Rename essentials_category to main_category

The field is used for top-level tabs in the node library, not just
the Essentials tab. Rename to main_category to support Partner Nodes,
Basic, and other main categories.

Amp-Thread-ID: https://ampcode.com/threads/T-019c2b69-81c1-71c3-8096-450a39e20910

comfy/model_management.py

@@ -1724,11 +1724,9 @@ def soft_empty_cache(force=False):
     elif is_mlu():
         torch.mlu.empty_cache()
     elif torch.cuda.is_available():
-        if comfy.memory_management.aimdo_allocator is None:
-            #Pytorch 2.7 and earlier crashes if you try and empty_cache when mempools exist
-            torch.cuda.synchronize()
-            torch.cuda.empty_cache()
-            torch.cuda.ipc_collect()
+        torch.cuda.synchronize()
+        torch.cuda.empty_cache()
+        torch.cuda.ipc_collect()
 
 def unload_all_models():
     free_memory(1e30, get_torch_device())

comfy/model_patcher.py

@@ -1400,7 +1400,7 @@ class ModelPatcher:
                 continue
             key = "diffusion_model." + k
             unet_state_dict[k] = LazyCastingParam(self, key, comfy.utils.get_attr(self.model, key))
-        return self.model.state_dict_for_saving(unet_state_dict)
+        return self.model.state_dict_for_saving(unet_state_dict, clip_state_dict=clip_state_dict, vae_state_dict=vae_state_dict, clip_vision_state_dict=clip_vision_state_dict)
 
     def __del__(self):
         self.unpin_all_weights()

comfy/sd.py

@@ -1444,7 +1444,12 @@ def load_text_encoder_state_dicts(state_dicts=[], embedding_directory=None, clip
             tokenizer_data["gemma_spiece_model"] = clip_data_gemma.get("spiece_model", None)
             tokenizer_data["jina_spiece_model"] = clip_data_jina.get("spiece_model", None)
         elif clip_type == CLIPType.ACE:
-            clip_target.clip = comfy.text_encoders.ace15.te(**llama_detect(clip_data))
+            te_models = [detect_te_model(clip_data[0]), detect_te_model(clip_data[1])]
+            if TEModel.QWEN3_4B in te_models:
+                model_type = "qwen3_4b"
+            else:
+                model_type = "qwen3_2b"
+            clip_target.clip = comfy.text_encoders.ace15.te(lm_model=model_type, **llama_detect(clip_data))
             clip_target.tokenizer = comfy.text_encoders.ace15.ACE15Tokenizer
         else:
             clip_target.clip = sdxl_clip.SDXLClipModel

comfy/supported_models.py

@@ -1625,8 +1625,16 @@ class ACEStep15(supported_models_base.BASE):
 
     def clip_target(self, state_dict={}):
         pref = self.text_encoder_key_prefix[0]
-        hunyuan_detect = comfy.text_encoders.hunyuan_video.llama_detect(state_dict, "{}qwen3_2b.transformer.".format(pref))
-        return supported_models_base.ClipTarget(comfy.text_encoders.ace15.ACE15Tokenizer, comfy.text_encoders.ace15.te(**hunyuan_detect))
+        detect_2b = comfy.text_encoders.hunyuan_video.llama_detect(state_dict, "{}qwen3_2b.transformer.".format(pref))
+        detect_4b = comfy.text_encoders.hunyuan_video.llama_detect(state_dict, "{}qwen3_4b.transformer.".format(pref))
+        if "dtype_llama" in detect_2b:
+            detect = detect_2b
+            detect["lm_model"] = "qwen3_2b"
+        elif "dtype_llama" in detect_4b:
+            detect = detect_4b
+            detect["lm_model"] = "qwen3_4b"
+
+        return supported_models_base.ClipTarget(comfy.text_encoders.ace15.ACE15Tokenizer, comfy.text_encoders.ace15.te(**detect))
 
