Merge branch 'master' into assets-redo

Author: Kosinkadink

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 @@ def get_resized_cond(self, cond_in: list[dict], x_in: torch.Tensor, window: Inde
         # 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 @@ def _forward(self, x, timesteps, context, num_tokens, attention_mask=None, trans
             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/ldm/qwen_image/model.py

@@ -61,7 +61,7 @@ def apply_rotary_emb(x, freqs_cis):
 
 
 class QwenTimestepProjEmbeddings(nn.Module):
-    def __init__(self, embedding_dim, pooled_projection_dim, dtype=None, device=None, operations=None):
+    def __init__(self, embedding_dim, pooled_projection_dim, use_additional_t_cond=False, dtype=None, device=None, operations=None):
         super().__init__()
         self.time_proj = Timesteps(num_channels=256, flip_sin_to_cos=True, downscale_freq_shift=0, scale=1000)
         self.timestep_embedder = TimestepEmbedding(
@@ -72,9 +72,19 @@ def __init__(self, embedding_dim, pooled_projection_dim, dtype=None, device=None
             operations=operations
         )
 
-    def forward(self, timestep, hidden_states):
+        self.use_additional_t_cond = use_additional_t_cond
+        if self.use_additional_t_cond:
+            self.addition_t_embedding = operations.Embedding(2, embedding_dim, device=device, dtype=dtype)
+
+    def forward(self, timestep, hidden_states, addition_t_cond=None):
         timesteps_proj = self.time_proj(timestep)
         timesteps_emb = self.timestep_embedder(timesteps_proj.to(dtype=hidden_states.dtype))
+
+        if self.use_additional_t_cond:
+            if addition_t_cond is None:
+                addition_t_cond = torch.zeros((timesteps_emb.shape[0]), device=timesteps_emb.device, dtype=torch.long)
+            timesteps_emb += self.addition_t_embedding(addition_t_cond, out_dtype=timesteps_emb.dtype)
+
         return timesteps_emb
 
 
@@ -320,11 +330,11 @@ def __init__(
         num_attention_heads: int = 24,
         joint_attention_dim: int = 3584,
         pooled_projection_dim: int = 768,
-        guidance_embeds: bool = False,
         axes_dims_rope: Tuple[int, int, int] = (16, 56, 56),
         default_ref_method="index",
         image_model=None,
         final_layer=True,
+        use_additional_t_cond=False,
         dtype=None,
         device=None,
         operations=None,
@@ -342,6 +352,7 @@ def __init__(
         self.time_text_embed = QwenTimestepProjEmbeddings(
             embedding_dim=self.inner_dim,
             pooled_projection_dim=pooled_projection_dim,
+            use_additional_t_cond=use_additional_t_cond,
             dtype=dtype,
             device=device,
             operations=operations
@@ -375,36 +386,42 @@ def process_img(self, x, index=0, h_offset=0, w_offset=0):
         patch_size = self.patch_size
         hidden_states = comfy.ldm.common_dit.pad_to_patch_size(x, (1, self.patch_size, self.patch_size))
         orig_shape = hidden_states.shape
-        hidden_states = hidden_states.view(orig_shape[0], orig_shape[1], orig_shape[-2] // 2, 2, orig_shape[-1] // 2, 2)
-        hidden_states = hidden_states.permute(0, 2, 4, 1, 3, 5)
-        hidden_states = hidden_states.reshape(orig_shape[0], (orig_shape[-2] // 2) * (orig_shape[-1] // 2), orig_shape[1] * 4)
+        hidden_states = hidden_states.view(orig_shape[0], orig_shape[1], orig_shape[-3], orig_shape[-2] // 2, 2, orig_shape[-1] // 2, 2)
+        hidden_states = hidden_states.permute(0, 2, 3, 5, 1, 4, 6)
+        hidden_states = hidden_states.reshape(orig_shape[0], orig_shape[-3] * (orig_shape[-2] // 2) * (orig_shape[-1] // 2), orig_shape[1] * 4)
+        t_len = t
         h_len = ((h + (patch_size // 2)) // patch_size)
         w_len = ((w + (patch_size // 2)) // patch_size)
 
         h_offset = ((h_offset + (patch_size // 2)) // patch_size)
         w_offset = ((w_offset + (patch_size // 2)) // patch_size)
 
-        img_ids = torch.zeros((h_len, w_len, 3), device=x.device)
-        img_ids[:, :, 0] = img_ids[:, :, 1] + index
-        img_ids[:, :, 1] = img_ids[:, :, 1] + torch.linspace(h_offset, h_len - 1 + h_offset, steps=h_len, device=x.device, dtype=x.dtype).unsqueeze(1) - (h_len // 2)
-        img_ids[:, :, 2] = img_ids[:, :, 2] + torch.linspace(w_offset, w_len - 1 + w_offset, steps=w_len, device=x.device, dtype=x.dtype).unsqueeze(0) - (w_len // 2)
-        return hidden_states, repeat(img_ids, "h w c -> b (h w) c", b=bs), orig_shape
+        img_ids = torch.zeros((t_len, h_len, w_len, 3), device=x.device)
+
+        if t_len > 1:
+            img_ids[:, :, :, 0] = img_ids[:, :, :, 0] + torch.linspace(0, t_len - 1, steps=t_len, device=x.device, dtype=x.dtype).unsqueeze(1).unsqueeze(1)
+        else:
+            img_ids[:, :, :, 0] = img_ids[:, :, :, 0] + index
+
+        img_ids[:, :, :, 1] = img_ids[:, :, :, 1] + torch.linspace(h_offset, h_len - 1 + h_offset, steps=h_len, device=x.device, dtype=x.dtype).unsqueeze(1).unsqueeze(0) - (h_len // 2)
+        img_ids[:, :, :, 2] = img_ids[:, :, :, 2] + torch.linspace(w_offset, w_len - 1 + w_offset, steps=w_len, device=x.device, dtype=x.dtype).unsqueeze(0).unsqueeze(0) - (w_len // 2)
+        return hidden_states, repeat(img_ids, "t h w c -> b (t h w) c", b=bs), orig_shape
 
-    def forward(self, x, timestep, context, attention_mask=None, guidance=None, ref_latents=None, transformer_options={}, **kwargs):
+    def forward(self, x, timestep, context, attention_mask=None, ref_latents=None, additional_t_cond=None, transformer_options={}, **kwargs):
         return comfy.patcher_extension.WrapperExecutor.new_class_executor(
             self._forward,
             self,
             comfy.patcher_extension.get_all_wrappers(comfy.patcher_extension.WrappersMP.DIFFUSION_MODEL, transformer_options)
-        ).execute(x, timestep, context, attention_mask, guidance, ref_latents, transformer_options, **kwargs)
+        ).execute(x, timestep, context, attention_mask, ref_latents, additional_t_cond, transformer_options, **kwargs)
 
     def _forward(
         self,
         x,
         timesteps,
         context,
         attention_mask=None,
-        guidance: torch.Tensor = None,
         ref_latents=None,
+        additional_t_cond=None,
         transformer_options={},
         control=None,
         **kwargs
@@ -423,12 +440,17 @@ def _forward(
             index = 0
             ref_method = kwargs.get("ref_latents_method", self.default_ref_method)
             index_ref_method = (ref_method == "index") or (ref_method == "index_timestep_zero")
+            negative_ref_method = ref_method == "negative_index"
             timestep_zero = ref_method == "index_timestep_zero"
             for ref in ref_latents:
                 if index_ref_method:
                     index += 1
                     h_offset = 0
                     w_offset = 0
+                elif negative_ref_method:
+                    index -= 1
+                    h_offset = 0
+                    w_offset = 0
                 else:
                     index = 1
                     h_offset = 0
@@ -458,14 +480,7 @@ def _forward(
         encoder_hidden_states = self.txt_norm(encoder_hidden_states)
         encoder_hidden_states = self.txt_in(encoder_hidden_states)
 
-        if guidance is not None:
-            guidance = guidance * 1000
-
-        temb = (
-            self.time_text_embed(timestep, hidden_states)
-            if guidance is None
-            else self.time_text_embed(timestep, guidance, hidden_states)
-        )
+        temb = self.time_text_embed(timestep, hidden_states, additional_t_cond)
 
         patches_replace = transformer_options.get("patches_replace", {})
         patches = transformer_options.get("patches", {})
@@ -513,6 +528,6 @@ def block_wrap(args):
         hidden_states = self.norm_out(hidden_states, temb)
         hidden_states = self.proj_out(hidden_states)
 
-        hidden_states = hidden_states[:, :num_embeds].view(orig_shape[0], orig_shape[-2] // 2, orig_shape[-1] // 2, orig_shape[1], 2, 2)
-        hidden_states = hidden_states.permute(0, 3, 1, 4, 2, 5)
+        hidden_states = hidden_states[:, :num_embeds].view(orig_shape[0], orig_shape[-3], orig_shape[-2] // 2, orig_shape[-1] // 2, orig_shape[1], 2, 2)
+        hidden_states = hidden_states.permute(0, 4, 1, 2, 5, 3, 6)
         return hidden_states.reshape(orig_shape)[:, :, :, :x.shape[-2], :x.shape[-1]]

comfy/model_base.py

@@ -1110,7 +1110,7 @@ def extra_conds(self, **kwargs):
             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
@@ -620,6 +621,9 @@ def detect_unet_config(state_dict, key_prefix, metadata=None):
         dit_config["num_layers"] = count_blocks(state_dict_keys, '{}transformer_blocks.'.format(key_prefix) + '{}.')
         if "{}__index_timestep_zero__".format(key_prefix) in state_dict_keys:  # 2511
             dit_config["default_ref_method"] = "index_timestep_zero"
+        if "{}time_text_embed.addition_t_embedding.weight".format(key_prefix) in state_dict_keys:  # Layered
+            dit_config["use_additional_t_cond"] = True
+            dit_config["default_ref_method"] = "negative_index"
         return dit_config
 
     if '{}visual_transformer_blocks.0.cross_attention.key_norm.weight'.format(key_prefix) in state_dict_keys: # Kandinsky 5

comfy/model_management.py

@@ -26,6 +26,7 @@
 import platform
 import weakref
 import gc
+import os
 
 class VRAMState(Enum):
     DISABLED = 0    #No vram present: no need to move models to vram
@@ -333,13 +334,15 @@ def amd_min_version(device=None, min_rdna_version=0):
 SUPPORT_FP8_OPS = args.supports_fp8_compute
 
 AMD_RDNA2_AND_OLDER_ARCH = ["gfx1030", "gfx1031", "gfx1010", "gfx1011", "gfx1012", "gfx906", "gfx900", "gfx803"]
+AMD_ENABLE_MIOPEN_ENV = 'COMFYUI_ENABLE_MIOPEN'
 
 try:
     if is_amd():
         arch = torch.cuda.get_device_properties(get_torch_device()).gcnArchName
         if not (any((a in arch) for a in AMD_RDNA2_AND_OLDER_ARCH)):
-            torch.backends.cudnn.enabled = False  # Seems to improve things a lot on AMD
-            logging.info("Set: torch.backends.cudnn.enabled = False for better AMD performance.")
+            if os.getenv(AMD_ENABLE_MIOPEN_ENV) != '1':
+                torch.backends.cudnn.enabled = False  # Seems to improve things a lot on AMD
+                logging.info("Set: torch.backends.cudnn.enabled = False for better AMD performance.")
 
         try:
             rocm_version = tuple(map(int, str(torch.version.hip).split(".")[:2]))

comfy/samplers.py

@@ -984,9 +984,6 @@ def outer_sample(self, noise, latent_image, sampler, sigmas, denoise_mask=None,
         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 @@ def sample(self, noise, latent_image, sampler, sigmas, denoise_mask=None, callba
         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.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 @@ class EmptyClass:
         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 @@ class EmptyClass:
         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 @@ def __init__(self, tokenizer_path=None, max_length=77, pad_with_end=True, embedd
         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 @@ def tokenize_with_weights(self, text:str, return_word_ids=False, tokenizer_optio
         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)