Skip to content

The most powerful and modular diffusion model GUI, api and backend with a graph/nodes interface. The fastest local inference engine in the world.

Support some z image lora formats. (#10978)

comfyanonymous52a32e2b323b
diff --git a/comfy/lora.py b/comfy/lora.py
index 36d26293a..360cd128f 100644
--- a/comfy/lora.py
+++ b/comfy/lora.py
@@ -313,6 +313,14 @@ def model_lora_keys_unet(model, key_map={}):
key_map["transformer.{}".format(key_lora)] = k
key_map["lycoris_{}".format(key_lora.replace(".", "_"))] = k #SimpleTuner lycoris format
+ if isinstance(model, comfy.model_base.Lumina2):
+ diffusers_keys = comfy.utils.z_image_to_diffusers(model.model_config.unet_config, output_prefix="diffusion_model.")
+ for k in diffusers_keys:
+ to = diffusers_keys[k]
+ key_lora = k[:-len(".weight")]
+ key_map["diffusion_model.{}".format(key_lora)] = to
+ key_map["lycoris_{}".format(key_lora.replace(".", "_"))] = to
+
return key_map
diff --git a/comfy/utils.py b/comfy/utils.py
index 4bd281057..21bd6e8cf 100644
--- a/comfy/utils.py
+++ b/comfy/utils.py
@@ -675,6 +675,51 @@ def flux_to_diffusers(mmdit_config, output_prefix=""):
return key_map
+def z_image_to_diffusers(mmdit_config, output_prefix=""):
+ n_layers = mmdit_config.get("n_layers", 0)
+ hidden_size = mmdit_config.get("dim", 0)
+
+ key_map = {}
+
+ for index in range(n_layers):
+ prefix_from = "layers.{}".format(index)
+ prefix_to = "{}layers.{}".format(output_prefix, index)
+
+ for end in ("weight", "bias"):
+ k = "{}.attention.".format(prefix_from)
+ qkv = "{}.attention.qkv.{}".format(prefix_to, end)
+
+ key_map["{}to_q.{}".format(k, end)] = (qkv, (0, 0, hidden_size))
+ key_map["{}to_k.{}".format(k, end)] = (qkv, (0, hidden_size, hidden_size))
+ key_map["{}to_v.{}".format(k, end)] = (qkv, (0, hidden_size * 2, hidden_size))
+
+ block_map = {
+ "attention.norm_q.weight": "attention.q_norm.weight",
+ "attention.norm_k.weight": "attention.k_norm.weight",
+ "attention.to_out.0.weight": "attention.out.weight",
+ "attention.to_out.0.bias": "attention.out.bias",
+ }
+
+ for k in block_map:
+ key_map["{}.{}".format(prefix_from, k)] = "{}.{}".format(prefix_to, block_map[k])
+
+ MAP_BASIC = {
+ # Final layer
+ ("final_layer.linear.weight", "all_final_layer.2-1.linear.weight"),
+ ("final_layer.linear.bias", "all_final_layer.2-1.linear.bias"),
+ ("final_layer.adaLN_modulation.1.weight", "all_final_layer.2-1.adaLN_modulation.1.weight"),
+ ("final_layer.adaLN_modulation.1.bias", "all_final_layer.2-1.adaLN_modulation.1.bias"),
+ # X embedder
+ ("x_embedder.weight", "all_x_embedder.2-1.weight"),
+ ("x_embedder.bias", "all_x_embedder.2-1.bias"),
+ }
+
+ for k in MAP_BASIC:
+ key_map[k[1]] = "{}{}".format(output_prefix, k[0])
+
+ return key_map
+
+
def repeat_to_batch_size(tensor, batch_size, dim=0):
if tensor.shape[dim] > batch_size:
return tensor.narrow(dim, 0, batch_size)