Merge 2d3a8d523d into e89b22993a

comfy/latent_formats.py

@@ -8,7 +8,6 @@ class LatentFormat:
     latent_rgb_factors_bias = None
     latent_rgb_factors_reshape = None
     taesd_decoder_name = None
-    spacial_downscale_ratio = 8
 
     def process_in(self, latent):
         return latent * self.scale_factor
@@ -182,7 +181,6 @@ class Flux(SD3):
 
 class Flux2(LatentFormat):
     latent_channels = 128
-    spacial_downscale_ratio = 16
 
     def __init__(self):
         self.latent_rgb_factors =[
@@ -751,7 +749,6 @@ class ACEAudio(LatentFormat):
 
 class ChromaRadiance(LatentFormat):
     latent_channels = 3
-    spacial_downscale_ratio = 1
 
     def __init__(self):
         self.latent_rgb_factors = [

comfy/ldm/wan/vae.py

@@ -5,7 +5,7 @@ import torch
 import torch.nn as nn
 import torch.nn.functional as F
 from einops import rearrange
-from comfy.ldm.modules.diffusionmodules.model import vae_attention, torch_cat_if_needed
+from comfy.ldm.modules.diffusionmodules.model import vae_attention
 
 import comfy.ops
 ops = comfy.ops.disable_weight_init
@@ -20,29 +20,22 @@ class CausalConv3d(ops.Conv3d):
 
     def __init__(self, *args, **kwargs):
         super().__init__(*args, **kwargs)
-        self._padding = 2 * self.padding[0]
-        self.padding = (0, self.padding[1], self.padding[2])
+        self._padding = (self.padding[2], self.padding[2], self.padding[1],
+                         self.padding[1], 2 * self.padding[0], 0)
+        self.padding = (0, 0, 0)
 
     def forward(self, x, cache_x=None, cache_list=None, cache_idx=None):
         if cache_list is not None:
             cache_x = cache_list[cache_idx]
             cache_list[cache_idx] = None
 
-        if cache_x is None and x.shape[2] == 1:
-            #Fast path - the op will pad for use by truncating the weight
-            #and save math on a pile of zeros.
-            return super().forward(x, autopad="causal_zero")
-
-        if self._padding > 0:
-            padding_needed = self._padding
-            if cache_x is not None:
-                cache_x = cache_x.to(x.device)
-                padding_needed = max(0, padding_needed - cache_x.shape[2])
-            padding_shape = list(x.shape)
-            padding_shape[2] = padding_needed
-            padding = torch.zeros(padding_shape, device=x.device, dtype=x.dtype)
-            x = torch_cat_if_needed([padding, cache_x, x], dim=2)
+        padding = list(self._padding)
+        if cache_x is not None and self._padding[4] > 0:
+            cache_x = cache_x.to(x.device)
+            x = torch.cat([cache_x, x], dim=2)
+            padding[4] -= cache_x.shape[2]
             del cache_x
+        x = F.pad(x, padding)
 
         return super().forward(x)
 

comfy/ops.py

@@ -203,9 +203,7 @@ class disable_weight_init:
         def reset_parameters(self):
             return None
 
-        def _conv_forward(self, input, weight, bias, autopad=None, *args, **kwargs):
-            if autopad == "causal_zero":
-                weight = weight[:, :, -input.shape[2]:, :, :]
+        def _conv_forward(self, input, weight, bias, *args, **kwargs):
             if NVIDIA_MEMORY_CONV_BUG_WORKAROUND and weight.dtype in (torch.float16, torch.bfloat16):
                 out = torch.cudnn_convolution(input, weight, self.padding, self.stride, self.dilation, self.groups, benchmark=False, deterministic=False, allow_tf32=True)
                 if bias is not None:
@@ -214,15 +212,15 @@ class disable_weight_init:
             else:
                 return super()._conv_forward(input, weight, bias, *args, **kwargs)
 
-        def forward_comfy_cast_weights(self, input, autopad=None):
+        def forward_comfy_cast_weights(self, input):
             weight, bias, offload_stream = cast_bias_weight(self, input, offloadable=True)
-            x = self._conv_forward(input, weight, bias, autopad=autopad)
+            x = self._conv_forward(input, weight, bias)
             uncast_bias_weight(self, weight, bias, offload_stream)
             return x
 
         def forward(self, *args, **kwargs):
             run_every_op()
-            if self.comfy_cast_weights or len(self.weight_function) > 0 or len(self.bias_function) > 0 or "autopad" in kwargs:
+            if self.comfy_cast_weights or len(self.weight_function) > 0 or len(self.bias_function) > 0:
                 return self.forward_comfy_cast_weights(*args, **kwargs)
             else:
                 return super().forward(*args, **kwargs)

comfy/sample.py

@@ -37,18 +37,12 @@ def prepare_noise(latent_image, seed, noise_inds=None):
 
     return noises
 
-def fix_empty_latent_channels(model, latent_image, downscale_ratio_spacial=None):
+def fix_empty_latent_channels(model, latent_image):
     if latent_image.is_nested:
         return latent_image
     latent_format = model.get_model_object("latent_format") #Resize the empty latent image so it has the right number of channels
-    if torch.count_nonzero(latent_image) == 0:
-        if latent_format.latent_channels != latent_image.shape[1]:
-            latent_image = comfy.utils.repeat_to_batch_size(latent_image, latent_format.latent_channels, dim=1)
-        if downscale_ratio_spacial is not None:
-            if downscale_ratio_spacial != latent_format.spacial_downscale_ratio:
-                ratio = downscale_ratio_spacial / latent_format.spacial_downscale_ratio
-                latent_image = comfy.utils.common_upscale(latent_image, round(latent_image.shape[-1] * ratio), round(latent_image.shape[-2] * ratio), "nearest-exact", crop="disabled")
-
+    if latent_format.latent_channels != latent_image.shape[1] and torch.count_nonzero(latent_image) == 0:
+        latent_image = comfy.utils.repeat_to_batch_size(latent_image, latent_format.latent_channels, dim=1)
     if latent_format.latent_dimensions == 3 and latent_image.ndim == 4:
         latent_image = latent_image.unsqueeze(2)
     return latent_image

comfy_extras/nodes_custom_sampler.py

@@ -741,7 +741,7 @@ class SamplerCustom(io.ComfyNode):
         latent = latent_image
         latent_image = latent["samples"]
         latent = latent.copy()
-        latent_image = comfy.sample.fix_empty_latent_channels(model, latent_image, latent.get("downscale_ratio_spacial", None))
+        latent_image = comfy.sample.fix_empty_latent_channels(model, latent_image)
         latent["samples"] = latent_image
 
         if not add_noise:
@@ -760,7 +760,6 @@ class SamplerCustom(io.ComfyNode):
         samples = comfy.sample.sample_custom(model, noise, cfg, sampler, sigmas, positive, negative, latent_image, noise_mask=noise_mask, callback=callback, disable_pbar=disable_pbar, seed=noise_seed)
 
         out = latent.copy()
-        out.pop("downscale_ratio_spacial", None)
         out["samples"] = samples
         if "x0" in x0_output:
             x0_out = model.model.process_latent_out(x0_output["x0"].cpu())
@@ -940,7 +939,7 @@ class SamplerCustomAdvanced(io.ComfyNode):
         latent = latent_image
         latent_image = latent["samples"]
         latent = latent.copy()
-        latent_image = comfy.sample.fix_empty_latent_channels(guider.model_patcher, latent_image, latent.get("downscale_ratio_spacial", None))
+        latent_image = comfy.sample.fix_empty_latent_channels(guider.model_patcher, latent_image)
         latent["samples"] = latent_image
 
         noise_mask = None
@@ -955,7 +954,6 @@ class SamplerCustomAdvanced(io.ComfyNode):
         samples = samples.to(comfy.model_management.intermediate_device())
 
         out = latent.copy()
-        out.pop("downscale_ratio_spacial", None)
         out["samples"] = samples
         if "x0" in x0_output:
             x0_out = guider.model_patcher.model.process_latent_out(x0_output["x0"].cpu())

comfy_extras/nodes_sd3.py

@@ -55,7 +55,7 @@ class EmptySD3LatentImage(io.ComfyNode):
     @classmethod
     def execute(cls, width, height, batch_size=1) -> io.NodeOutput:
         latent = torch.zeros([batch_size, 16, height // 8, width // 8], device=comfy.model_management.intermediate_device())
-        return io.NodeOutput({"samples": latent, "downscale_ratio_spacial": 8})
+        return io.NodeOutput({"samples":latent})
 
     generate = execute  # TODO: remove
 

nodes.py

@@ -1230,7 +1230,7 @@ class EmptyLatentImage:
 
     def generate(self, width, height, batch_size=1):
         latent = torch.zeros([batch_size, 4, height // 8, width // 8], device=self.device)
-        return ({"samples": latent, "downscale_ratio_spacial": 8}, )
+        return ({"samples":latent}, )
 
 
 class LatentFromBatch:
@@ -1538,7 +1538,7 @@ class SetLatentNoiseMask:
 
 def common_ksampler(model, seed, steps, cfg, sampler_name, scheduler, positive, negative, latent, denoise=1.0, disable_noise=False, start_step=None, last_step=None, force_full_denoise=False):
     latent_image = latent["samples"]
-    latent_image = comfy.sample.fix_empty_latent_channels(model, latent_image, latent.get("downscale_ratio_spacial", None))
+    latent_image = comfy.sample.fix_empty_latent_channels(model, latent_image)
 
     if disable_noise:
         noise = torch.zeros(latent_image.size(), dtype=latent_image.dtype, layout=latent_image.layout, device="cpu")
@@ -1556,7 +1556,6 @@ def common_ksampler(model, seed, steps, cfg, sampler_name, scheduler, positive,
                                   denoise=denoise, disable_noise=disable_noise, start_step=start_step, last_step=last_step,
                                   force_full_denoise=force_full_denoise, noise_mask=noise_mask, callback=callback, disable_pbar=disable_pbar, seed=seed)
     out = latent.copy()
-    out.pop("downscale_ratio_spacial", None)
     out["samples"] = samples
     return (out, )