The most powerful and modular diffusion model GUI, api and backend with a graph/nodes interface. The fastest local inference engine in the world.
Fix lumina 2 batches.
comfyanonymousa57d635c5f36
diff --git a/comfy/ldm/lumina/model.py b/comfy/ldm/lumina/model.pyindex 4eb7164d1..e4b0d34a6 100644--- a/comfy/ldm/lumina/model.py+++ b/comfy/ldm/lumina/model.py@@ -581,12 +581,13 @@ class NextDiT(nn.Module):flat_x.append(img)x = flat_xpadded_img_embed = torch.zeros(bsz, max_img_len, x[0].shape[-1], device=device, dtype=x[0].dtype)- padded_img_mask = torch.zeros(bsz, max_img_len, dtype=torch.bool, device=device)+ padded_img_mask = torch.zeros(bsz, max_img_len, dtype=dtype, device=device)for i in range(bsz):padded_img_embed[i, :l_effective_img_len[i]] = x[i]- padded_img_mask[i, :l_effective_img_len[i]] = True+ padded_img_mask[i, l_effective_img_len[i]:] = -torch.finfo(dtype).maxpadded_img_embed = self.x_embedder(padded_img_embed)+ padded_img_mask = padded_img_mask.unsqueeze(1)for layer in self.noise_refiner:padded_img_embed = layer(padded_img_embed, padded_img_mask, img_freqs_cis, t)