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 directly to vram with CLIPLoader node.
comfyanonymous9acfe4df41b3
diff --git a/comfy/sd.py b/comfy/sd.pyindex bbd9412d0..c88f5a30d 100644--- a/comfy/sd.py+++ b/comfy/sd.py@@ -82,6 +82,8 @@ class CLIP:self.tokenizer = tokenizer(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data)self.patcher = comfy.model_patcher.ModelPatcher(self.cond_stage_model, load_device=load_device, offload_device=offload_device)+ if params['device'] == load_device:+ model_management.load_model_gpu(self.patcher)self.layer_idx = Nonelogging.debug("CLIP model load device: {}, offload device: {}, current: {}".format(load_device, offload_device, params['device']))@@ -455,7 +457,11 @@ def load_clip(ckpt_paths, embedding_directory=None, clip_type=CLIPType.STABLE_DIclip_target.clip = comfy.text_encoders.sd3_clip.SD3ClipModelclip_target.tokenizer = comfy.text_encoders.sd3_clip.SD3Tokenizer- clip = CLIP(clip_target, embedding_directory=embedding_directory)+ parameters = 0+ for c in clip_data:+ parameters += comfy.utils.calculate_parameters(c)++ clip = CLIP(clip_target, embedding_directory=embedding_directory, parameters=parameters)for c in clip_data:m, u = clip.load_sd(c)if len(m) > 0: