Merge branch 'master' into v3-improvements

README.md

@@ -119,6 +119,9 @@ ComfyUI follows a weekly release cycle targeting Monday but this regularly chang
 
 1. **[ComfyUI Core](https://github.com/comfyanonymous/ComfyUI)**
    - Releases a new stable version (e.g., v0.7.0) roughly every week.
+   - Starting from v0.4.0 patch versions will be used for fixes backported onto the current stable release.
+   - Minor versions will be used for releases off the master branch.
+   - Patch versions may still be used for releases on the master branch in cases where a backport would not make sense.
    - Commits outside of the stable release tags may be very unstable and break many custom nodes.
    - Serves as the foundation for the desktop release
 

comfy/context_windows.py

@@ -143,7 +143,7 @@ class IndexListContextHandler(ContextHandlerABC):
         # if multiple conds, split based on primary region
         if self.split_conds_to_windows and len(cond_in) > 1:
             region = window.get_region_index(len(cond_in))
-            logging.info(f"Splitting conds to windows; using region {region} for window {window[0]}-{window[-1]} with center ratio {window.center_ratio:.3f}")
+            logging.info(f"Splitting conds to windows; using region {region} for window {window.index_list[0]}-{window.index_list[-1]} with center ratio {window.center_ratio:.3f}")
             cond_in = [cond_in[region]]
         # cond object is a list containing a dict - outer list is irrelevant, so just loop through it
         for actual_cond in cond_in:

comfy/ldm/lumina/model.py

@@ -625,7 +625,7 @@ class NextDiT(nn.Module):
             if pooled is not None:
                 pooled = self.clip_text_pooled_proj(pooled)
             else:
-                pooled = torch.zeros((1, self.clip_text_dim), device=x.device, dtype=x.dtype)
+                pooled = torch.zeros((x.shape[0], self.clip_text_dim), device=x.device, dtype=x.dtype)
 
             adaln_input = self.time_text_embed(torch.cat((t, pooled), dim=-1))
 

comfy/model_base.py

@@ -1110,7 +1110,7 @@ class Lumina2(BaseModel):
             if 'num_tokens' not in out:
                 out['num_tokens'] = comfy.conds.CONDConstant(cross_attn.shape[1])
 
-        clip_text_pooled = kwargs["pooled_output"]  # Newbie
+        clip_text_pooled = kwargs.get("pooled_output", None)  # NewBie
         if clip_text_pooled is not None:
             out['clip_text_pooled'] = comfy.conds.CONDRegular(clip_text_pooled)
 

comfy/model_detection.py

@@ -430,8 +430,9 @@ def detect_unet_config(state_dict, key_prefix, metadata=None):
             dit_config["rope_theta"] = 10000.0
             dit_config["ffn_dim_multiplier"] = 4.0
             ctd_weight = state_dict.get('{}clip_text_pooled_proj.0.weight'.format(key_prefix), None)
-            if ctd_weight is not None:
+            if ctd_weight is not None:  # NewBie
                 dit_config["clip_text_dim"] = ctd_weight.shape[0]
+                # NewBie also sets axes_lens = [1024, 512, 512] but it's not used in ComfyUI
         elif dit_config["dim"] == 3840:  # Z image
             dit_config["n_heads"] = 30
             dit_config["n_kv_heads"] = 30

comfy/samplers.py

@@ -984,9 +984,6 @@ class CFGGuider:
         self.inner_model, self.conds, self.loaded_models = comfy.sampler_helpers.prepare_sampling(self.model_patcher, noise.shape, self.conds, self.model_options)
         device = self.model_patcher.load_device
 
-        if denoise_mask is not None:
-            denoise_mask = comfy.sampler_helpers.prepare_mask(denoise_mask, noise.shape, device)
-
         noise = noise.to(device)
         latent_image = latent_image.to(device)
         sigmas = sigmas.to(device)
@@ -1013,6 +1010,24 @@ class CFGGuider:
         else:
             latent_shapes = [latent_image.shape]
 
+        if denoise_mask is not None:
+            if denoise_mask.is_nested:
+                denoise_masks = denoise_mask.unbind()
+                denoise_masks = denoise_masks[:len(latent_shapes)]
+            else:
+                denoise_masks = [denoise_mask]
+
+            for i in range(len(denoise_masks), len(latent_shapes)):
+                denoise_masks.append(torch.ones(latent_shapes[i]))
+
+            for i in range(len(denoise_masks)):
+                denoise_masks[i] = comfy.sampler_helpers.prepare_mask(denoise_masks[i], latent_shapes[i], self.model_patcher.load_device)
+
+            if len(denoise_masks) > 1:
+                denoise_mask, _ = comfy.utils.pack_latents(denoise_masks)
+            else:
+                denoise_mask = denoise_masks[0]
+
         self.conds = {}
         for k in self.original_conds:
             self.conds[k] = list(map(lambda a: a.copy(), self.original_conds[k]))

comfy/sd.py

@@ -55,6 +55,8 @@ import comfy.text_encoders.hunyuan_image
 import comfy.text_encoders.z_image
 import comfy.text_encoders.ovis
 import comfy.text_encoders.kandinsky5
+import comfy.text_encoders.jina_clip_2
+import comfy.text_encoders.newbie
 
 import comfy.model_patcher
 import comfy.lora
@@ -1008,6 +1010,7 @@ class CLIPType(Enum):
     OVIS = 21
     KANDINSKY5 = 22
     KANDINSKY5_IMAGE = 23
+    NEWBIE = 24
 
 
 def load_clip(ckpt_paths, embedding_directory=None, clip_type=CLIPType.STABLE_DIFFUSION, model_options={}):
@@ -1038,6 +1041,7 @@ class TEModel(Enum):
     MISTRAL3_24B_PRUNED_FLUX2 = 15
     QWEN3_4B = 16
     QWEN3_2B = 17
+    JINA_CLIP_2 = 18
 
 
 def detect_te_model(sd):
@@ -1047,6 +1051,8 @@ def detect_te_model(sd):
         return TEModel.CLIP_H
     if "text_model.encoder.layers.0.mlp.fc1.weight" in sd:
         return TEModel.CLIP_L
+    if "model.encoder.layers.0.mixer.Wqkv.weight" in sd:
+        return TEModel.JINA_CLIP_2
     if "encoder.block.23.layer.1.DenseReluDense.wi_1.weight" in sd:
         weight = sd["encoder.block.23.layer.1.DenseReluDense.wi_1.weight"]
         if weight.shape[-1] == 4096:
@@ -1207,6 +1213,9 @@ def load_text_encoder_state_dicts(state_dicts=[], embedding_directory=None, clip
         elif te_model == TEModel.QWEN3_2B:
             clip_target.clip = comfy.text_encoders.ovis.te(**llama_detect(clip_data))
             clip_target.tokenizer = comfy.text_encoders.ovis.OvisTokenizer
+        elif te_model == TEModel.JINA_CLIP_2:
+            clip_target.clip = comfy.text_encoders.jina_clip_2.JinaClip2TextModelWrapper
+            clip_target.tokenizer = comfy.text_encoders.jina_clip_2.JinaClip2TokenizerWrapper
         else:
             # clip_l
             if clip_type == CLIPType.SD3:
@@ -1262,6 +1271,17 @@ def load_text_encoder_state_dicts(state_dicts=[], embedding_directory=None, clip
         elif clip_type == CLIPType.KANDINSKY5_IMAGE:
             clip_target.clip = comfy.text_encoders.kandinsky5.te(**llama_detect(clip_data))
             clip_target.tokenizer = comfy.text_encoders.kandinsky5.Kandinsky5TokenizerImage
+        elif clip_type == CLIPType.NEWBIE:
+            clip_target.clip = comfy.text_encoders.newbie.te(**llama_detect(clip_data))
+            clip_target.tokenizer = comfy.text_encoders.newbie.NewBieTokenizer
+            if "model.layers.0.self_attn.q_norm.weight" in clip_data[0]:
+                clip_data_gemma = clip_data[0]
+                clip_data_jina = clip_data[1]
+            else:
+                clip_data_gemma = clip_data[1]
+                clip_data_jina = clip_data[0]
+            tokenizer_data["gemma_spiece_model"] = clip_data_gemma.get("spiece_model", None)
+            tokenizer_data["jina_spiece_model"] = clip_data_jina.get("spiece_model", None)
         else:
             clip_target.clip = sdxl_clip.SDXLClipModel
             clip_target.tokenizer = sdxl_clip.SDXLTokenizer

comfy/sd1_clip.py

@@ -466,7 +466,7 @@ def load_embed(embedding_name, embedding_directory, embedding_size, embed_key=No
     return embed_out
 
 class SDTokenizer:
-    def __init__(self, tokenizer_path=None, max_length=77, pad_with_end=True, embedding_directory=None, embedding_size=768, embedding_key='clip_l', tokenizer_class=CLIPTokenizer, has_start_token=True, has_end_token=True, pad_to_max_length=True, min_length=None, pad_token=None, end_token=None, min_padding=None, pad_left=False, tokenizer_data={}, tokenizer_args={}):
+    def __init__(self, tokenizer_path=None, max_length=77, pad_with_end=True, embedding_directory=None, embedding_size=768, embedding_key='clip_l', tokenizer_class=CLIPTokenizer, has_start_token=True, has_end_token=True, pad_to_max_length=True, min_length=None, pad_token=None, end_token=None, min_padding=None, pad_left=False, disable_weights=False, tokenizer_data={}, tokenizer_args={}):
         if tokenizer_path is None:
             tokenizer_path = os.path.join(os.path.dirname(os.path.realpath(__file__)), "sd1_tokenizer")
         self.tokenizer = tokenizer_class.from_pretrained(tokenizer_path, **tokenizer_args)
@@ -513,6 +513,8 @@ class SDTokenizer:
         self.embedding_size = embedding_size
         self.embedding_key = embedding_key
 
+        self.disable_weights = disable_weights
+
     def _try_get_embedding(self, embedding_name:str):
         '''
         Takes a potential embedding name and tries to retrieve it.
@@ -547,7 +549,7 @@ class SDTokenizer:
         min_padding = tokenizer_options.get("{}_min_padding".format(self.embedding_key), self.min_padding)
 
         text = escape_important(text)
-        if kwargs.get("disable_weights", False):
+        if kwargs.get("disable_weights", self.disable_weights):
             parsed_weights = [(text, 1.0)]
         else:
             parsed_weights = token_weights(text, 1.0)

comfy/text_encoders/jina_clip_2.py

@@ -0,0 +1,219 @@
+# Jina CLIP v2 and Jina Embeddings v3 both use their modified XLM-RoBERTa architecture. Reference implementation:
+# Jina CLIP v2 (both text and vision): https://huggingface.co/jinaai/jina-clip-implementation/blob/39e6a55ae971b59bea6e44675d237c99762e7ee2/modeling_clip.py
+# Jina XLM-RoBERTa (text only): http://huggingface.co/jinaai/xlm-roberta-flash-implementation/blob/2b6bc3f30750b3a9648fe9b63448c09920efe9be/modeling_xlm_roberta.py
+
+from dataclasses import dataclass
+
+import torch
+from torch import nn as nn
+from torch.nn import functional as F
+
+import comfy.model_management
+import comfy.ops
+from comfy import sd1_clip
+from .spiece_tokenizer import SPieceTokenizer
+
+class JinaClip2Tokenizer(sd1_clip.SDTokenizer):
+    def __init__(self, embedding_directory=None, tokenizer_data={}):
+        tokenizer = tokenizer_data.get("spiece_model", None)
+        # The official NewBie uses max_length=8000, but Jina Embeddings v3 actually supports 8192
+        super().__init__(tokenizer, pad_with_end=False, embedding_size=1024, embedding_key='jina_clip_2', tokenizer_class=SPieceTokenizer, has_start_token=True, has_end_token=True, pad_to_max_length=False, max_length=8192, min_length=1, pad_token=1, end_token=2, tokenizer_args={"add_bos": True, "add_eos": True}, tokenizer_data=tokenizer_data)
+
+    def state_dict(self):
+        return {"spiece_model": self.tokenizer.serialize_model()}
+
+class JinaClip2TokenizerWrapper(sd1_clip.SD1Tokenizer):
+    def __init__(self, embedding_directory=None, tokenizer_data={}):
+        super().__init__(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data, tokenizer=JinaClip2Tokenizer, name="jina_clip_2")
+
+# https://huggingface.co/jinaai/jina-embeddings-v3/blob/343dbf534c76fe845f304fa5c2d1fd87e1e78918/config.json
+@dataclass
+class XLMRobertaConfig:
+    vocab_size: int = 250002
+    type_vocab_size: int = 1
+    hidden_size: int = 1024
+    num_hidden_layers: int = 24
+    num_attention_heads: int = 16
+    rotary_emb_base: float = 20000.0
+    intermediate_size: int = 4096
+    hidden_act: str = "gelu"
+    hidden_dropout_prob: float = 0.1
+    attention_probs_dropout_prob: float = 0.1
+    layer_norm_eps: float = 1e-05
+    bos_token_id: int = 0
+    eos_token_id: int = 2
+    pad_token_id: int = 1
+
+class XLMRobertaEmbeddings(nn.Module):
+    def __init__(self, config, device=None, dtype=None, ops=None):
+        super().__init__()
+        embed_dim = config.hidden_size
+        self.word_embeddings = ops.Embedding(config.vocab_size, embed_dim, padding_idx=config.pad_token_id, device=device, dtype=dtype)
+        self.token_type_embeddings = ops.Embedding(config.type_vocab_size, embed_dim, device=device, dtype=dtype)
+
+    def forward(self, input_ids=None, embeddings=None):
+        if input_ids is not None and embeddings is None:
+            embeddings = self.word_embeddings(input_ids)
+
+        if embeddings is not None:
+            token_type_ids = torch.zeros(embeddings.shape[1], device=embeddings.device, dtype=torch.int32)
+            token_type_embeddings = self.token_type_embeddings(token_type_ids)
+            embeddings = embeddings + token_type_embeddings
+        return embeddings
+
+class RotaryEmbedding(nn.Module):
+    def __init__(self, dim, base, device=None):
+        super().__init__()
+        inv_freq = 1.0 / (base ** (torch.arange(0, dim, 2, device=device, dtype=torch.float32) / dim))
+        self.register_buffer("inv_freq", inv_freq, persistent=False)
+        self._seq_len_cached = 0
+        self._cos_cached = None
+        self._sin_cached = None
+
+    def _update_cos_sin_cache(self, seqlen, device=None, dtype=None):
+        if seqlen > self._seq_len_cached or self._cos_cached is None or self._cos_cached.device != device or self._cos_cached.dtype != dtype:
+            self._seq_len_cached = seqlen
+            t = torch.arange(seqlen, device=device, dtype=torch.float32)
+            freqs = torch.outer(t, self.inv_freq.to(device=t.device))
+            emb = torch.cat((freqs, freqs), dim=-1)
+            self._cos_cached = emb.cos().to(dtype)
+            self._sin_cached = emb.sin().to(dtype)
+
+    def forward(self, q, k):
+        batch, seqlen, heads, head_dim = q.shape
+        self._update_cos_sin_cache(seqlen, device=q.device, dtype=q.dtype)
+
+        cos = self._cos_cached[:seqlen].view(1, seqlen, 1, head_dim)
+        sin = self._sin_cached[:seqlen].view(1, seqlen, 1, head_dim)
+
+        def rotate_half(x):
+            size = x.shape[-1] // 2
+            x1, x2 = x[..., :size], x[..., size:]
+            return torch.cat((-x2, x1), dim=-1)
+
+        q_embed = (q * cos) + (rotate_half(q) * sin)
+        k_embed = (k * cos) + (rotate_half(k) * sin)
+        return q_embed, k_embed
+
+class MHA(nn.Module):
+    def __init__(self, config, device=None, dtype=None, ops=None):
+        super().__init__()
+        embed_dim = config.hidden_size
+        self.num_heads = config.num_attention_heads
+        self.head_dim = embed_dim // config.num_attention_heads
+
+        self.rotary_emb = RotaryEmbedding(self.head_dim, config.rotary_emb_base, device=device)
+        self.Wqkv = ops.Linear(embed_dim, 3 * embed_dim, device=device, dtype=dtype)
+        self.out_proj = ops.Linear(embed_dim, embed_dim, device=device, dtype=dtype)
+
+    def forward(self, x, mask=None, optimized_attention=None):
+        qkv = self.Wqkv(x)
+        batch_size, seq_len, _ = qkv.shape
+        qkv = qkv.view(batch_size, seq_len, 3, self.num_heads, self.head_dim)
+        q, k, v = qkv.unbind(2)
+
+        q, k = self.rotary_emb(q, k)
+
+        # NHD -> HND
+        q = q.transpose(1, 2)
+        k = k.transpose(1, 2)
+        v = v.transpose(1, 2)
+
+        out = optimized_attention(q, k, v, heads=self.num_heads, mask=mask, skip_reshape=True)
+        return self.out_proj(out)
+
+class MLP(nn.Module):
+    def __init__(self, config, device=None, dtype=None, ops=None):
+        super().__init__()
+        self.fc1 = ops.Linear(config.hidden_size, config.intermediate_size, device=device, dtype=dtype)
+        self.activation = F.gelu
+        self.fc2 = ops.Linear(config.intermediate_size, config.hidden_size, device=device, dtype=dtype)
+
+    def forward(self, x):
+        x = self.fc1(x)
+        x = self.activation(x)
+        x = self.fc2(x)
+        return x
+
+class Block(nn.Module):
+    def __init__(self, config, device=None, dtype=None, ops=None):
+        super().__init__()
+        self.mixer = MHA(config, device=device, dtype=dtype, ops=ops)
+        self.dropout1 = nn.Dropout(config.hidden_dropout_prob)
+        self.norm1 = ops.LayerNorm(config.hidden_size, eps=config.layer_norm_eps, device=device, dtype=dtype)
+        self.mlp = MLP(config, device=device, dtype=dtype, ops=ops)
+        self.dropout2 = nn.Dropout(config.hidden_dropout_prob)
+        self.norm2 = ops.LayerNorm(config.hidden_size, eps=config.layer_norm_eps, device=device, dtype=dtype)
+
+    def forward(self, hidden_states, mask=None, optimized_attention=None):
+        mixer_out = self.mixer(hidden_states, mask=mask, optimized_attention=optimized_attention)
+        hidden_states = self.norm1(self.dropout1(mixer_out) + hidden_states)
+        mlp_out = self.mlp(hidden_states)
+        hidden_states = self.norm2(self.dropout2(mlp_out) + hidden_states)
+        return hidden_states
+
+class XLMRobertaEncoder(nn.Module):
+    def __init__(self, config, device=None, dtype=None, ops=None):
+        super().__init__()
+        self.layers = nn.ModuleList([Block(config, device=device, dtype=dtype, ops=ops) for _ in range(config.num_hidden_layers)])
+
+    def forward(self, hidden_states, attention_mask=None):
+        optimized_attention = comfy.ldm.modules.attention.optimized_attention_for_device(hidden_states.device, mask=attention_mask is not None, small_input=True)
+        for layer in self.layers:
+            hidden_states = layer(hidden_states, mask=attention_mask, optimized_attention=optimized_attention)
+        return hidden_states
+
+class XLMRobertaModel_(nn.Module):
+    def __init__(self, config, device=None, dtype=None, ops=None):
+        super().__init__()
+        self.embeddings = XLMRobertaEmbeddings(config, device=device, dtype=dtype, ops=ops)
+        self.emb_ln = ops.LayerNorm(config.hidden_size, eps=config.layer_norm_eps, device=device, dtype=dtype)
+        self.emb_drop = nn.Dropout(config.hidden_dropout_prob)
+        self.encoder = XLMRobertaEncoder(config, device=device, dtype=dtype, ops=ops)
+
+    def forward(self, input_ids, attention_mask=None, embeds=None, num_tokens=None, intermediate_output=None, final_layer_norm_intermediate=True, dtype=None, embeds_info=[]):
+        x = self.embeddings(input_ids=input_ids, embeddings=embeds)
+        x = self.emb_ln(x)
+        x = self.emb_drop(x)
+
+        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]))
+            mask = mask.masked_fill(mask.to(torch.bool), -torch.finfo(x.dtype).max)
+
+        sequence_output = self.encoder(x, attention_mask=mask)
+
+        # Mean pool, see https://huggingface.co/jinaai/jina-clip-implementation/blob/39e6a55ae971b59bea6e44675d237c99762e7ee2/hf_model.py
+        pooled_output = None
+        if attention_mask is None:
+            pooled_output = sequence_output.mean(dim=1)
+        else:
+            attention_mask = attention_mask.to(sequence_output.dtype)
+            pooled_output = (sequence_output * attention_mask.unsqueeze(-1)).sum(dim=1) / attention_mask.sum(dim=-1, keepdim=True)
+
+        # Intermediate output is not yet implemented, use None for placeholder
+        return sequence_output, None, pooled_output
+
+class XLMRobertaModel(nn.Module):
+    def __init__(self, config_dict, dtype, device, operations):
+        super().__init__()
+        self.config = XLMRobertaConfig(**config_dict)
+        self.model = XLMRobertaModel_(self.config, device=device, dtype=dtype, ops=operations)
+        self.num_layers = self.config.num_hidden_layers
+
+    def get_input_embeddings(self):
+        return self.model.embeddings.word_embeddings
+
+    def set_input_embeddings(self, embeddings):
+        self.model.embeddings.word_embeddings = embeddings
+
+    def forward(self, *args, **kwargs):
+        return self.model(*args, **kwargs)
+
+class JinaClip2TextModel(sd1_clip.SDClipModel):
+    def __init__(self, device="cpu", dtype=None, model_options={}):
+        super().__init__(device=device, dtype=dtype, textmodel_json_config={}, model_class=XLMRobertaModel, special_tokens={"start": 0, "end": 2, "pad": 1}, enable_attention_masks=True, return_attention_masks=True, model_options=model_options)
+
+class JinaClip2TextModelWrapper(sd1_clip.SD1ClipModel):
+    def __init__(self, device="cpu", dtype=None, model_options={}):
+        super().__init__(device=device, dtype=dtype, clip_model=JinaClip2TextModel, name="jina_clip_2", model_options=model_options)