 
 models = [LotusD, Stable_Zero123, SD15_instructpix2pix, SD15, SD20, SD21UnclipL, SD21UnclipH, SDXL_instructpix2pix, SDXLRefiner, SDXL, SSD1B, KOALA_700M, KOALA_1B, Segmind_Vega, SD_X4Upscaler, Stable_Cascade_C, Stable_Cascade_B, SV3D_u, SV3D_p, SD3, StableAudio, AuraFlow, PixArtAlpha, PixArtSigma, HunyuanDiT, HunyuanDiT1, FluxInpaint, Flux, FluxSchnell, GenmoMochi, LTXV, LTXAV, HunyuanVideo15_SR_Distilled, HunyuanVideo15, HunyuanImage21Refiner, HunyuanImage21, HunyuanVideoSkyreelsI2V, HunyuanVideoI2V, HunyuanVideo, CosmosT2V, CosmosI2V, CosmosT2IPredict2, CosmosI2VPredict2, ZImage, Lumina2, WAN22_T2V, WAN21_T2V, WAN21_I2V, WAN21_FunControl2V, WAN21_Vace, WAN21_Camera, WAN22_Camera, WAN22_S2V, WAN21_HuMo, WAN22_Animate, Hunyuan3Dv2mini, Hunyuan3Dv2, Hunyuan3Dv2_1, HiDream, Chroma, ChromaRadiance, ACEStep, ACEStep15, Omnigen2, QwenImage, Flux2, Kandinsky5Image, Kandinsky5, Anima]

comfy/text_encoders/ace15.py

@@ -19,6 +19,7 @@ def sample_manual_loop_no_classes(
     min_tokens: int = 1,
     max_new_tokens: int = 2048,
     audio_start_id: int = 151669,  # The cutoff ID for audio codes
+    audio_end_id: int = 215669,
     eos_token_id: int = 151645,
 ):
     device = model.execution_device
@@ -60,6 +61,7 @@ def sample_manual_loop_no_classes(
         remove_logit_value = torch.finfo(cfg_logits.dtype).min
         # Only generate audio tokens
         cfg_logits[:, :audio_start_id] = remove_logit_value
+        cfg_logits[:, audio_end_id:] = remove_logit_value
 
         if eos_token_id is not None and eos_token_id < audio_start_id and min_tokens < step:
             cfg_logits[:, eos_token_id] = eos_score
@@ -162,14 +164,34 @@ class Qwen3_2B_ACE15(sd1_clip.SDClipModel):
 
         super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config={}, dtype=dtype, special_tokens={"pad": 151643}, layer_norm_hidden_state=False, model_class=comfy.text_encoders.llama.Qwen3_2B_ACE15_lm, enable_attention_masks=attention_mask, return_attention_masks=attention_mask, model_options=model_options)
 
+class Qwen3_4B_ACE15(sd1_clip.SDClipModel):
+    def __init__(self, device="cpu", layer="last", layer_idx=None, dtype=None, attention_mask=True, model_options={}):
+        llama_quantization_metadata = model_options.get("llama_quantization_metadata", None)
+        if llama_quantization_metadata is not None:
+            model_options = model_options.copy()
+            model_options["quantization_metadata"] = llama_quantization_metadata
+
+        super().__init__(device=device, layer=layer, layer_idx=layer_idx, textmodel_json_config={}, dtype=dtype, special_tokens={"pad": 151643}, layer_norm_hidden_state=False, model_class=comfy.text_encoders.llama.Qwen3_4B_ACE15_lm, enable_attention_masks=attention_mask, return_attention_masks=attention_mask, model_options=model_options)
+
 class ACE15TEModel(torch.nn.Module):
-    def __init__(self, device="cpu", dtype=None, dtype_llama=None, model_options={}):
+    def __init__(self, device="cpu", dtype=None, dtype_llama=None, lm_model=None, model_options={}):
         super().__init__()
         if dtype_llama is None:
             dtype_llama = dtype
 
+        model = None
+        self.constant = 0.4375
+        if lm_model == "qwen3_4b":
+            model = Qwen3_4B_ACE15
+            self.constant = 0.5625
+        elif lm_model == "qwen3_2b":
+            model = Qwen3_2B_ACE15
+
+        self.lm_model = lm_model
         self.qwen3_06b = Qwen3_06BModel(device=device, dtype=dtype, model_options=model_options)
-        self.qwen3_2b = Qwen3_2B_ACE15(device=device, dtype=dtype_llama, model_options=model_options)
+        if model is not None:
+            setattr(self, self.lm_model, model(device=device, dtype=dtype_llama, model_options=model_options))
+
         self.dtypes = set([dtype, dtype_llama])
 
     def encode_token_weights(self, token_weight_pairs):
@@ -182,17 +204,21 @@ class ACE15TEModel(torch.nn.Module):
         lyrics_embeds, _, extra_l = self.qwen3_06b.encode_token_weights(token_weight_pairs_lyrics)
 
         lm_metadata = token_weight_pairs["lm_metadata"]
-        audio_codes = generate_audio_codes(self.qwen3_2b, token_weight_pairs["lm_prompt"], token_weight_pairs["lm_prompt_negative"], min_tokens=lm_metadata["min_tokens"], max_tokens=lm_metadata["min_tokens"], seed=lm_metadata["seed"])
+        audio_codes = generate_audio_codes(getattr(self, self.lm_model, self.qwen3_06b), token_weight_pairs["lm_prompt"], token_weight_pairs["lm_prompt_negative"], min_tokens=lm_metadata["min_tokens"], max_tokens=lm_metadata["min_tokens"], seed=lm_metadata["seed"])
 
         return base_out, None, {"conditioning_lyrics": lyrics_embeds[:, 0], "audio_codes": [audio_codes]}
 
     def set_clip_options(self, options):
         self.qwen3_06b.set_clip_options(options)
-        self.qwen3_2b.set_clip_options(options)
+        lm_model = getattr(self, self.lm_model, None)
+        if lm_model is not None:
+            lm_model.set_clip_options(options)
 
     def reset_clip_options(self):
         self.qwen3_06b.reset_clip_options()
-        self.qwen3_2b.reset_clip_options()
+        lm_model = getattr(self, self.lm_model, None)
+        if lm_model is not None:
+            lm_model.reset_clip_options()
 
     def load_sd(self, sd):
         if "model.layers.0.post_attention_layernorm.weight" in sd:
@@ -200,11 +226,11 @@ class ACE15TEModel(torch.nn.Module):
             if shape[0] == 1024:
                 return self.qwen3_06b.load_sd(sd)
             else:
-                return self.qwen3_2b.load_sd(sd)
+                return getattr(self, self.lm_model).load_sd(sd)
 
     def memory_estimation_function(self, token_weight_pairs, device=None):
         lm_metadata = token_weight_pairs["lm_metadata"]
-        constant = 0.4375
+        constant = self.constant
         if comfy.model_management.should_use_bf16(device):
             constant *= 0.5
 
@@ -213,11 +239,11 @@ class ACE15TEModel(torch.nn.Module):
         num_tokens += lm_metadata['min_tokens']
         return num_tokens * constant * 1024 * 1024
 
-def te(dtype_llama=None, llama_quantization_metadata=None):
+def te(dtype_llama=None, llama_quantization_metadata=None, lm_model="qwen3_2b"):
     class ACE15TEModel_(ACE15TEModel):
         def __init__(self, device="cpu", dtype=None, model_options={}):
             if llama_quantization_metadata is not None:
                 model_options = model_options.copy()
                 model_options["llama_quantization_metadata"] = llama_quantization_metadata
-            super().__init__(device=device, dtype_llama=dtype_llama, dtype=dtype, model_options=model_options)
+            super().__init__(device=device, dtype_llama=dtype_llama, lm_model=lm_model, dtype=dtype, model_options=model_options)
     return ACE15TEModel_

comfy/text_encoders/llama.py

@@ -150,6 +150,29 @@ class Qwen3_2B_ACE15_lm_Config:
     final_norm: bool = True
     lm_head: bool = False
 
+@dataclass
+class Qwen3_4B_ACE15_lm_Config:
+    vocab_size: int = 217204
+    hidden_size: int = 2560
+    intermediate_size: int = 9728
+    num_hidden_layers: int = 36
+    num_attention_heads: int = 32
+    num_key_value_heads: int = 8
+    max_position_embeddings: int = 40960
+    rms_norm_eps: float = 1e-6
+    rope_theta: float = 1000000.0
+    transformer_type: str = "llama"
+    head_dim = 128
+    rms_norm_add = False
+    mlp_activation = "silu"
+    qkv_bias = False
+    rope_dims = None
+    q_norm = "gemma3"
+    k_norm = "gemma3"
+    rope_scale = None
+    final_norm: bool = True
+    lm_head: bool = False
+
 @dataclass
 class Qwen3_4BConfig:
     vocab_size: int = 151936
@@ -628,10 +651,10 @@ class Llama2_(nn.Module):
         mask = None
         if attention_mask is not None:
             mask = 1.0 - attention_mask.to(x.dtype).reshape((attention_mask.shape[0], 1, -1, attention_mask.shape[-1])).expand(attention_mask.shape[0], 1, seq_len, attention_mask.shape[-1])
-            mask = mask.masked_fill(mask.to(torch.bool), torch.finfo(x.dtype).min)
+            mask = mask.masked_fill(mask.to(torch.bool), torch.finfo(x.dtype).min / 4)
 
         if seq_len > 1:
-            causal_mask = torch.empty(past_len + seq_len, past_len + seq_len, dtype=x.dtype, device=x.device).fill_(torch.finfo(x.dtype).min).triu_(1)
+            causal_mask = torch.empty(past_len + seq_len, past_len + seq_len, dtype=x.dtype, device=x.device).fill_(torch.finfo(x.dtype).min / 4).triu_(1)
             if mask is not None:
                 mask += causal_mask
             else:
@@ -739,6 +762,21 @@ class BaseLlama:
     def forward(self, input_ids, *args, **kwargs):
         return self.model(input_ids, *args, **kwargs)
 
+class BaseQwen3:
+    def logits(self, x):
+        input = x[:, -1:]
+        module = self.model.embed_tokens
+
+        offload_stream = None
+        if module.comfy_cast_weights:
+            weight, _, offload_stream = comfy.ops.cast_bias_weight(module, input, offloadable=True)
+        else:
+            weight = self.model.embed_tokens.weight.to(x)
+
+        x = torch.nn.functional.linear(input, weight, None)
+
+        comfy.ops.uncast_bias_weight(module, weight, None, offload_stream)
+        return x
 
 class Llama2(BaseLlama, torch.nn.Module):
     def __init__(self, config_dict, dtype, device, operations):
@@ -767,7 +805,7 @@ class Qwen25_3B(BaseLlama, torch.nn.Module):
         self.model = Llama2_(config, device=device, dtype=dtype, ops=operations)
         self.dtype = dtype
 
-class Qwen3_06B(BaseLlama, torch.nn.Module):
+class Qwen3_06B(BaseLlama, BaseQwen3, torch.nn.Module):
     def __init__(self, config_dict, dtype, device, operations):
         super().__init__()
         config = Qwen3_06BConfig(**config_dict)
@@ -776,7 +814,7 @@ class Qwen3_06B(BaseLlama, torch.nn.Module):
         self.model = Llama2_(config, device=device, dtype=dtype, ops=operations)
         self.dtype = dtype
 
-class Qwen3_06B_ACE15(BaseLlama, torch.nn.Module):
+class Qwen3_06B_ACE15(BaseLlama, BaseQwen3, torch.nn.Module):
     def __init__(self, config_dict, dtype, device, operations):
         super().__init__()
         config = Qwen3_06B_ACE15_Config(**config_dict)
@@ -785,7 +823,7 @@ class Qwen3_06B_ACE15(BaseLlama, torch.nn.Module):
         self.model = Llama2_(config, device=device, dtype=dtype, ops=operations)
         self.dtype = dtype
 
-class Qwen3_2B_ACE15_lm(BaseLlama, torch.nn.Module):
+class Qwen3_2B_ACE15_lm(BaseLlama, BaseQwen3, torch.nn.Module):
     def __init__(self, config_dict, dtype, device, operations):
         super().__init__()
         config = Qwen3_2B_ACE15_lm_Config(**config_dict)
@@ -794,22 +832,7 @@ class Qwen3_2B_ACE15_lm(BaseLlama, torch.nn.Module):
         self.model = Llama2_(config, device=device, dtype=dtype, ops=operations)
         self.dtype = dtype
 
-    def logits(self, x):
-        input = x[:, -1:]
-        module = self.model.embed_tokens
-
-        offload_stream = None
-        if module.comfy_cast_weights:
-            weight, _, offload_stream = comfy.ops.cast_bias_weight(module, input, offloadable=True)
-        else:
-            weight = self.model.embed_tokens.weight.to(x)
-
-        x = torch.nn.functional.linear(input, weight, None)
-
-        comfy.ops.uncast_bias_weight(module, weight, None, offload_stream)
-        return x
-
-class Qwen3_4B(BaseLlama, torch.nn.Module):
+class Qwen3_4B(BaseLlama, BaseQwen3, torch.nn.Module):
     def __init__(self, config_dict, dtype, device, operations):
         super().__init__()
         config = Qwen3_4BConfig(**config_dict)
@@ -818,7 +841,16 @@ class Qwen3_4B(BaseLlama, torch.nn.Module):
         self.model = Llama2_(config, device=device, dtype=dtype, ops=operations)
         self.dtype = dtype
 
