feat(api-nodes): allow to use "IMAGE+TEXT" in NanoBanana2 (#12729)

Comfy Aimdo 0.2.4 fixes a VRAM buffer alignment issue that happens in
someworkflows where action is able to bypass the pytorch allocator
and go straight to the cuda hook.

app/frontend_management.py

@@ -17,7 +17,7 @@ from importlib.metadata import version
 import requests
 from typing_extensions import NotRequired
 
-from utils.install_util import get_missing_requirements_message, requirements_path
+from utils.install_util import get_missing_requirements_message, get_required_packages_versions
 
 from comfy.cli_args import DEFAULT_VERSION_STRING
 import app.logger
@@ -45,25 +45,7 @@ def get_installed_frontend_version():
 
 
 def get_required_frontend_version():
-    """Get the required frontend version from requirements.txt."""
-    try:
-        with open(requirements_path, "r", encoding="utf-8") as f:
-            for line in f:
-                line = line.strip()
-                if line.startswith("comfyui-frontend-package=="):
-                    version_str = line.split("==")[-1]
-                    if not is_valid_version(version_str):
-                        logging.error(f"Invalid version format in requirements.txt: {version_str}")
-                        return None
-                    return version_str
-            logging.error("comfyui-frontend-package not found in requirements.txt")
-            return None
-    except FileNotFoundError:
-        logging.error("requirements.txt not found. Cannot determine required frontend version.")
-        return None
-    except Exception as e:
-        logging.error(f"Error reading requirements.txt: {e}")
-        return None
+    return get_required_packages_versions().get("comfyui-frontend-package", None)
 
 
 def check_frontend_version():
@@ -217,25 +199,7 @@ class FrontendManager:
 
     @classmethod
     def get_required_templates_version(cls) -> str:
-        """Get the required workflow templates version from requirements.txt."""
-        try:
-            with open(requirements_path, "r", encoding="utf-8") as f:
-                for line in f:
-                    line = line.strip()
-                    if line.startswith("comfyui-workflow-templates=="):
-                        version_str = line.split("==")[-1]
-                        if not is_valid_version(version_str):
-                            logging.error(f"Invalid templates version format in requirements.txt: {version_str}")
-                            return None
-                        return version_str
-                logging.error("comfyui-workflow-templates not found in requirements.txt")
-                return None
-        except FileNotFoundError:
-            logging.error("requirements.txt not found. Cannot determine required templates version.")
-            return None
-        except Exception as e:
-            logging.error(f"Error reading requirements.txt: {e}")
-            return None
+        return get_required_packages_versions().get("comfyui-workflow-templates", None)
 
     @classmethod
     def default_frontend_path(cls) -> str:

comfy/cli_args.py

@@ -146,6 +146,7 @@ parser.add_argument("--reserve-vram", type=float, default=None, help="Set the am
 
 parser.add_argument("--async-offload", nargs='?', const=2, type=int, default=None, metavar="NUM_STREAMS", help="Use async weight offloading. An optional argument controls the amount of offload streams. Default is 2. Enabled by default on Nvidia.")
 parser.add_argument("--disable-async-offload", action="store_true", help="Disable async weight offloading.")
+parser.add_argument("--disable-dynamic-vram", action="store_true", help="Disable dynamic VRAM and use estimate based model loading.")
 
 parser.add_argument("--force-non-blocking", action="store_true", help="Force ComfyUI to use non-blocking operations for all applicable tensors. This may improve performance on some non-Nvidia systems but can cause issues with some workflows.")
 
@@ -159,7 +160,6 @@ class PerformanceFeature(enum.Enum):
     Fp8MatrixMultiplication = "fp8_matrix_mult"
     CublasOps = "cublas_ops"
     AutoTune = "autotune"
-    DynamicVRAM = "dynamic_vram"
 
 parser.add_argument("--fast", nargs="*", type=PerformanceFeature, help="Enable some untested and potentially quality deteriorating optimizations. This is used to test new features so using it might crash your comfyui. --fast with no arguments enables everything. You can pass a list specific optimizations if you only want to enable specific ones. Current valid optimizations: {}".format(" ".join(map(lambda c: c.value, PerformanceFeature))))
 
@@ -260,4 +260,4 @@ else:
     args.fast = set(args.fast)
 
 def enables_dynamic_vram():
-    return PerformanceFeature.DynamicVRAM in args.fast and not args.highvram and not args.gpu_only
+    return not args.disable_dynamic_vram and not args.highvram and not args.gpu_only and not args.novram and not args.cpu

comfy/context_windows.py

@@ -214,7 +214,7 @@ class IndexListContextHandler(ContextHandlerABC):
         mask = torch.isclose(model_options["transformer_options"]["sample_sigmas"], timestep[0], rtol=0.0001)
         matches = torch.nonzero(mask)
         if torch.numel(matches) == 0:
-            raise Exception("No sample_sigmas matched current timestep; something went wrong.")
+            return  # substep from multi-step sampler: keep self._step from the last full step
         self._step = int(matches[0].item())
 
     def get_context_windows(self, model: BaseModel, x_in: torch.Tensor, model_options: dict[str]) -> list[IndexListContextWindow]:

comfy/ldm/wan/model.py

@@ -1621,3 +1621,118 @@ class HumoWanModel(WanModel):
         # unpatchify
         x = self.unpatchify(x, grid_sizes)
         return x
+
+class SCAILWanModel(WanModel):
+    def __init__(self, model_type="scail", patch_size=(1, 2, 2), in_dim=20, dim=5120, operations=None, device=None, dtype=None, **kwargs):
+        super().__init__(model_type='i2v', patch_size=patch_size, in_dim=in_dim, dim=dim, operations=operations, device=device, dtype=dtype, **kwargs)
+
+        self.patch_embedding_pose = operations.Conv3d(in_dim, dim, kernel_size=patch_size, stride=patch_size, device=device, dtype=torch.float32)
+
+    def forward_orig(self, x, t, context, clip_fea=None, freqs=None, transformer_options={}, pose_latents=None, reference_latent=None, **kwargs):
+
+        if reference_latent is not None:
+            x = torch.cat((reference_latent, x), dim=2)
+
+        # embeddings
+        x = self.patch_embedding(x.float()).to(x.dtype)
+        grid_sizes = x.shape[2:]
+        transformer_options["grid_sizes"] = grid_sizes
+        x = x.flatten(2).transpose(1, 2)
+
+        scail_pose_seq_len = 0
+        if pose_latents is not None:
+            scail_x = self.patch_embedding_pose(pose_latents.float()).to(x.dtype)
+            scail_x = scail_x.flatten(2).transpose(1, 2)
+            scail_pose_seq_len = scail_x.shape[1]
+            x = torch.cat([x, scail_x], dim=1)
+            del scail_x
+
+        # time embeddings
+        e = self.time_embedding(sinusoidal_embedding_1d(self.freq_dim, t.flatten()).to(dtype=x[0].dtype))
+        e = e.reshape(t.shape[0], -1, e.shape[-1])
+        e0 = self.time_projection(e).unflatten(2, (6, self.dim))
+
+        # context
+        context = self.text_embedding(context)
+
+        context_img_len = None
+        if clip_fea is not None:
+            if self.img_emb is not None:
+                context_clip = self.img_emb(clip_fea)  # bs x 257 x dim
+                context = torch.cat([context_clip, context], dim=1)
+            context_img_len = clip_fea.shape[-2]
+
+        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"], context_img_len=context_img_len, transformer_options=args["transformer_options"])
+                    return out
+                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, context_img_len=context_img_len, transformer_options=transformer_options)
+
+        # head
+        x = self.head(x, e)
+
+        if scail_pose_seq_len > 0:
+            x = x[:, :-scail_pose_seq_len]
+
+        # unpatchify
+        x = self.unpatchify(x, grid_sizes)
+
+        if reference_latent is not None:
+            x = x[:, :, reference_latent.shape[2]:]
+
+        return x
+
+    def rope_encode(self, t, h, w, t_start=0, steps_t=None, steps_h=None, steps_w=None, device=None, dtype=None, pose_latents=None, reference_latent=None, transformer_options={}):
+        main_freqs = super().rope_encode(t, h, w, t_start=t_start, steps_t=steps_t, steps_h=steps_h, steps_w=steps_w, device=device, dtype=dtype, transformer_options=transformer_options)
+
+        if pose_latents is None:
+            return main_freqs
+
+        ref_t_patches = 0
+        if reference_latent is not None:
+            ref_t_patches = (reference_latent.shape[2] + (self.patch_size[0] // 2)) // self.patch_size[0]
+
+        F_pose, H_pose, W_pose = pose_latents.shape[-3], pose_latents.shape[-2], pose_latents.shape[-1]
+
+        # if pose is at half resolution, scale_y/scale_x=2 stretches the position range to cover the same RoPE extent as the main frames
+        h_scale = h / H_pose
+        w_scale = w / W_pose
+
+        # 120 w-offset and shift 0.5 to place positions at midpoints (0.5, 2.5, ...) to match the original code
+        h_shift = (h_scale - 1) / 2
+        w_shift = (w_scale - 1) / 2
+        pose_transformer_options = {"rope_options": {"shift_y": h_shift, "shift_x": 120.0 + w_shift, "scale_y": h_scale, "scale_x": w_scale}}
+        pose_freqs = super().rope_encode(F_pose, H_pose, W_pose, t_start=t_start+ref_t_patches, device=device, dtype=dtype, transformer_options=pose_transformer_options)
+
+        return torch.cat([main_freqs, pose_freqs], dim=1)
+
+    def _forward(self, x, timestep, context, clip_fea=None, time_dim_concat=None, transformer_options={}, pose_latents=None, **kwargs):
+        bs, c, t, h, w = x.shape
+        x = comfy.ldm.common_dit.pad_to_patch_size(x, self.patch_size)
+
+        if pose_latents is not None:
+            pose_latents = comfy.ldm.common_dit.pad_to_patch_size(pose_latents, self.patch_size)
+
+        t_len = t
+        if time_dim_concat is not None:
+            time_dim_concat = comfy.ldm.common_dit.pad_to_patch_size(time_dim_concat, self.patch_size)
+            x = torch.cat([x, time_dim_concat], dim=2)
+            t_len = x.shape[2]
+
+        reference_latent = None
+        if "reference_latent" in kwargs:
+            reference_latent = comfy.ldm.common_dit.pad_to_patch_size(kwargs.pop("reference_latent"), self.patch_size)
+            t_len += reference_latent.shape[2]
+
+        freqs = self.rope_encode(t_len, h, w, device=x.device, dtype=x.dtype, transformer_options=transformer_options, pose_latents=pose_latents, reference_latent=reference_latent)
+        return self.forward_orig(x, timestep, context, clip_fea=clip_fea, freqs=freqs, transformer_options=transformer_options, pose_latents=pose_latents, reference_latent=reference_latent, **kwargs)[:, :, :t, :h, :w]

comfy/model_base.py

@@ -1502,6 +1502,44 @@ class WAN21_FlowRVS(WAN21):
         super(WAN21, self).__init__(model_config, model_type, device=device, unet_model=comfy.ldm.wan.model.WanModel)
         self.image_to_video = image_to_video
 
+class WAN21_SCAIL(WAN21):
+    def __init__(self, model_config, model_type=ModelType.FLOW, image_to_video=False, device=None):
+        super(WAN21, self).__init__(model_config, model_type, device=device, unet_model=comfy.ldm.wan.model.SCAILWanModel)
+        self.memory_usage_factor_conds = ("reference_latent", "pose_latents")
+        self.memory_usage_shape_process = {"pose_latents": lambda shape: [shape[0], shape[1], 1.5, shape[-2], shape[-1]]}
+        self.image_to_video = image_to_video
+
+    def extra_conds(self, **kwargs):
+        out = super().extra_conds(**kwargs)
+
+        reference_latents = kwargs.get("reference_latents", None)
+        if reference_latents is not None:
+            ref_latent = self.process_latent_in(reference_latents[-1])
+            ref_mask = torch.ones_like(ref_latent[:, :4])
+            ref_latent = torch.cat([ref_latent, ref_mask], dim=1)
+            out['reference_latent'] = comfy.conds.CONDRegular(ref_latent)
+
+        pose_latents = kwargs.get("pose_video_latent", None)
+        if pose_latents is not None:
+            pose_latents = self.process_latent_in(pose_latents)
+            pose_mask = torch.ones_like(pose_latents[:, :4])
+            pose_latents = torch.cat([pose_latents, pose_mask], dim=1)
+            out['pose_latents'] = comfy.conds.CONDRegular(pose_latents)
+
+        return out
+
+    def extra_conds_shapes(self, **kwargs):
+        out = {}
+        ref_latents = kwargs.get("reference_latents", None)
+        if ref_latents is not None:
+            out['reference_latent'] = list([1, 20, sum(map(lambda a: math.prod(a.size()), ref_latents)) // 16])
+
+        pose_latents = kwargs.get("pose_video_latent", None)
+        if pose_latents is not None:
+            out['pose_latents'] = [pose_latents.shape[0], 20, *pose_latents.shape[2:]]
+
+        return out
+
 class Hunyuan3Dv2(BaseModel):
     def __init__(self, model_config, model_type=ModelType.FLOW, device=None):
         super().__init__(model_config, model_type, device=device, unet_model=comfy.ldm.hunyuan3d.model.Hunyuan3Dv2)

comfy/model_detection.py

@@ -498,6 +498,8 @@ def detect_unet_config(state_dict, key_prefix, metadata=None):
             dit_config["model_type"] = "humo"
         elif '{}face_adapter.fuser_blocks.0.k_norm.weight'.format(key_prefix) in state_dict_keys:
             dit_config["model_type"] = "animate"
+        elif '{}patch_embedding_pose.weight'.format(key_prefix) in state_dict_keys:
+            dit_config["model_type"] = "scail"
         else:
             if '{}img_emb.proj.0.bias'.format(key_prefix) in state_dict_keys:
                 dit_config["model_type"] = "i2v"

comfy/model_management.py

@@ -180,6 +180,14 @@ def is_ixuca():
         return True
     return False
 
+def is_wsl():
+    version = platform.uname().release
+    if version.endswith("-Microsoft"):
+        return True
+    elif version.endswith("microsoft-standard-WSL2"):
+        return True
+    return False
+
 def get_torch_device():
     global directml_enabled
     global cpu_state
@@ -631,12 +639,11 @@ def free_memory(memory_required, device, keep_loaded=[], for_dynamic=False, ram_
         if not DISABLE_SMART_MEMORY:
             memory_to_free = memory_required - get_free_memory(device)
             ram_to_free = ram_required - get_free_ram()
-
-        if current_loaded_models[i].model.is_dynamic() and for_dynamic:
-            #don't actually unload dynamic models for the sake of other dynamic models
-            #as that works on-demand.
-            memory_required -= current_loaded_models[i].model.loaded_size()
-            memory_to_free = 0
+            if current_loaded_models[i].model.is_dynamic() and for_dynamic:
+                #don't actually unload dynamic models for the sake of other dynamic models
+                #as that works on-demand.
+                memory_required -= current_loaded_models[i].model.loaded_size()
+                memory_to_free = 0
         if memory_to_free > 0 and current_loaded_models[i].model_unload(memory_to_free):
             logging.debug(f"Unloading {current_loaded_models[i].model.model.__class__.__name__}")
             unloaded_model.append(i)

comfy/model_patcher.py

@@ -308,15 +308,22 @@ class ModelPatcher:
     def get_free_memory(self, device):
         return comfy.model_management.get_free_memory(device)
 
-    def clone(self, disable_dynamic=False):
+    def get_clone_model_override(self):
+        return self.model, (self.backup, self.object_patches_backup, self.pinned)
+
+    def clone(self, disable_dynamic=False, model_override=None):
         class_ = self.__class__
-        model = self.model
         if self.is_dynamic() and disable_dynamic:
             class_ = ModelPatcher
-            temp_model_patcher = self.cached_patcher_init[0](*self.cached_patcher_init[1], disable_dynamic=True)
-            model = temp_model_patcher.model
+            if model_override is None:
+                if self.cached_patcher_init is None:
+                    raise RuntimeError("Cannot create non-dynamic delegate: cached_patcher_init is not initialized.")
+                temp_model_patcher = self.cached_patcher_init[0](*self.cached_patcher_init[1], disable_dynamic=True)
+                model_override = temp_model_patcher.get_clone_model_override()
+        if model_override is None:
+            model_override = self.get_clone_model_override()
 
-        n = class_(model, self.load_device, self.offload_device, self.model_size(), weight_inplace_update=self.weight_inplace_update)
+        n = class_(model_override[0], self.load_device, self.offload_device, self.model_size(), weight_inplace_update=self.weight_inplace_update)
         n.patches = {}
         for k in self.patches:
             n.patches[k] = self.patches[k][:]
@@ -325,13 +332,12 @@ class ModelPatcher:
         n.object_patches = self.object_patches.copy()
         n.weight_wrapper_patches = self.weight_wrapper_patches.copy()
         n.model_options = comfy.utils.deepcopy_list_dict(self.model_options)
-        n.backup = self.backup
-        n.object_patches_backup = self.object_patches_backup
         n.parent = self
-        n.pinned = self.pinned
 
         n.force_cast_weights = self.force_cast_weights
 
+        n.backup, n.object_patches_backup, n.pinned = model_override[1]
+
         # attachments
         n.attachments = {}
         for k in self.attachments:
@@ -1435,6 +1441,7 @@ class ModelPatcherDynamic(ModelPatcher):
             del self.model.model_loaded_weight_memory
         if not hasattr(self.model, "dynamic_vbars"):
             self.model.dynamic_vbars = {}
+        self.non_dynamic_delegate_model = None
         assert load_device is not None
 
     def is_dynamic(self):
@@ -1669,4 +1676,10 @@ class ModelPatcherDynamic(ModelPatcher):
     def unpatch_hooks(self, whitelist_keys_set: set[str]=None) -> None:
         pass
 
+    def get_non_dynamic_delegate(self):
+        model_patcher = self.clone(disable_dynamic=True, model_override=self.non_dynamic_delegate_model)
+        self.non_dynamic_delegate_model = model_patcher.get_clone_model_override()
+        return model_patcher
+
+
 CoreModelPatcher = ModelPatcher

comfy/sampler_helpers.py

@@ -66,6 +66,18 @@ def convert_cond(cond):
         out.append(temp)
     return out
 
+def cond_has_hooks(cond):
+    for c in cond:
+        temp = c[1]
+        if "hooks" in temp:
+            return True
+        if "control" in temp:
+            control = temp["control"]
+            extra_hooks = control.get_extra_hooks()
+            if len(extra_hooks) > 0:
+                return True
+    return False
+
 def get_additional_models(conds, dtype):
     """loads additional models in conditioning"""
     cnets: list[ControlBase] = []

comfy/samplers.py

@@ -946,6 +946,8 @@ class CFGGuider:
 
     def inner_set_conds(self, conds):
         for k in conds:
+            if self.model_patcher.is_dynamic() and comfy.sampler_helpers.cond_has_hooks(conds[k]):
+                self.model_patcher = self.model_patcher.get_non_dynamic_delegate()
             self.original_conds[k] = comfy.sampler_helpers.convert_cond(conds[k])
 
     def __call__(self, *args, **kwargs):

comfy/sd.py

@@ -204,7 +204,7 @@ def load_bypass_lora_for_models(model, clip, lora, strength_model, strength_clip
 
 
 class CLIP:
-    def __init__(self, target=None, embedding_directory=None, no_init=False, tokenizer_data={}, parameters=0, state_dict=[], model_options={}):
+    def __init__(self, target=None, embedding_directory=None, no_init=False, tokenizer_data={}, parameters=0, state_dict=[], model_options={}, disable_dynamic=False):
         if no_init:
             return
         params = target.params.copy()
@@ -233,7 +233,8 @@ class CLIP:
         model_management.archive_model_dtypes(self.cond_stage_model)
 
         self.tokenizer = tokenizer(embedding_directory=embedding_directory, tokenizer_data=tokenizer_data)
-        self.patcher = comfy.model_patcher.CoreModelPatcher(self.cond_stage_model, load_device=load_device, offload_device=offload_device)
+        ModelPatcher = comfy.model_patcher.ModelPatcher if disable_dynamic else comfy.model_patcher.CoreModelPatcher
+        self.patcher = ModelPatcher(self.cond_stage_model, load_device=load_device, offload_device=offload_device)
         #Match torch.float32 hardcode upcast in TE implemention
         self.patcher.set_model_compute_dtype(torch.float32)
         self.patcher.hook_mode = comfy.hooks.EnumHookMode.MinVram
@@ -267,9 +268,9 @@ class CLIP:
         logging.info("CLIP/text encoder model load device: {}, offload device: {}, current: {}, dtype: {}".format(load_device, offload_device, params['device'], dtype))
         self.tokenizer_options = {}
 