comfy/text_encoders/llama.py

@@ -3,7 +3,6 @@ import torch.nn as nn
 from dataclasses import dataclass
 from typing import Optional, Any
 import math
-import logging
 
 from comfy.ldm.modules.attention import optimized_attention_for_device
 import comfy.model_management
@@ -177,7 +176,7 @@ class Gemma3_4B_Config:
     num_key_value_heads: int = 4
     max_position_embeddings: int = 131072
     rms_norm_eps: float = 1e-6
-    rope_theta = [10000.0, 1000000.0]
+    rope_theta = [1000000.0, 10000.0]
     transformer_type: str = "gemma3"
     head_dim = 256
     rms_norm_add = True
@@ -186,8 +185,8 @@ class Gemma3_4B_Config:
     rope_dims = None
     q_norm = "gemma3"
     k_norm = "gemma3"
-    sliding_attention = [False, False, False, False, False, 1024]
-    rope_scale = [1.0, 8.0]
+    sliding_attention = [1024, 1024, 1024, 1024, 1024, False]
+    rope_scale = [8.0, 1.0]
     final_norm: bool = True
 
 class RMSNorm(nn.Module):
@@ -370,7 +369,7 @@ class TransformerBlockGemma2(nn.Module):
         self.pre_feedforward_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps, add=config.rms_norm_add, device=device, dtype=dtype)
         self.post_feedforward_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps, add=config.rms_norm_add, device=device, dtype=dtype)
 
