The most powerful and modular diffusion model GUI, api and backend with a graph/nodes interface. The fastest local inference engine in the world.
Implement the USO subject identity lora. (#9674)
comfyanonymous27e067ce505c
diff --git a/comfy/ldm/flux/model.py b/comfy/ldm/flux/model.pyindex 1344c3a57..1e62f4626 100644--- a/comfy/ldm/flux/model.py+++ b/comfy/ldm/flux/model.py@@ -233,12 +233,18 @@ class Flux(nn.Module):h = 0w = 0index = 0- index_ref_method = kwargs.get("ref_latents_method", "offset") == "index"+ ref_latents_method = kwargs.get("ref_latents_method", "offset")for ref in ref_latents:- if index_ref_method:+ if ref_latents_method == "index":index += 1h_offset = 0w_offset = 0+ elif ref_latents_method == "uso":+ index = 0+ h_offset = h_len * patch_size + h+ w_offset = w_len * patch_size + w+ h += ref.shape[-2]+ w += ref.shape[-1]else:index = 1h_offset = 0diff --git a/comfy/lora.py b/comfy/lora.pyindex 00358884b..4a44f1318 100644--- a/comfy/lora.py+++ b/comfy/lora.py@@ -260,6 +260,10 @@ def model_lora_keys_unet(model, key_map={}):key_map["transformer.{}".format(k[:-len(".weight")])] = to #simpletrainer and probably regular diffusers flux lora formatkey_map["lycoris_{}".format(k[:-len(".weight")].replace(".", "_"))] = to #simpletrainer lycoriskey_map["lora_transformer_{}".format(k[:-len(".weight")].replace(".", "_"))] = to #onetrainer+ for k in sdk:+ hidden_size = model.model_config.unet_config.get("hidden_size", 0)+ if k.endswith(".weight") and ".linear1." in k:+ key_map["{}".format(k.replace(".linear1.weight", ".linear1_qkv"))] = (k, (0, 0, hidden_size * 3))if isinstance(model, comfy.model_base.GenmoMochi):for k in sdk:diff --git a/comfy/lora_convert.py b/comfy/lora_convert.pyindex 3e00b63db..9d8d21efe 100644--- a/comfy/lora_convert.py+++ b/comfy/lora_convert.py@@ -15,10 +15,29 @@ def convert_lora_bfl_control(sd): #BFL loras for Fluxdef convert_lora_wan_fun(sd): #Wan Fun lorasreturn comfy.utils.state_dict_prefix_replace(sd, {"lora_unet__": "lora_unet_"})+def convert_uso_lora(sd):+ sd_out = {}+ for k in sd:+ tensor = sd[k]+ k_to = "diffusion_model.{}".format(k.replace(".down.weight", ".lora_down.weight")+ .replace(".up.weight", ".lora_up.weight")+ .replace(".qkv_lora2.", ".txt_attn.qkv.")+ .replace(".qkv_lora1.", ".img_attn.qkv.")+ .replace(".proj_lora1.", ".img_attn.proj.")+ .replace(".proj_lora2.", ".txt_attn.proj.")+ .replace(".qkv_lora.", ".linear1_qkv.")+ .replace(".proj_lora.", ".linear2.")+ .replace(".processor.", ".")+ )+ sd_out[k_to] = tensor+ return sd_out+def convert_lora(sd):if "img_in.lora_A.weight" in sd and "single_blocks.0.norm.key_norm.scale" in sd:return convert_lora_bfl_control(sd)if "lora_unet__blocks_0_cross_attn_k.lora_down.weight" in sd:return convert_lora_wan_fun(sd)+ if "single_blocks.37.processor.qkv_lora.up.weight" in sd and "double_blocks.18.processor.qkv_lora2.up.weight" in sd:+ return convert_uso_lora(sd)return sddiff --git a/comfy_extras/nodes_flux.py b/comfy_extras/nodes_flux.pyindex c8db75bb3..1bf7ddd92 100644--- a/comfy_extras/nodes_flux.py+++ b/comfy_extras/nodes_flux.py@@ -105,7 +105,7 @@ class FluxKontextMultiReferenceLatentMethod:def INPUT_TYPES(s):return {"required": {"conditioning": ("CONDITIONING", ),- "reference_latents_method": (("offset", "index"), ),+ "reference_latents_method": (("offset", "index", "uso"), ),}}RETURN_TYPES = ("CONDITIONING",)