The most powerful and modular diffusion model GUI, api and backend with a graph/nodes interface. The fastest local inference engine in the world.
Auto reshape 2d to 3d latent for single image generation on video model.
comfyanonymousa618f768e06d
diff --git a/comfy/latent_formats.py b/comfy/latent_formats.pyindex 40234b64c..a50a70aec 100644--- a/comfy/latent_formats.py+++ b/comfy/latent_formats.py@@ -3,6 +3,7 @@ import torchclass LatentFormat:scale_factor = 1.0latent_channels = 4+ latent_dimensions = 2latent_rgb_factors = Nonelatent_rgb_factors_bias = Nonetaesd_decoder_name = None@@ -143,6 +144,7 @@ class SD3(LatentFormat):class StableAudio1(LatentFormat):latent_channels = 64+ latent_dimensions = 1class Flux(SD3):latent_channels = 16@@ -178,6 +180,7 @@ class Flux(SD3):class Mochi(LatentFormat):latent_channels = 12+ latent_dimensions = 3def __init__(self):self.scale_factor = 1.0@@ -219,6 +222,8 @@ class Mochi(LatentFormat):class LTXV(LatentFormat):latent_channels = 128+ latent_dimensions = 3+def __init__(self):self.latent_rgb_factors = [[ 1.1202e-02, -6.3815e-04, -1.0021e-02],@@ -355,6 +360,7 @@ class LTXV(LatentFormat):class HunyuanVideo(LatentFormat):latent_channels = 16+ latent_dimensions = 3scale_factor = 0.476986latent_rgb_factors = [[-0.0395, -0.0331, 0.0445],diff --git a/comfy/sample.py b/comfy/sample.pyindex 9974e0657..be5a7e246 100644--- a/comfy/sample.py+++ b/comfy/sample.py@@ -25,9 +25,11 @@ def prepare_noise(latent_image, seed, noise_inds=None):return noisesdef fix_empty_latent_channels(model, latent_image):- latent_channels = model.get_model_object("latent_format").latent_channels #Resize the empty latent image so it has the right number of channels- if latent_channels != latent_image.shape[1] and torch.count_nonzero(latent_image) == 0:- latent_image = comfy.utils.repeat_to_batch_size(latent_image, latent_channels, dim=1)+ latent_format = model.get_model_object("latent_format") #Resize the empty latent image so it has the right number of channels+ 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_imagedef prepare_sampling(model, noise_shape, positive, negative, noise_mask):