-        if config.sliding_attention is not None:  # TODO: implement. (Not that necessary since models are trained on less than 1024 tokens)
+        if config.sliding_attention is not None:
             self.sliding_attention = config.sliding_attention[index % len(config.sliding_attention)]
         else:
             self.sliding_attention = False
@@ -387,7 +386,12 @@ class TransformerBlockGemma2(nn.Module):
         if self.transformer_type == 'gemma3':
             if self.sliding_attention:
                 if x.shape[1] > self.sliding_attention:
-                    logging.warning("Warning: sliding attention not implemented, results may be incorrect")
+                    sliding_mask = torch.full((x.shape[1], x.shape[1]), float("-inf"), device=x.device, dtype=x.dtype)
+                    sliding_mask.tril_(diagonal=-self.sliding_attention)
+                    if attention_mask is not None:
+                        attention_mask = attention_mask + sliding_mask
+                    else:
+                        attention_mask = sliding_mask
                 freqs_cis = freqs_cis[1]
             else:
                 freqs_cis = freqs_cis[0]

comfy/text_encoders/lumina2.py

@@ -14,7 +14,7 @@ class Gemma2BTokenizer(sd1_clip.SDTokenizer):
 class Gemma3_4BTokenizer(sd1_clip.SDTokenizer):
     def __init__(self, embedding_directory=None, tokenizer_data={}):
         tokenizer = tokenizer_data.get("spiece_model", None)
-        super().__init__(tokenizer, pad_with_end=False, embedding_size=2560, embedding_key='gemma3_4b', tokenizer_class=SPieceTokenizer, has_end_token=False, pad_to_max_length=False, max_length=99999999, min_length=1, tokenizer_args={"add_bos": True, "add_eos": False}, tokenizer_data=tokenizer_data)
+        super().__init__(tokenizer, pad_with_end=False, embedding_size=2560, embedding_key='gemma3_4b', tokenizer_class=SPieceTokenizer, has_end_token=False, pad_to_max_length=False, max_length=99999999, min_length=1, tokenizer_args={"add_bos": True, "add_eos": False}, disable_weights=True, tokenizer_data=tokenizer_data)
 
     def state_dict(self):
         return {"spiece_model": self.tokenizer.serialize_model()}
@@ -33,6 +33,11 @@ class Gemma2_2BModel(sd1_clip.SDClipModel):
 
 class Gemma3_4BModel(sd1_clip.SDClipModel):
     def __init__(self, device="cpu", layer="hidden", layer_idx=-2, 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={"start": 2, "pad": 0}, layer_norm_hidden_state=False, model_class=comfy.text_encoders.llama.Gemma3_4B, enable_attention_masks=attention_mask, return_attention_masks=attention_mask, model_options=model_options)
 
 class LuminaModel(sd1_clip.SD1ClipModel):

comfy/text_encoders/newbie.py

@@ -0,0 +1,62 @@
+import torch
+
+import comfy.model_management
+import comfy.text_encoders.jina_clip_2
+import comfy.text_encoders.lumina2
+
+class NewBieTokenizer:
+    def __init__(self, embedding_directory=None, tokenizer_data={}):
+        self.gemma = comfy.text_encoders.lumina2.Gemma3_4BTokenizer(embedding_directory=embedding_directory, tokenizer_data={"spiece_model": tokenizer_data["gemma_spiece_model"]})
+        self.jina = comfy.text_encoders.jina_clip_2.JinaClip2Tokenizer(embedding_directory=embedding_directory, tokenizer_data={"spiece_model": tokenizer_data["jina_spiece_model"]})
+
+    def tokenize_with_weights(self, text:str, return_word_ids=False, **kwargs):
+        out = {}
+        out["gemma"] = self.gemma.tokenize_with_weights(text, return_word_ids, **kwargs)
+        out["jina"] = self.jina.tokenize_with_weights(text, return_word_ids, **kwargs)
+        return out
+
+    def untokenize(self, token_weight_pair):
+        raise NotImplementedError
+
+    def state_dict(self):
+        return {}
+
+class NewBieTEModel(torch.nn.Module):
+    def __init__(self, dtype_gemma=None, device="cpu", dtype=None, model_options={}):
+        super().__init__()
+        dtype_gemma = comfy.model_management.pick_weight_dtype(dtype_gemma, dtype, device)
+        self.gemma = comfy.text_encoders.lumina2.Gemma3_4BModel(device=device, dtype=dtype_gemma, model_options=model_options)
+        self.jina = comfy.text_encoders.jina_clip_2.JinaClip2TextModel(device=device, dtype=dtype, model_options=model_options)
+        self.dtypes = {dtype, dtype_gemma}
+
+    def set_clip_options(self, options):
+        self.gemma.set_clip_options(options)
+        self.jina.set_clip_options(options)
+
+    def reset_clip_options(self):
+        self.gemma.reset_clip_options()
+        self.jina.reset_clip_options()
+
+    def encode_token_weights(self, token_weight_pairs):
+        token_weight_pairs_gemma = token_weight_pairs["gemma"]
+        token_weight_pairs_jina = token_weight_pairs["jina"]
+
+        gemma_out, gemma_pooled, gemma_extra = self.gemma.encode_token_weights(token_weight_pairs_gemma)
+        jina_out, jina_pooled, jina_extra = self.jina.encode_token_weights(token_weight_pairs_jina)
+
+        return gemma_out, jina_pooled, gemma_extra
+
+    def load_sd(self, sd):
+        if "model.layers.0.self_attn.q_norm.weight" in sd:
+            return self.gemma.load_sd(sd)
+        else:
+            return self.jina.load_sd(sd)
+
+def te(dtype_llama=None, llama_quantization_metadata=None):
+    class NewBieTEModel_(NewBieTEModel):
+        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__(dtype_gemma=dtype_llama, device=device, dtype=dtype, model_options=model_options)
+    return NewBieTEModel_

comfy_api_nodes/apis/bytedance_api.py

@@ -10,7 +10,7 @@ class Text2ImageTaskCreationRequest(BaseModel):
     size: str | None = Field(None)
     seed: int | None = Field(0, ge=0, le=2147483647)
     guidance_scale: float | None = Field(..., ge=1.0, le=10.0)
-    watermark: bool | None = Field(True)
+    watermark: bool | None = Field(False)
 
 
 class Image2ImageTaskCreationRequest(BaseModel):
@@ -21,7 +21,7 @@ class Image2ImageTaskCreationRequest(BaseModel):
     size: str | None = Field("adaptive")
     seed: int | None = Field(..., ge=0, le=2147483647)
     guidance_scale: float | None = Field(..., ge=1.0, le=10.0)
-    watermark: bool | None = Field(True)
+    watermark: bool | None = Field(False)
 
 
 class Seedream4Options(BaseModel):
@@ -37,7 +37,7 @@ class Seedream4TaskCreationRequest(BaseModel):
     seed: int = Field(..., ge=0, le=2147483647)
     sequential_image_generation: str = Field("disabled")
     sequential_image_generation_options: Seedream4Options = Field(Seedream4Options(max_images=15))
-    watermark: bool = Field(True)
+    watermark: bool = Field(False)
 
 
 class ImageTaskCreationResponse(BaseModel):

comfy_api_nodes/nodes_bytedance.py

@@ -112,7 +112,7 @@ class ByteDanceImageNode(IO.ComfyNode):
                 ),
                 IO.Boolean.Input(
                     "watermark",
-                    default=True,
+                    default=False,
                     tooltip='Whether to add an "AI generated" watermark to the image',
                     optional=True,
                 ),
@@ -215,7 +215,7 @@ class ByteDanceImageEditNode(IO.ComfyNode):
                 ),
                 IO.Boolean.Input(
                     "watermark",
-                    default=True,
+                    default=False,
                     tooltip='Whether to add an "AI generated" watermark to the image',
                     optional=True,
                 ),
@@ -346,7 +346,7 @@ class ByteDanceSeedreamNode(IO.ComfyNode):
                 ),
                 IO.Boolean.Input(
                     "watermark",
-                    default=True,
+                    default=False,
                     tooltip='Whether to add an "AI generated" watermark to the image.',
                     optional=True,
                 ),
@@ -380,7 +380,7 @@ class ByteDanceSeedreamNode(IO.ComfyNode):
         sequential_image_generation: str = "disabled",
         max_images: int = 1,
         seed: int = 0,
-        watermark: bool = True,
+        watermark: bool = False,
         fail_on_partial: bool = True,
     ) -> IO.NodeOutput:
         validate_string(prompt, strip_whitespace=True, min_length=1)
@@ -507,7 +507,7 @@ class ByteDanceTextToVideoNode(IO.ComfyNode):
                 ),
                 IO.Boolean.Input(
                     "watermark",
-                    default=True,
+                    default=False,
                     tooltip='Whether to add an "AI generated" watermark to the video.',
                     optional=True,
                 ),
@@ -617,7 +617,7 @@ class ByteDanceImageToVideoNode(IO.ComfyNode):
                 ),
                 IO.Boolean.Input(
                     "watermark",
-                    default=True,
+                    default=False,
                     tooltip='Whether to add an "AI generated" watermark to the video.',
                     optional=True,
                 ),
@@ -739,7 +739,7 @@ class ByteDanceFirstLastFrameNode(IO.ComfyNode):
                 ),
                 IO.Boolean.Input(
                     "watermark",
-                    default=True,
+                    default=False,
                     tooltip='Whether to add an "AI generated" watermark to the video.',
                     optional=True,
                 ),
@@ -862,7 +862,7 @@ class ByteDanceImageReferenceNode(IO.ComfyNode):
                 ),
                 IO.Boolean.Input(
                     "watermark",
-                    default=True,
+                    default=False,
                     tooltip='Whether to add an "AI generated" watermark to the video.',
                     optional=True,
                 ),

comfy_api_nodes/nodes_kling.py

@@ -858,7 +858,7 @@ class OmniProFirstLastFrameNode(IO.ComfyNode):
                     tooltip="A text prompt describing the video content. "
                     "This can include both positive and negative descriptions.",
                 ),
-                IO.Combo.Input("duration", options=["5", "10"]),
+                IO.Int.Input("duration", default=5, min=3, max=10, display_mode=IO.NumberDisplay.slider),
                 IO.Image.Input("first_frame"),
                 IO.Image.Input(
                     "end_frame",
@@ -897,6 +897,10 @@ class OmniProFirstLastFrameNode(IO.ComfyNode):
         validate_string(prompt, min_length=1, max_length=2500)
         if end_frame is not None and reference_images is not None:
             raise ValueError("The 'end_frame' input cannot be used simultaneously with 'reference_images'.")
+        if duration not in (5, 10) and end_frame is None and reference_images is None:
+            raise ValueError(
+                "Duration is only supported for 5 or 10 seconds if there is no end frame or reference images."
+            )
         validate_image_dimensions(first_frame, min_width=300, min_height=300)
         validate_image_aspect_ratio(first_frame, (1, 2.5), (2.5, 1))
         image_list: list[OmniParamImage] = [

comfy_api_nodes/nodes_topaz.py

@@ -23,10 +23,6 @@ UPSCALER_MODELS_MAP = {
     "Starlight (Astra) Fast": "slf-1",
     "Starlight (Astra) Creative": "slc-1",
 }
-UPSCALER_VALUES_MAP = {
-    "FullHD (1080p)": 1920,
-    "4K (2160p)": 3840,
-}
 