-class Qwen3_8B(BaseLlama, torch.nn.Module):
+class Qwen3_4B_ACE15_lm(BaseLlama, BaseQwen3, torch.nn.Module):
+    def __init__(self, config_dict, dtype, device, operations):
+        super().__init__()
+        config = Qwen3_4B_ACE15_lm_Config(**config_dict)
+        self.num_layers = config.num_hidden_layers
+
+        self.model = Llama2_(config, device=device, dtype=dtype, ops=operations)
+        self.dtype = dtype
+
+class Qwen3_8B(BaseLlama, BaseQwen3, torch.nn.Module):
     def __init__(self, config_dict, dtype, device, operations):
         super().__init__()
         config = Qwen3_8BConfig(**config_dict)

comfy/utils.py

@@ -82,14 +82,12 @@ _TYPES = {
 def load_safetensors(ckpt):
     f = open(ckpt, "rb")
     mapping = mmap.mmap(f.fileno(), 0, access=mmap.ACCESS_READ)
+    mv = memoryview(mapping)
 
     header_size = struct.unpack("<Q", mapping[:8])[0]
     header = json.loads(mapping[8:8+header_size].decode("utf-8"))
 
-    with warnings.catch_warnings():
-        #We are working with read-only RAM by design
-        warnings.filterwarnings("ignore", message="The given buffer is not writable")
-        data_area = torch.frombuffer(mapping, dtype=torch.uint8)[8 + header_size:]
+    mv = mv[8 + header_size:]
 
     sd = {}
     for name, info in header.items():
@@ -97,7 +95,13 @@ def load_safetensors(ckpt):
             continue
 
         start, end = info["data_offsets"]
-        sd[name] = data_area[start:end].view(_TYPES[info["dtype"]]).view(info["shape"])
+        if start == end:
+            sd[name] = torch.empty(info["shape"], dtype =_TYPES[info["dtype"]])
+        else:
+            with warnings.catch_warnings():
+                #We are working with read-only RAM by design
+                warnings.filterwarnings("ignore", message="The given buffer is not writable")
+                sd[name] = torch.frombuffer(mv[start:end], dtype=_TYPES[info["dtype"]]).view(info["shape"])
 
     return sd, header.get("__metadata__", {}),
 

comfy_api/latest/_io.py

@@ -1309,6 +1309,7 @@ class NodeInfoV1:
     api_node: bool=None
     price_badge: dict | None = None
     search_aliases: list[str]=None
+    main_category: str=None
 
 
 @dataclass
@@ -1430,6 +1431,8 @@ class Schema:
     """Flags a node as expandable, allowing NodeOutput to include 'expand' property."""
     accept_all_inputs: bool=False
     """When True, all inputs from the prompt will be passed to the node as kwargs, even if not defined in the schema."""
+    main_category: str | None = None
+    """Optional main category for top-level tabs in the node library (e.g., 'Basic', 'Image Tools', 'Partner Nodes')."""
 
     def validate(self):
         '''Validate the schema:
@@ -1536,6 +1539,7 @@ class Schema:
             python_module=getattr(cls, "RELATIVE_PYTHON_MODULE", "nodes"),
             price_badge=self.price_badge.as_dict(self.inputs) if self.price_badge is not None else None,
             search_aliases=self.search_aliases if self.search_aliases else None,
+            main_category=self.main_category,
         )
         return info
 

comfy_api_nodes/nodes_hunyuan3d.py

@@ -37,6 +37,7 @@ class TencentTextToModelNode(IO.ComfyNode):
             node_id="TencentTextToModelNode",
             display_name="Hunyuan3D: Text to Model (Pro)",
             category="api node/3d/Tencent",
+            main_category="3D",
             inputs=[
                 IO.Combo.Input(
                     "model",
@@ -147,6 +148,7 @@ class TencentImageToModelNode(IO.ComfyNode):
             node_id="TencentImageToModelNode",
             display_name="Hunyuan3D: Image(s) to Model (Pro)",
             category="api node/3d/Tencent",
+            main_category="3D",
             inputs=[
                 IO.Combo.Input(
                     "model",

comfy_api_nodes/nodes_kling.py

@@ -1936,6 +1936,7 @@ class KlingLipSyncAudioToVideoNode(IO.ComfyNode):
             node_id="KlingLipSyncAudioToVideoNode",
             display_name="Kling Lip Sync Video with Audio",
             category="api node/video/Kling",
+            main_category="Video Generation",
             description="Kling Lip Sync Audio to Video Node. Syncs mouth movements in a video file to the audio content of an audio file. When using, ensure that the audio contains clearly distinguishable vocals and that the video contains a distinct face. The audio file should not be larger than 5MB. The video file should not be larger than 100MB, should have height/width between 720px and 1920px, and should be between 2s and 10s in length.",
             inputs=[
                 IO.Video.Input("video"),

comfy_api_nodes/nodes_openai.py

@@ -576,6 +576,7 @@ class OpenAIChatNode(IO.ComfyNode):
             node_id="OpenAIChatNode",
             display_name="OpenAI ChatGPT",
             category="api node/text/OpenAI",
+            main_category="Text Generation",
             description="Generate text responses from an OpenAI model.",
             inputs=[
                 IO.String.Input(

comfy_api_nodes/nodes_recraft.py

@@ -963,6 +963,7 @@ class RecraftRemoveBackgroundNode(IO.ComfyNode):
             node_id="RecraftRemoveBackgroundNode",
             display_name="Recraft Remove Background",
             category="api node/image/Recraft",
+            main_category="Image Tools",
             description="Remove background from image, and return processed image and mask.",
             inputs=[
                 IO.Image.Input("image"),

comfy_api_nodes/nodes_stability.py

@@ -624,6 +624,7 @@ class StabilityTextToAudio(IO.ComfyNode):
             node_id="StabilityTextToAudio",
             display_name="Stability AI Text To Audio",
             category="api node/audio/Stability AI",
+            main_category="Audio",
             description=cleandoc(cls.__doc__ or ""),
             inputs=[
                 IO.Combo.Input(

comfy_extras/nodes_audio.py

@@ -129,6 +129,7 @@ class SaveAudio(IO.ComfyNode):
             search_aliases=["export flac"],
             display_name="Save Audio (FLAC)",
             category="audio",
+            main_category="Audio",
             inputs=[
                 IO.Audio.Input("audio"),
                 IO.String.Input("filename_prefix", default="audio/ComfyUI"),
@@ -270,6 +271,7 @@ class LoadAudio(IO.ComfyNode):
             search_aliases=["import audio", "open audio", "audio file"],
             display_name="Load Audio",
             category="audio",
+            main_category="Audio",
             inputs=[
                 IO.Combo.Input("audio", upload=IO.UploadType.audio, options=sorted(files)),
             ],

comfy_extras/nodes_canny.py

@@ -12,6 +12,7 @@ class Canny(io.ComfyNode):
             node_id="Canny",
             search_aliases=["edge detection", "outline", "contour detection", "line art"],
             category="image/preprocessors",
+            main_category="Image Tools/Preprocessing",
             inputs=[
                 io.Image.Input("image"),
                 io.Float.Input("low_threshold", default=0.4, min=0.01, max=0.99, step=0.01),

comfy_extras/nodes_hunyuan3d.py

@@ -618,16 +618,24 @@ class SaveGLB(IO.ComfyNode):
     def define_schema(cls):
         return IO.Schema(
             node_id="SaveGLB",
+            display_name="Save 3D Model",
             search_aliases=["export 3d model", "save mesh"],
             category="3d",
+            main_category="Basic",
             is_output_node=True,
             inputs=[
                 IO.MultiType.Input(
                     IO.Mesh.Input("mesh"),
                     types=[
                         IO.File3DGLB,
+                        IO.File3DGLTF,
+                        IO.File3DOBJ,
+                        IO.File3DFBX,
+                        IO.File3DSTL,
+                        IO.File3DUSDZ,
+                        IO.File3DAny,
                     ],
-                    tooltip="Mesh or GLB file to save",
+                    tooltip="Mesh or 3D file to save",
                 ),
                 IO.String.Input("filename_prefix", default="mesh/ComfyUI"),
             ],
@@ -649,7 +657,8 @@ class SaveGLB(IO.ComfyNode):
 
         if isinstance(mesh, Types.File3D):
             # Handle File3D input - save BytesIO data to output folder
-            f = f"{filename}_{counter:05}_.glb"
+            ext = mesh.format or "glb"
+            f = f"{filename}_{counter:05}_.{ext}"
             mesh.save_to(os.path.join(full_output_folder, f))
             results.append({
                 "filename": f,

comfy_extras/nodes_images.py

@@ -25,6 +25,7 @@ class ImageCrop(IO.ComfyNode):
             search_aliases=["trim"],
             display_name="Image Crop",
             category="image/transform",
+            main_category="Image Tools",
             inputs=[
                 IO.Image.Input("image"),
                 IO.Int.Input("width", default=512, min=1, max=nodes.MAX_RESOLUTION, step=1),
@@ -537,6 +538,7 @@ class ImageRotate(IO.ComfyNode):
             node_id="ImageRotate",
             search_aliases=["turn", "flip orientation"],
             category="image/transform",
+            main_category="Image Tools",
             inputs=[
                 IO.Image.Input("image"),
                 IO.Combo.Input("rotation", options=["none", "90 degrees", "180 degrees", "270 degrees"]),

comfy_extras/nodes_load_3d.py

@@ -31,6 +31,7 @@ class Load3D(IO.ComfyNode):
             node_id="Load3D",
             display_name="Load 3D & Animation",
             category="3d",
+            main_category="Basic",
             is_experimental=True,
             inputs=[
                 IO.Combo.Input("model_file", options=sorted(files), upload=IO.UploadType.model),
@@ -45,6 +46,7 @@ class Load3D(IO.ComfyNode):
                 IO.Image.Output(display_name="normal"),
                 IO.Load3DCamera.Output(display_name="camera_info"),
                 IO.Video.Output(display_name="recording_video"),
+                IO.File3DAny.Output(display_name="model_3d"),
             ],
         )
 
@@ -66,7 +68,8 @@ class Load3D(IO.ComfyNode):
 
             video = InputImpl.VideoFromFile(recording_video_path)
 
-        return IO.NodeOutput(output_image, output_mask, model_file, normal_image, image['camera_info'], video)
+        file_3d = Types.File3D(folder_paths.get_annotated_filepath(model_file))
+        return IO.NodeOutput(output_image, output_mask, model_file, normal_image, image['camera_info'], video, file_3d)
 
     process = execute  # TODO: remove
 

comfy_extras/nodes_post_processing.py

@@ -77,6 +77,7 @@ class Blur(io.ComfyNode):
         return io.Schema(
             node_id="ImageBlur",
             category="image/postprocessing",
+            main_category="Image Tools",
             inputs=[
                 io.Image.Input("image"),
                 io.Int.Input("blur_radius", default=1, min=1, max=31, step=1),

comfy_extras/nodes_video.py

@@ -73,6 +73,7 @@ class SaveVideo(io.ComfyNode):
             search_aliases=["export video"],
             display_name="Save Video",
             category="image/video",
+            main_category="Basic",
             description="Saves the input images to your ComfyUI output directory.",
             inputs=[
                 io.Video.Input("video", tooltip="The video to save."),
@@ -146,6 +147,7 @@ class GetVideoComponents(io.ComfyNode):
             search_aliases=["extract frames", "split video", "video to images", "demux"],
             display_name="Get Video Components",
             category="image/video",
+            main_category="Video Tools",
             description="Extracts all components from a video: frames, audio, and framerate.",
             inputs=[
                 io.Video.Input("video", tooltip="The video to extract components from."),
@@ -174,6 +176,7 @@ class LoadVideo(io.ComfyNode):
             search_aliases=["import video", "open video", "video file"],
             display_name="Load Video",
             category="image/video",
+            main_category="Basic",
             inputs=[
                 io.Combo.Input("file", options=sorted(files), upload=io.UploadType.video),
             ],

comfyui_version.py

@@ -1,3 +1,3 @@
 # This file is automatically generated by the build process when version is
 # updated in pyproject.toml.
-__version__ = "0.12.1"
+__version__ = "0.12.2"

main.py

@@ -192,7 +192,10 @@ import comfy_aimdo.control
 import comfy_aimdo.torch
 
 if enables_dynamic_vram():
-    if comfy_aimdo.control.init_device(comfy.model_management.get_torch_device().index):
+    if comfy.model_management.torch_version_numeric < (2, 8):
+        logging.warning("Unsupported Pytorch detected. DynamicVRAM support requires Pytorch version 2.8 or later. Falling back to legacy ModelPatcher. VRAM estimates may be unreliable especially on Windows")
+        comfy.memory_management.aimdo_allocator = None
+    elif comfy_aimdo.control.init_device(comfy.model_management.get_torch_device().index):
         if args.verbose == 'DEBUG':
             comfy_aimdo.control.set_log_debug()
         elif args.verbose == 'CRITICAL':
@@ -208,7 +211,7 @@ if enables_dynamic_vram():
         comfy.memory_management.aimdo_allocator = comfy_aimdo.torch.get_torch_allocator()
         logging.info("DynamicVRAM support detected and enabled")
     else:
-        logging.info("No working comfy-aimdo install detected. DynamicVRAM support disabled. Falling back to legacy ModelPatcher. VRAM estimates may be unreliable especially on Windows")
+        logging.warning("No working comfy-aimdo install detected. DynamicVRAM support disabled. Falling back to legacy ModelPatcher. VRAM estimates may be unreliable especially on Windows")
         comfy.memory_management.aimdo_allocator = None
 
 

nodes.py

@@ -69,6 +69,7 @@ class CLIPTextEncode(ComfyNodeABC):
     FUNCTION = "encode"
 
     CATEGORY = "conditioning"
+    MAIN_CATEGORY = "Basic"
     DESCRIPTION = "Encodes a text prompt using a CLIP model into an embedding that can be used to guide the diffusion model towards generating specific images."
     SEARCH_ALIASES = ["text", "prompt", "text prompt", "positive prompt", "negative prompt", "encode text", "text encoder", "encode prompt"]
 
@@ -667,6 +668,8 @@ class CLIPSetLastLayer:
         return (clip,)
 
 class LoraLoader:
+    MAIN_CATEGORY = "Image Generation"
+
     def __init__(self):
         self.loaded_lora = None
 
@@ -1648,6 +1651,7 @@ class SaveImage:
     OUTPUT_NODE = True
 
     CATEGORY = "image"
+    MAIN_CATEGORY = "Basic"
     DESCRIPTION = "Saves the input images to your ComfyUI output directory."
     SEARCH_ALIASES = ["save", "save image", "export image", "output image", "write image", "download"]
 
@@ -1706,6 +1710,7 @@ class LoadImage:
                 }
 
     CATEGORY = "image"
+    MAIN_CATEGORY = "Basic"
     SEARCH_ALIASES = ["load image", "open image", "import image", "image input", "upload image", "read image", "image loader"]
 
     RETURN_TYPES = ("IMAGE", "MASK")
@@ -1863,6 +1868,7 @@ class ImageScale:
     FUNCTION = "upscale"
 
     CATEGORY = "image/upscaling"
+    MAIN_CATEGORY = "Image Tools"
     SEARCH_ALIASES = ["resize", "resize image", "scale image", "image resize", "zoom", "zoom in", "change size"]
 
     def upscale(self, image, upscale_method, width, height, crop):
@@ -1902,6 +1908,7 @@ class ImageScaleBy:
 
 class ImageInvert:
     SEARCH_ALIASES = ["reverse colors"]
+    MAIN_CATEGORY = "Image Tools"
 
     @classmethod
     def INPUT_TYPES(s):
@@ -1918,6 +1925,7 @@ class ImageInvert:
 
 class ImageBatch:
     SEARCH_ALIASES = ["combine images", "merge images", "stack images"]
+    MAIN_CATEGORY = "Image Tools"
 
     @classmethod
     def INPUT_TYPES(s):

pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "ComfyUI"
-version = "0.12.1"
+version = "0.12.2"
 readme = "README.md"
 license = { file = "LICENSE" }
 requires-python = ">=3.10"

server.py

@@ -687,6 +687,10 @@ class PromptServer():
                 info['api_node'] = obj_class.API_NODE
 
             info['search_aliases'] = getattr(obj_class, 'SEARCH_ALIASES', [])
+
+            if hasattr(obj_class, 'MAIN_CATEGORY'):
+                info['main_category'] = obj_class.MAIN_CATEGORY
+
             return info
 
         @routes.get("/object_info")