-    def clone(self):
+    def clone(self, disable_dynamic=False):
         n = CLIP(no_init=True)
-        n.patcher = self.patcher.clone()
+        n.patcher = self.patcher.clone(disable_dynamic=disable_dynamic)
         n.cond_stage_model = self.cond_stage_model
         n.tokenizer = self.tokenizer
         n.layer_idx = self.layer_idx
@@ -1164,14 +1165,21 @@ class CLIPType(Enum):
     LONGCAT_IMAGE = 26
 
 
-def load_clip(ckpt_paths, embedding_directory=None, clip_type=CLIPType.STABLE_DIFFUSION, model_options={}):
+
+def load_clip_model_patcher(ckpt_paths, embedding_directory=None, clip_type=CLIPType.STABLE_DIFFUSION, model_options={}, disable_dynamic=False):
+    clip = load_clip(ckpt_paths, embedding_directory, clip_type, model_options, disable_dynamic)
+    return clip.patcher
+
+def load_clip(ckpt_paths, embedding_directory=None, clip_type=CLIPType.STABLE_DIFFUSION, model_options={}, disable_dynamic=False):
     clip_data = []
     for p in ckpt_paths:
         sd, metadata = comfy.utils.load_torch_file(p, safe_load=True, return_metadata=True)
         if model_options.get("custom_operations", None) is None:
             sd, metadata = comfy.utils.convert_old_quants(sd, model_prefix="", metadata=metadata)
         clip_data.append(sd)
-    return load_text_encoder_state_dicts(clip_data, embedding_directory=embedding_directory, clip_type=clip_type, model_options=model_options)
+    clip = load_text_encoder_state_dicts(clip_data, embedding_directory=embedding_directory, clip_type=clip_type, model_options=model_options, disable_dynamic=disable_dynamic)
+    clip.patcher.cached_patcher_init = (load_clip_model_patcher, (ckpt_paths, embedding_directory, clip_type, model_options))
+    return clip
 
 
 class TEModel(Enum):
@@ -1276,7 +1284,7 @@ def llama_detect(clip_data):
 
     return {}
 
-def load_text_encoder_state_dicts(state_dicts=[], embedding_directory=None, clip_type=CLIPType.STABLE_DIFFUSION, model_options={}):
+def load_text_encoder_state_dicts(state_dicts=[], embedding_directory=None, clip_type=CLIPType.STABLE_DIFFUSION, model_options={}, disable_dynamic=False):
     clip_data = state_dicts
 
     class EmptyClass:
@@ -1496,7 +1504,7 @@ def load_text_encoder_state_dicts(state_dicts=[], embedding_directory=None, clip
         parameters += comfy.utils.calculate_parameters(c)
         tokenizer_data, model_options = comfy.text_encoders.long_clipl.model_options_long_clip(c, tokenizer_data, model_options)
 
-    clip = CLIP(clip_target, embedding_directory=embedding_directory, parameters=parameters, tokenizer_data=tokenizer_data, state_dict=clip_data, model_options=model_options)
+    clip = CLIP(clip_target, embedding_directory=embedding_directory, parameters=parameters, tokenizer_data=tokenizer_data, state_dict=clip_data, model_options=model_options, disable_dynamic=disable_dynamic)
     return clip
 
 def load_gligen(ckpt_path):
@@ -1541,8 +1549,10 @@ def load_checkpoint_guess_config(ckpt_path, output_vae=True, output_clip=True, o
     out = load_state_dict_guess_config(sd, output_vae, output_clip, output_clipvision, embedding_directory, output_model, model_options, te_model_options=te_model_options, metadata=metadata, disable_dynamic=disable_dynamic)
     if out is None:
         raise RuntimeError("ERROR: Could not detect model type of: {}\n{}".format(ckpt_path, model_detection_error_hint(ckpt_path, sd)))
-    if output_model:
+    if output_model and out[0] is not None:
         out[0].cached_patcher_init = (load_checkpoint_guess_config_model_only, (ckpt_path, embedding_directory, model_options, te_model_options))
+    if output_clip and out[1] is not None:
+        out[1].patcher.cached_patcher_init = (load_checkpoint_guess_config_clip_only, (ckpt_path, embedding_directory, model_options, te_model_options))
     return out
 
 def load_checkpoint_guess_config_model_only(ckpt_path, embedding_directory=None, model_options={}, te_model_options={}, disable_dynamic=False):
@@ -1553,6 +1563,14 @@ def load_checkpoint_guess_config_model_only(ckpt_path, embedding_directory=None,
             disable_dynamic=disable_dynamic)
     return model
 