 
 class TopazImageEnhance(IO.ComfyNode):
@@ -214,7 +210,7 @@ class TopazVideoEnhance(IO.ComfyNode):
                 IO.Video.Input("video"),
                 IO.Boolean.Input("upscaler_enabled", default=True),
                 IO.Combo.Input("upscaler_model", options=list(UPSCALER_MODELS_MAP.keys())),
-                IO.Combo.Input("upscaler_resolution", options=list(UPSCALER_VALUES_MAP.keys())),
+                IO.Combo.Input("upscaler_resolution", options=["FullHD (1080p)", "4K (2160p)"]),
                 IO.Combo.Input(
                     "upscaler_creativity",
                     options=["low", "middle", "high"],
@@ -306,8 +302,33 @@ class TopazVideoEnhance(IO.ComfyNode):
         target_frame_rate = src_frame_rate
         filters = []
         if upscaler_enabled:
-            target_width = UPSCALER_VALUES_MAP[upscaler_resolution]
-            target_height = UPSCALER_VALUES_MAP[upscaler_resolution]
+            if "1080p" in upscaler_resolution:
+                target_pixel_p = 1080
+                max_long_side = 1920
+            else:
+                target_pixel_p = 2160
+                max_long_side = 3840
+            ar = src_width / src_height
+            if src_width >= src_height:
+                # Landscape or Square; Attempt to set height to target (e.g., 2160), calculate width
+                target_height = target_pixel_p
+                target_width = int(target_height * ar)
+                # Check if width exceeds standard bounds (for ultra-wide e.g., 21:9 ARs)
+                if target_width > max_long_side:
+                    target_width = max_long_side
+                    target_height = int(target_width / ar)
+            else:
+                # Portrait; Attempt to set width to target (e.g., 2160), calculate height
+                target_width = target_pixel_p
+                target_height = int(target_width / ar)
+                # Check if height exceeds standard bounds
+                if target_height > max_long_side:
+                    target_height = max_long_side
+                    target_width = int(target_height * ar)
+            if target_width % 2 != 0:
+                target_width += 1
+            if target_height % 2 != 0:
+                target_height += 1
             filters.append(
                 topaz_api.VideoEnhancementFilter(
                     model=UPSCALER_MODELS_MAP[upscaler_model],

comfy_api_nodes/nodes_wan.py

@@ -46,14 +46,14 @@ class Txt2ImageParametersField(BaseModel):
     n: int = Field(1, description="Number of images to generate.")  # we support only value=1
     seed: int = Field(..., ge=0, le=2147483647)
     prompt_extend: bool = Field(True)
-    watermark: bool = Field(True)
+    watermark: bool = Field(False)
 
 
 class Image2ImageParametersField(BaseModel):
     size: str | None = Field(None)
     n: int = Field(1, description="Number of images to generate.")  # we support only value=1
     seed: int = Field(..., ge=0, le=2147483647)
-    watermark: bool = Field(True)
+    watermark: bool = Field(False)
 
 
 class Text2VideoParametersField(BaseModel):
@@ -61,7 +61,7 @@ class Text2VideoParametersField(BaseModel):
     seed: int = Field(..., ge=0, le=2147483647)
     duration: int = Field(5, ge=5, le=15)
     prompt_extend: bool = Field(True)
-    watermark: bool = Field(True)
+    watermark: bool = Field(False)
     audio: bool = Field(False, description="Whether to generate audio automatically.")
     shot_type: str = Field("single")
 
@@ -71,7 +71,7 @@ class Image2VideoParametersField(BaseModel):
     seed: int = Field(..., ge=0, le=2147483647)
     duration: int = Field(5, ge=5, le=15)
     prompt_extend: bool = Field(True)
-    watermark: bool = Field(True)
+    watermark: bool = Field(False)
     audio: bool = Field(False, description="Whether to generate audio automatically.")
     shot_type: str = Field("single")
 
@@ -208,7 +208,7 @@ class WanTextToImageApi(IO.ComfyNode):
                 ),
                 IO.Boolean.Input(
                     "watermark",
-                    default=True,
+                    default=False,
                     tooltip="Whether to add an AI-generated watermark to the result.",
                     optional=True,
                 ),
@@ -234,7 +234,7 @@ class WanTextToImageApi(IO.ComfyNode):
         height: int = 1024,
         seed: int = 0,
         prompt_extend: bool = True,
-        watermark: bool = True,
+        watermark: bool = False,
     ):
         initial_response = await sync_op(
             cls,
@@ -327,7 +327,7 @@ class WanImageToImageApi(IO.ComfyNode):
                 ),
                 IO.Boolean.Input(
                     "watermark",
-                    default=True,
+                    default=False,
                     tooltip="Whether to add an AI-generated watermark to the result.",
                     optional=True,
                 ),
@@ -353,7 +353,7 @@ class WanImageToImageApi(IO.ComfyNode):
         # width: int = 1024,
         # height: int = 1024,
         seed: int = 0,
-        watermark: bool = True,
+        watermark: bool = False,
     ):
         n_images = get_number_of_images(image)
         if n_images not in (1, 2):
@@ -476,7 +476,7 @@ class WanTextToVideoApi(IO.ComfyNode):
                 ),
                 IO.Boolean.Input(
                     "watermark",
-                    default=True,
+                    default=False,
                     tooltip="Whether to add an AI-generated watermark to the result.",
                     optional=True,
                 ),
@@ -512,7 +512,7 @@ class WanTextToVideoApi(IO.ComfyNode):
         seed: int = 0,
         generate_audio: bool = False,
         prompt_extend: bool = True,
-        watermark: bool = True,
+        watermark: bool = False,
         shot_type: str = "single",
     ):
         if "480p" in size and model == "wan2.6-t2v":
@@ -637,7 +637,7 @@ class WanImageToVideoApi(IO.ComfyNode):
                 ),
                 IO.Boolean.Input(
                     "watermark",
-                    default=True,
+                    default=False,
                     tooltip="Whether to add an AI-generated watermark to the result.",
                     optional=True,
                 ),
@@ -674,7 +674,7 @@ class WanImageToVideoApi(IO.ComfyNode):
         seed: int = 0,
         generate_audio: bool = False,
         prompt_extend: bool = True,
-        watermark: bool = True,
+        watermark: bool = False,
         shot_type: str = "single",
     ):
         if get_number_of_images(image) != 1:

comfy_extras/nodes_custom_sampler.py

@@ -760,8 +760,12 @@ class SamplerCustom(io.ComfyNode):
         out = latent.copy()
         out["samples"] = samples
         if "x0" in x0_output:
+            x0_out = model.model.process_latent_out(x0_output["x0"].cpu())
+            if samples.is_nested:
+                latent_shapes = [x.shape for x in samples.unbind()]
+                x0_out = comfy.nested_tensor.NestedTensor(comfy.utils.unpack_latents(x0_out, latent_shapes))
             out_denoised = latent.copy()
-            out_denoised["samples"] = model.model.process_latent_out(x0_output["x0"].cpu())
+            out_denoised["samples"] = x0_out
         else:
             out_denoised = out
         return io.NodeOutput(out, out_denoised)
@@ -948,8 +952,12 @@ class SamplerCustomAdvanced(io.ComfyNode):
         out = latent.copy()
         out["samples"] = samples
         if "x0" in x0_output:
+            x0_out = guider.model_patcher.model.process_latent_out(x0_output["x0"].cpu())
+            if samples.is_nested:
+                latent_shapes = [x.shape for x in samples.unbind()]
+                x0_out = comfy.nested_tensor.NestedTensor(comfy.utils.unpack_latents(x0_out, latent_shapes))
             out_denoised = latent.copy()
-            out_denoised["samples"] = guider.model_patcher.model.process_latent_out(x0_output["x0"].cpu())
+            out_denoised["samples"] = x0_out
         else:
             out_denoised = out
         return io.NodeOutput(out, out_denoised)

comfy_extras/nodes_model_patch.py

@@ -348,7 +348,7 @@ class ZImageControlPatch:
             if self.mask is None:
                 mask_ = torch.zeros_like(inpaint_image_latent)[:, :1]
             else:
-                mask_ = comfy.utils.common_upscale(self.mask.view(self.mask.shape[0], -1, self.mask.shape[-2], self.mask.shape[-1]).mean(dim=1, keepdim=True), inpaint_image_latent.shape[-1], inpaint_image_latent.shape[-2], "nearest", "center")
+                mask_ = comfy.utils.common_upscale(self.mask.view(self.mask.shape[0], -1, self.mask.shape[-2], self.mask.shape[-1]).mean(dim=1, keepdim=True).to(device=inpaint_image_latent.device), inpaint_image_latent.shape[-1], inpaint_image_latent.shape[-2], "nearest", "center")
 
             if latent_image is None:
                 latent_image = comfy.latent_formats.Flux().process_in(self.vae.encode(torch.ones_like(inpaint_image) * 0.5))

comfy_extras/nodes_qwen.py

@@ -3,7 +3,9 @@ import comfy.utils
 import math
 from typing_extensions import override
 from comfy_api.latest import ComfyExtension, io
-
+import comfy.model_management
+import torch
+import nodes
 
 class TextEncodeQwenImageEdit(io.ComfyNode):
     @classmethod
@@ -104,12 +106,37 @@ class TextEncodeQwenImageEditPlus(io.ComfyNode):
         return io.NodeOutput(conditioning)
 
 
+class EmptyQwenImageLayeredLatentImage(io.ComfyNode):
+    @classmethod
+    def define_schema(cls):
+        return io.Schema(
+            node_id="EmptyQwenImageLayeredLatentImage",
+            display_name="Empty Qwen Image Layered Latent",
+            category="latent/qwen",
+            inputs=[
+                io.Int.Input("width", default=640, min=16, max=nodes.MAX_RESOLUTION, step=16),
+                io.Int.Input("height", default=640, min=16, max=nodes.MAX_RESOLUTION, step=16),
+                io.Int.Input("layers", default=3, min=0, max=nodes.MAX_RESOLUTION, step=1),
+                io.Int.Input("batch_size", default=1, min=1, max=4096),
+            ],
+            outputs=[
+                io.Latent.Output(),
+            ],
+        )
+
+    @classmethod
+    def execute(cls, width, height, layers, batch_size=1) -> io.NodeOutput:
+        latent = torch.zeros([batch_size, 16, layers + 1, height // 8, width // 8], device=comfy.model_management.intermediate_device())
+        return io.NodeOutput({"samples": latent})
+
+
 class QwenExtension(ComfyExtension):
     @override
     async def get_node_list(self) -> list[type[io.ComfyNode]]:
         return [
             TextEncodeQwenImageEdit,
             TextEncodeQwenImageEditPlus,
+            EmptyQwenImageLayeredLatentImage,
         ]
 
 

nodes.py

@@ -970,7 +970,7 @@ class DualCLIPLoader:
     def INPUT_TYPES(s):
         return {"required": { "clip_name1": (folder_paths.get_filename_list("text_encoders"), ),
                               "clip_name2": (folder_paths.get_filename_list("text_encoders"), ),
-                              "type": (["sdxl", "sd3", "flux", "hunyuan_video", "hidream", "hunyuan_image", "hunyuan_video_15", "kandinsky5", "kandinsky5_image"], ),
+                              "type": (["sdxl", "sd3", "flux", "hunyuan_video", "hidream", "hunyuan_image", "hunyuan_video_15", "kandinsky5", "kandinsky5_image", "newbie"], ),
                               },
                 "optional": {
                               "device": (["default", "cpu"], {"advanced": True}),
@@ -980,7 +980,7 @@ class DualCLIPLoader:
 
     CATEGORY = "advanced/loaders"
 
-    DESCRIPTION = "[Recipes]\n\nsdxl: clip-l, clip-g\nsd3: clip-l, clip-g / clip-l, t5 / clip-g, t5\nflux: clip-l, t5\nhidream: at least one of t5 or llama, recommended t5 and llama\nhunyuan_image: qwen2.5vl 7b and byt5 small"
+    DESCRIPTION = "[Recipes]\n\nsdxl: clip-l, clip-g\nsd3: clip-l, clip-g / clip-l, t5 / clip-g, t5\nflux: clip-l, t5\nhidream: at least one of t5 or llama, recommended t5 and llama\nhunyuan_image: qwen2.5vl 7b and byt5 small\nnewbie: gemma-3-4b-it, jina clip v2"
 
     def load_clip(self, clip_name1, clip_name2, type, device="default"):
         clip_type = getattr(comfy.sd.CLIPType, type.upper(), comfy.sd.CLIPType.STABLE_DIFFUSION)

requirements.txt

@@ -1,5 +1,5 @@
 comfyui-frontend-package==1.34.9
-comfyui-workflow-templates==0.7.60
+comfyui-workflow-templates==0.7.62
 comfyui-embedded-docs==0.3.1
 torch
 torchsde