True, "tooltip": "Override cfg schedule with a list of ints"}),
@@ -489,7 +490,9 @@ class MochiSampler:
         mm.soft_empty_cache()
 
         return ({"samples": latents},)
-    
+#endregion    
+#region Latents
+
 class MochiDecode:
     @classmethod
     def INPUT_TYPES(s):
@@ -688,12 +691,18 @@ class MochiDecodeSpatialTiling:
 
         return (frames,)
     
+
 class MochiImageEncode:
     @classmethod
     def INPUT_TYPES(s):
         return {"required": {
             "encoder": ("MOCHIVAE",),
             "images": ("IMAGE", ),
+            "enable_vae_tiling": ("BOOLEAN", {"default": False, "tooltip": "Drastically reduces memory use but may introduce seams"}),
+            "num_tiles_w": ("INT", {"default": 4, "min": 2, "max": 64, "step": 2, "tooltip": "Number of horizontal tiles"}),
+            "num_tiles_h": ("INT", {"default": 4, "min": 2, "max": 64, "step": 2, "tooltip": "Number of vertical tiles"}),
+            "overlap": ("INT", {"default": 16, "min": 0, "max": 256, "step": 1, "tooltip": "Number of pixel of overlap between adjacent tiles"}),
+            "min_block_size": ("INT", {"default": 1, "min": 1, "max": 256, "step": 1, "tooltip": "Minimum number of pixels in each dimension when subdividing"}),
             },
         }
 
@@ -702,23 +711,25 @@ class MochiImageEncode:
     FUNCTION = "decode"
     CATEGORY = "MochiWrapper"
 
-    def decode(self, encoder, images):
+    def decode(self, encoder, images, enable_vae_tiling, num_tiles_w, num_tiles_h, overlap, min_block_size):
         device = mm.get_torch_device()
         offload_device = mm.unet_offload_device()
         intermediate_device = mm.intermediate_device()
-
+        from .mochi_preview.vae.model import apply_tiled
         B, H, W, C = images.shape
         
         images = images.unsqueeze(0) * 2 - 1
         images = rearrange(images, "t b h w c -> t c b h w")
         images = images.to(encoder.dtype).to(device)
         print(images.shape)
-        
         encoder.to(device)
         print("images before encoding", images.shape)
         with torch.autocast(mm.get_autocast_device(device), dtype=encoder.dtype):
-                video = add_fourier_features(images)          
+                video = add_fourier_features(images)
-                latents = encoder(video).sample()
+                if enable_vae_tiling:     
+                    latents = apply_tiled(encoder, video, num_tiles_w = num_tiles_w, num_tiles_h = num_tiles_h, overlap=overlap, min_block_size=min_block_size)
+                else:
+                    latents = encoder(video)
         latents = vae_latents_to_dit_latents(latents)
         print("encoder output",latents.shape)
 
@@ -784,6 +795,8 @@ class MochiLatentPreview:
         
         return (latent_images.float().cpu(),)    
 
+#endregion
+#region NodeMappings
 NODE_CLASS_MAPPINGS = {
     "DownloadAndLoadMochiModel": DownloadAndLoadMochiModel,
     "MochiSampler": MochiSampler,