The most powerful and modular diffusion model GUI, api and backend with a graph/nodes interface. The fastest local inference engine in the world.
Qwen rotary embeddings should now match reference code. (#9437)
comfyanonymous7cd2c4bd6ab2
diff --git a/comfy/ldm/qwen_image/model.py b/comfy/ldm/qwen_image/model.pyindex a3c726299..bf3940313 100644--- a/comfy/ldm/qwen_image/model.py+++ b/comfy/ldm/qwen_image/model.py@@ -349,8 +349,8 @@ class QwenImageTransformer2DModel(nn.Module):img_ids = torch.zeros((h_len, w_len, 3), device=x.device, dtype=x.dtype)img_ids[:, :, 0] = img_ids[:, :, 1] + index- img_ids[:, :, 1] = img_ids[:, :, 1] + torch.linspace(h_offset, h_len - 1 + h_offset, steps=h_len, device=x.device, dtype=x.dtype).unsqueeze(1)- img_ids[:, :, 2] = img_ids[:, :, 2] + torch.linspace(w_offset, w_len - 1 + w_offset, steps=w_len, device=x.device, dtype=x.dtype).unsqueeze(0)+ img_ids[:, :, 1] = img_ids[:, :, 1] + torch.linspace(h_offset, h_len - 1 + h_offset, steps=h_len, device=x.device, dtype=x.dtype).unsqueeze(1) - (h_len // 2)+ img_ids[:, :, 2] = img_ids[:, :, 2] + torch.linspace(w_offset, w_len - 1 + w_offset, steps=w_len, device=x.device, dtype=x.dtype).unsqueeze(0) - (w_len // 2)return hidden_states, repeat(img_ids, "h w c -> b (h w) c", b=bs), orig_shapedef forward(@@ -396,7 +396,7 @@ class QwenImageTransformer2DModel(nn.Module):hidden_states = torch.cat([hidden_states, kontext], dim=1)img_ids = torch.cat([img_ids, kontext_ids], dim=1)- txt_start = round(max(((x.shape[-1] + (self.patch_size // 2)) // self.patch_size), ((x.shape[-2] + (self.patch_size // 2)) // self.patch_size)))+ txt_start = round(max(((x.shape[-1] + (self.patch_size // 2)) // self.patch_size) // 2, ((x.shape[-2] + (self.patch_size // 2)) // self.patch_size) // 2))txt_ids = torch.linspace(txt_start, txt_start + context.shape[1], steps=context.shape[1], device=x.device, dtype=x.dtype).reshape(1, -1, 1).repeat(x.shape[0], 1, 3)ids = torch.cat((txt_ids, img_ids), dim=1)image_rotary_emb = self.pe_embedder(ids).squeeze(1).unsqueeze(2).to(x.dtype)