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 loading flux 2 klein checkpoints saved with SaveCheckpoint. (#12033)
comfyanonymous09a2e67151c6
diff --git a/comfy/supported_models.py b/comfy/supported_models.pyindex 70abebf46..45d913fa6 100644--- a/comfy/supported_models.py+++ b/comfy/supported_models.py@@ -771,10 +771,24 @@ class Flux2(Flux):return outdef clip_target(self, state_dict={}):- return None # TODOpref = self.text_encoder_key_prefix[0]- t5_detect = comfy.text_encoders.sd3_clip.t5_xxl_detect(state_dict, "{}t5xxl.transformer.".format(pref))- return supported_models_base.ClipTarget(comfy.text_encoders.flux.FluxTokenizer, comfy.text_encoders.flux.flux_clip(**t5_detect))+ detect = comfy.text_encoders.hunyuan_video.llama_detect(state_dict, "{}qwen3_4b.transformer.".format(pref))+ if len(detect) > 0:+ detect["model_type"] = "qwen3_4b"+ return supported_models_base.ClipTarget(comfy.text_encoders.flux.KleinTokenizer, comfy.text_encoders.flux.klein_te(**detect))++ detect = comfy.text_encoders.hunyuan_video.llama_detect(state_dict, "{}qwen3_8b.transformer.".format(pref))+ if len(detect) > 0:+ detect["model_type"] = "qwen3_8b"+ return supported_models_base.ClipTarget(comfy.text_encoders.flux.KleinTokenizer8B, comfy.text_encoders.flux.klein_te(**detect))++ detect = comfy.text_encoders.hunyuan_video.llama_detect(state_dict, "{}mistral3_24b.transformer.".format(pref))+ if len(detect) > 0:+ if "{}mistral3_24b.transformer.model.layers.39.post_attention_layernorm.weight".format(pref) not in state_dict:+ detect["pruned"] = True+ return supported_models_base.ClipTarget(comfy.text_encoders.flux.Flux2Tokenizer, comfy.text_encoders.flux.flux2_te(**detect))++ return Noneclass GenmoMochi(supported_models_base.BASE):unet_config = {diff --git a/comfy/text_encoders/hunyuan_video.py b/comfy/text_encoders/hunyuan_video.pyindex a9a6c525e..2ddb4da60 100644--- a/comfy/text_encoders/hunyuan_video.py+++ b/comfy/text_encoders/hunyuan_video.py@@ -10,9 +10,11 @@ import comfy.utilsdef llama_detect(state_dict, prefix=""):out = {}- t5_key = "{}model.norm.weight".format(prefix)- if t5_key in state_dict:- out["dtype_llama"] = state_dict[t5_key].dtype+ norm_keys = ["{}model.norm.weight".format(prefix), "{}model.layers.0.input_layernorm.weight".format(prefix)]+ for norm_key in norm_keys:+ if norm_key in state_dict:+ out["dtype_llama"] = state_dict[norm_key].dtype+ breakquant = comfy.utils.detect_layer_quantization(state_dict, prefix)if quant is not None: