Merge branch 'master' into v3-improvements

Author: Kosinkadink

comfy/cli_args.py

@@ -97,6 +97,13 @@ class LatentPreviewMethod(enum.Enum):
     Latent2RGB = "latent2rgb"
     TAESD = "taesd"
 
+    @classmethod
+    def from_string(cls, value: str):
+        for member in cls:
+            if member.value == value:
+                return member
+        return None
+
 parser.add_argument("--preview-method", type=LatentPreviewMethod, default=LatentPreviewMethod.NoPreviews, help="Default preview method for sampler nodes.", action=EnumAction)
 
 parser.add_argument("--preview-size", type=int, default=512, help="Sets the maximum preview size for sampler nodes.")

comfy/context_windows.py

@@ -87,6 +87,7 @@ class IndexListCallbacks:
     COMBINE_CONTEXT_WINDOW_RESULTS = "combine_context_window_results"
     EXECUTE_START = "execute_start"
     EXECUTE_CLEANUP = "execute_cleanup"
+    RESIZE_COND_ITEM = "resize_cond_item"
 
     def init_callbacks(self):
         return {}
@@ -166,6 +167,18 @@ def get_resized_cond(self, cond_in: list[dict], x_in: torch.Tensor, window: Inde
                         new_cond_item = cond_item.copy()
                         # when in dictionary, look for tensors and CONDCrossAttn [comfy/conds.py] (has cond attr that is a tensor)
                         for cond_key, cond_value in new_cond_item.items():
+                            # Allow callbacks to handle custom conditioning items
+                            handled = False
+                            for callback in comfy.patcher_extension.get_all_callbacks(
+                                IndexListCallbacks.RESIZE_COND_ITEM, self.callbacks
+                            ):
+                                result = callback(cond_key, cond_value, window, x_in, device, new_cond_item)
+                                if result is not None:
+                                    new_cond_item[cond_key] = result
+                                    handled = True
+                                    break
+                            if handled:
+                                continue
                             if isinstance(cond_value, torch.Tensor):
                                 if (self.dim < cond_value.ndim and cond_value(self.dim) == x_in.size(self.dim)) or \
                                    (cond_value.ndim < self.dim and cond_value.size(0) == x_in.size(self.dim)):

comfy/ldm/lumina/model.py

@@ -634,8 +634,11 @@ def _forward(self, x, timesteps, context, num_tokens, attention_mask=None, trans
         img, mask, img_size, cap_size, freqs_cis = self.patchify_and_embed(x, cap_feats, cap_mask, adaln_input, num_tokens, transformer_options=transformer_options)
         freqs_cis = freqs_cis.to(img.device)
 
+        transformer_options["total_blocks"] = len(self.layers)
+        transformer_options["block_type"] = "double"
         img_input = img
         for i, layer in enumerate(self.layers):
+            transformer_options["block_index"] = i
             img = layer(img, mask, freqs_cis, adaln_input, transformer_options=transformer_options)
             if "double_block" in patches:
                 for p in patches["double_block"]:

comfy/ldm/qwen_image/model.py

@@ -218,9 +218,24 @@ def __init__(
             operations=operations,
         )
 
-    def _modulate(self, x: torch.Tensor, mod_params: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
+    def _apply_gate(self, x, y, gate, timestep_zero_index=None):
+        if timestep_zero_index is not None:
+            return y + torch.cat((x[:, :timestep_zero_index] * gate[0], x[:, timestep_zero_index:] * gate[1]), dim=1)
+        else:
+            return torch.addcmul(y, gate, x)
+
+    def _modulate(self, x: torch.Tensor, mod_params: torch.Tensor, timestep_zero_index=None) -> Tuple[torch.Tensor, torch.Tensor]:
         shift, scale, gate = torch.chunk(mod_params, 3, dim=-1)
-        return torch.addcmul(shift.unsqueeze(1), x, 1 + scale.unsqueeze(1)), gate.unsqueeze(1)
+        if timestep_zero_index is not None:
+            actual_batch = shift.size(0) // 2
+            shift, shift_0 = shift[:actual_batch], shift[actual_batch:]
+            scale, scale_0 = scale[:actual_batch], scale[actual_batch:]
+            gate, gate_0 = gate[:actual_batch], gate[actual_batch:]
+            reg = torch.addcmul(shift.unsqueeze(1), x[:, :timestep_zero_index], 1 + scale.unsqueeze(1))
+            zero = torch.addcmul(shift_0.unsqueeze(1), x[:, timestep_zero_index:], 1 + scale_0.unsqueeze(1))
+            return torch.cat((reg, zero), dim=1), (gate.unsqueeze(1), gate_0.unsqueeze(1))
+        else:
+            return torch.addcmul(shift.unsqueeze(1), x, 1 + scale.unsqueeze(1)), gate.unsqueeze(1)
 
     def forward(
         self,
@@ -229,14 +244,19 @@ def forward(
         encoder_hidden_states_mask: torch.Tensor,
         temb: torch.Tensor,
         image_rotary_emb: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
+        timestep_zero_index=None,
         transformer_options={},
     ) -> Tuple[torch.Tensor, torch.Tensor]:
         img_mod_params = self.img_mod(temb)
+
+        if timestep_zero_index is not None:
+            temb = temb.chunk(2, dim=0)[0]
+
         txt_mod_params = self.txt_mod(temb)
         img_mod1, img_mod2 = img_mod_params.chunk(2, dim=-1)
         txt_mod1, txt_mod2 = txt_mod_params.chunk(2, dim=-1)
 
-        img_modulated, img_gate1 = self._modulate(self.img_norm1(hidden_states), img_mod1)
+        img_modulated, img_gate1 = self._modulate(self.img_norm1(hidden_states), img_mod1, timestep_zero_index)
         del img_mod1
         txt_modulated, txt_gate1 = self._modulate(self.txt_norm1(encoder_hidden_states), txt_mod1)
         del txt_mod1
@@ -251,15 +271,15 @@ def forward(
         del img_modulated
         del txt_modulated
 
-        hidden_states = hidden_states + img_gate1 * img_attn_output
+        hidden_states = self._apply_gate(img_attn_output, hidden_states, img_gate1, timestep_zero_index)
         encoder_hidden_states = encoder_hidden_states + txt_gate1 * txt_attn_output
         del img_attn_output
         del txt_attn_output
         del img_gate1
         del txt_gate1
 
-        img_modulated2, img_gate2 = self._modulate(self.img_norm2(hidden_states), img_mod2)
-        hidden_states = torch.addcmul(hidden_states, img_gate2, self.img_mlp(img_modulated2))
+        img_modulated2, img_gate2 = self._modulate(self.img_norm2(hidden_states), img_mod2, timestep_zero_index)
+        hidden_states = self._apply_gate(self.img_mlp(img_modulated2), hidden_states, img_gate2, timestep_zero_index)
 
         txt_modulated2, txt_gate2 = self._modulate(self.txt_norm2(encoder_hidden_states), txt_mod2)
         encoder_hidden_states = torch.addcmul(encoder_hidden_states, txt_gate2, self.txt_mlp(txt_modulated2))
@@ -391,11 +411,14 @@ def _forward(
         hidden_states, img_ids, orig_shape = self.process_img(x)
         num_embeds = hidden_states.shape[1]
 
+        timestep_zero_index = None
         if ref_latents is not None:
             h = 0
             w = 0
             index = 0
-            index_ref_method = kwargs.get("ref_latents_method", "index") == "index"
+            ref_method = kwargs.get("ref_latents_method", "index")
+            index_ref_method = (ref_method == "index") or (ref_method == "index_timestep_zero")
+            timestep_zero = ref_method == "index_timestep_zero"
             for ref in ref_latents:
                 if index_ref_method:
                     index += 1
@@ -415,6 +438,10 @@ def _forward(
                 kontext, kontext_ids, _ = self.process_img(ref, index=index, h_offset=h_offset, w_offset=w_offset)
                 hidden_states = torch.cat([hidden_states, kontext], dim=1)
                 img_ids = torch.cat([img_ids, kontext_ids], dim=1)
+            if timestep_zero:
+                if index > 0:
+                    timestep = torch.cat([timestep, timestep * 0], dim=0)
+                    timestep_zero_index = num_embeds
 
         txt_start = round(max(((x.shape[-1] + (self.patch_size // 2)) // self.patch_size) // 2, ((x.shape[-2] + (self.patch_size // 2)) // self.patch_size) // 2))
         txt_ids = torch.arange(txt_start, txt_start + context.shape[1], device=x.device).reshape(1, -1, 1).repeat(x.shape[0], 1, 3)
@@ -446,7 +473,7 @@ def _forward(
             if ("double_block", i) in blocks_replace:
                 def block_wrap(args):
                     out = {}
-                    out["txt"], out["img"] = block(hidden_states=args["img"], encoder_hidden_states=args["txt"], encoder_hidden_states_mask=encoder_hidden_states_mask, temb=args["vec"], image_rotary_emb=args["pe"], transformer_options=args["transformer_options"])
+                    out["txt"], out["img"] = block(hidden_states=args["img"], encoder_hidden_states=args["txt"], encoder_hidden_states_mask=encoder_hidden_states_mask, temb=args["vec"], image_rotary_emb=args["pe"], timestep_zero_index=timestep_zero_index, transformer_options=args["transformer_options"])
                     return out
                 out = blocks_replace[("double_block", i)]({"img": hidden_states, "txt": encoder_hidden_states, "vec": temb, "pe": image_rotary_emb, "transformer_options": transformer_options}, {"original_block": block_wrap})
                 hidden_states = out["img"]
@@ -458,6 +485,7 @@ def block_wrap(args):
                     encoder_hidden_states_mask=encoder_hidden_states_mask,
                     temb=temb,
                     image_rotary_emb=image_rotary_emb,
+                    timestep_zero_index=timestep_zero_index,
                     transformer_options=transformer_options,
                 )
 
@@ -474,6 +502,9 @@ def block_wrap(args):
                     if add is not None:
                         hidden_states[:, :add.shape[1]] += add
 
+        if timestep_zero_index is not None:
+            temb = temb.chunk(2, dim=0)[0]
+
         hidden_states = self.norm_out(hidden_states, temb)
         hidden_states = self.proj_out(hidden_states)
 

comfy/ldm/wan/model.py

@@ -568,7 +568,10 @@ def forward_orig(
 
         patches_replace = transformer_options.get("patches_replace", {})
         blocks_replace = patches_replace.get("dit", {})
+        transformer_options["total_blocks"] = len(self.blocks)
+        transformer_options["block_type"] = "double"
         for i, block in enumerate(self.blocks):
+            transformer_options["block_index"] = i
             if ("double_block", i) in blocks_replace:
                 def block_wrap(args):
                     out = {}
@@ -763,7 +766,10 @@ def forward_orig(
 
         patches_replace = transformer_options.get("patches_replace", {})
         blocks_replace = patches_replace.get("dit", {})
+        transformer_options["total_blocks"] = len(self.blocks)
+        transformer_options["block_type"] = "double"
         for i, block in enumerate(self.blocks):
+            transformer_options["block_index"] = i
             if ("double_block", i) in blocks_replace:
                 def block_wrap(args):
                     out = {}
@@ -862,7 +868,10 @@ def forward_orig(
 
         patches_replace = transformer_options.get("patches_replace", {})
         blocks_replace = patches_replace.get("dit", {})
+        transformer_options["total_blocks"] = len(self.blocks)
+        transformer_options["block_type"] = "double"
         for i, block in enumerate(self.blocks):
+            transformer_options["block_index"] = i
             if ("double_block", i) in blocks_replace:
                 def block_wrap(args):
                     out = {}
@@ -1326,16 +1335,19 @@ def forward_orig(
 
         patches_replace = transformer_options.get("patches_replace", {})
         blocks_replace = patches_replace.get("dit", {})
+        transformer_options["total_blocks"] = len(self.blocks)
+        transformer_options["block_type"] = "double"
         for i, block in enumerate(self.blocks):
+            transformer_options["block_index"] = i
             if ("double_block", i) in blocks_replace:
                 def block_wrap(args):
                     out = {}
-                    out["img"] = block(args["img"], context=args["txt"], e=args["vec"], freqs=args["pe"])
+                    out["img"] = block(args["img"], context=args["txt"], e=args["vec"], freqs=args["pe"], transformer_options=args["transformer_options"])
                     return out
-                out = blocks_replace[("double_block", i)]({"img": x, "txt": context, "vec": e0, "pe": freqs}, {"original_block": block_wrap})
+                out = blocks_replace[("double_block", i)]({"img": x, "txt": context, "vec": e0, "pe": freqs, "transformer_options": transformer_options}, {"original_block": block_wrap})
                 x = out["img"]
             else:
-                x = block(x, e=e0, freqs=freqs, context=context)
+                x = block(x, e=e0, freqs=freqs, context=context, transformer_options=transformer_options)
             if audio_emb is not None:
                 x = self.audio_injector(x, i, audio_emb, audio_emb_global, seq_len)
         # head
@@ -1574,7 +1586,10 @@ def forward_orig(
 
         patches_replace = transformer_options.get("patches_replace", {})
         blocks_replace = patches_replace.get("dit", {})
+        transformer_options["total_blocks"] = len(self.blocks)
+        transformer_options["block_type"] = "double"
         for i, block in enumerate(self.blocks):
+            transformer_options["block_index"] = i
             if ("double_block", i) in blocks_replace:
                 def block_wrap(args):
                     out = {}

comfy/ldm/wan/model_animate.py

@@ -523,7 +523,10 @@ def forward_orig(
 
         patches_replace = transformer_options.get("patches_replace", {})
         blocks_replace = patches_replace.get("dit", {})
+        transformer_options["total_blocks"] = len(self.blocks)
+        transformer_options["block_type"] = "double"
         for i, block in enumerate(self.blocks):
+            transformer_options["block_index"] = i
             if ("double_block", i) in blocks_replace:
                 def block_wrap(args):
                     out = {}

comfy/supported_models.py

@@ -28,6 +28,7 @@
 from . import latent_formats
 
 from . import diffusers_convert
+import comfy.model_management
 
 class SD15(supported_models_base.BASE):
     unet_config = {
@@ -1028,7 +1029,13 @@ class ZImage(Lumina2):
 
     memory_usage_factor = 2.0
 
-    supported_inference_dtypes = [torch.bfloat16, torch.float16, torch.float32]
+    supported_inference_dtypes = [torch.bfloat16, torch.float32]
+
+    def __init__(self, unet_config):
+        super().__init__(unet_config)
+        if comfy.model_management.extended_fp16_support():
+            self.supported_inference_dtypes = self.supported_inference_dtypes.copy()
+            self.supported_inference_dtypes.insert(1, torch.float16)
 
     def clip_target(self, state_dict={}):
         pref = self.text_encoder_key_prefix[0]

comfy/utils.py

@@ -53,7 +53,7 @@ def scalar(*args, **kwargs):
     ALWAYS_SAFE_LOAD = True
     logging.info("Checkpoint files will always be loaded safely.")
 else:
-    logging.info("Warning, you are using an old pytorch version and some ckpt/pt files might be loaded unsafely. Upgrading to 2.4 or above is recommended.")
+    logging.warning("Warning, you are using an old pytorch version and some ckpt/pt files might be loaded unsafely. Upgrading to 2.4 or above is recommended as older versions of pytorch are no longer supported.")
 
 def load_torch_file(ckpt, safe_load=False, device=None, return_metadata=False):
     if device is None:

comfy_api/feature_flags.py

@@ -5,20 +5,20 @@
 allowing graceful protocol evolution while maintaining backward compatibility.
 """
 
-from typing import Any, Dict
+from typing import Any
 
 from comfy.cli_args import args
 
 # Default server capabilities
-SERVER_FEATURE_FLAGS: Dict[str, Any] = {
+SERVER_FEATURE_FLAGS: dict[str, Any] = {
     "supports_preview_metadata": True,
     "max_upload_size": args.max_upload_size * 1024 * 1024, # Convert MB to bytes
     "extension": {"manager": {"supports_v4": True}},
 }
 
 
 def get_connection_feature(
-    sockets_metadata: Dict[str, Dict[str, Any]],
+    sockets_metadata: dict[str, dict[str, Any]],
     sid: str,
     feature_name: str,
     default: Any = False
@@ -42,7 +42,7 @@ def get_connection_feature(
 
 
 def supports_feature(
-    sockets_metadata: Dict[str, Dict[str, Any]],
+    sockets_metadata: dict[str, dict[str, Any]],
     sid: str,
     feature_name: str
 ) -> bool:
@@ -60,7 +60,7 @@ def supports_feature(
     return get_connection_feature(sockets_metadata, sid, feature_name, False) is True
 
 
-def get_server_features() -> Dict[str, Any]:
+def get_server_features() -> dict[str, Any]:
     """
     Get the server's feature flags.
 

comfy_api/internal/api_registry.py

@@ -1,4 +1,4 @@
-from typing import Type, List, NamedTuple
+from typing import NamedTuple
 from comfy_api.internal.singleton import ProxiedSingleton
 from packaging import version as packaging_version
 
@@ -10,7 +10,7 @@ def __init__(self):
 
 class ComfyAPIWithVersion(NamedTuple):
     version: str
-    api_class: Type[ComfyAPIBase]
+    api_class: type[ComfyAPIBase]
 
 
 def parse_version(version_str: str) -> packaging_version.Version:
@@ -23,16 +23,16 @@ def parse_version(version_str: str) -> packaging_version.Version:
     return packaging_version.parse(version_str)
 
 
-registered_versions: List[ComfyAPIWithVersion] = []
+registered_versions: list[ComfyAPIWithVersion] = []
 
 
-def register_versions(versions: List[ComfyAPIWithVersion]):
+def register_versions(versions: list[ComfyAPIWithVersion]):
     versions.sort(key=lambda x: parse_version(x.version))
     global registered_versions
     registered_versions = versions
 
 
-def get_all_versions() -> List[ComfyAPIWithVersion]:
+def get_all_versions() -> list[ComfyAPIWithVersion]:
     """
     Returns a list of all registered ComfyAPI versions.
     """