The most powerful and modular diffusion model GUI, api and backend with a graph/nodes interface. The fastest local inference engine in the world.
Make lumina model work with any latent resolution.
comfyanonymous4027466c802d
diff --git a/comfy/ldm/lumina/model.py b/comfy/ldm/lumina/model.pyindex 3292bd2f0..ccd5d2c0e 100644--- a/comfy/ldm/lumina/model.py+++ b/comfy/ldm/lumina/model.py@@ -6,6 +6,7 @@ from typing import List, Optional, Tupleimport torchimport torch.nn as nnimport torch.nn.functional as F+import comfy.ldm.common_ditfrom comfy.ldm.modules.diffusionmodules.mmdit import TimestepEmbedder, RMSNormfrom comfy.ldm.modules.attention import optimized_attention_masked@@ -594,6 +595,8 @@ class NextDiT(nn.Module):t = 1.0 - timestepscap_feats = contextcap_mask = attention_mask+ bs, c, h, w = x.shape+ x = comfy.ldm.common_dit.pad_to_patch_size(x, (self.patch_size, self.patch_size))"""Forward pass of NextDiT.t: (N,) tensor of diffusion timesteps@@ -613,7 +616,7 @@ class NextDiT(nn.Module):x = layer(x, mask, freqs_cis, adaln_input)x = self.final_layer(x, adaln_input)- x = self.unpatchify(x, img_size, cap_size, return_tensor=x_is_tensor)+ x = self.unpatchify(x, img_size, cap_size, return_tensor=x_is_tensor)[:,:,:h,:w]return -x