+def load_checkpoint_guess_config_clip_only(ckpt_path, embedding_directory=None, model_options={}, te_model_options={}, disable_dynamic=False):
+    _, clip, *_ = load_checkpoint_guess_config(ckpt_path, False, True, False,
+            embedding_directory=embedding_directory, output_model=False,
+            model_options=model_options,
+            te_model_options=te_model_options,
+            disable_dynamic=disable_dynamic)
+    return clip.patcher
+
 def load_state_dict_guess_config(sd, output_vae=True, output_clip=True, output_clipvision=False, embedding_directory=None, output_model=True, model_options={}, te_model_options={}, metadata=None, disable_dynamic=False):
     clip = None
     clipvision = None
@@ -1638,7 +1656,7 @@ def load_state_dict_guess_config(sd, output_vae=True, output_clip=True, output_c
             clip_sd = model_config.process_clip_state_dict(sd)
             if len(clip_sd) > 0:
                 parameters = comfy.utils.calculate_parameters(clip_sd)
-                clip = CLIP(clip_target, embedding_directory=embedding_directory, tokenizer_data=clip_sd, parameters=parameters, state_dict=clip_sd, model_options=te_model_options)
+                clip = CLIP(clip_target, embedding_directory=embedding_directory, tokenizer_data=clip_sd, parameters=parameters, state_dict=clip_sd, model_options=te_model_options, disable_dynamic=disable_dynamic)
             else:
                 logging.warning("no CLIP/text encoder weights in checkpoint, the text encoder model will not be loaded.")
 

comfy/supported_models.py

@@ -1268,6 +1268,16 @@ class WAN21_FlowRVS(WAN21_T2V):
         out = model_base.WAN21_FlowRVS(self, image_to_video=True, device=device)
         return out
 
+class WAN21_SCAIL(WAN21_T2V):
+    unet_config = {
+        "image_model": "wan2.1",
+        "model_type": "scail",
+    }
+
+    def get_model(self, state_dict, prefix="", device=None):
+        out = model_base.WAN21_SCAIL(self, image_to_video=False, device=device)
+        return out
+
 class Hunyuan3Dv2(supported_models_base.BASE):
     unet_config = {
         "image_model": "hunyuan3d2",
@@ -1710,6 +1720,6 @@ class LongCatImage(supported_models_base.BASE):
         hunyuan_detect = comfy.text_encoders.hunyuan_video.llama_detect(state_dict, "{}qwen25_7b.transformer.".format(pref))
         return supported_models_base.ClipTarget(comfy.text_encoders.longcat_image.LongCatImageTokenizer, comfy.text_encoders.longcat_image.te(**hunyuan_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, LongCatImage, 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, WAN21_FlowRVS, Hunyuan3Dv2mini, Hunyuan3Dv2, Hunyuan3Dv2_1, HiDream, Chroma, ChromaRadiance, ACEStep, ACEStep15, Omnigen2, QwenImage, Flux2, Kandinsky5Image, Kandinsky5, Anima]
+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, LongCatImage, 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, WAN21_FlowRVS, WAN21_SCAIL, Hunyuan3Dv2mini, Hunyuan3Dv2, Hunyuan3Dv2_1, HiDream, Chroma, ChromaRadiance, ACEStep, ACEStep15, Omnigen2, QwenImage, Flux2, Kandinsky5Image, Kandinsky5, Anima]
 
 models += [SVD_img2vid]

comfy/text_encoders/ace15.py

@@ -328,14 +328,14 @@ class ACE15TEModel(torch.nn.Module):
                 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"]
+        lm_metadata = token_weight_pairs.get("lm_metadata", {})
         constant = self.constant
         if comfy.model_management.should_use_bf16(device):
             constant *= 0.5
 
         token_weight_pairs = token_weight_pairs.get("lm_prompt", [])
         num_tokens = sum(map(lambda a: len(a), token_weight_pairs))
-        num_tokens += lm_metadata['min_tokens']
+        num_tokens += lm_metadata.get("min_tokens", 0)
         return num_tokens * constant * 1024 * 1024
 
 def te(dtype_llama=None, llama_quantization_metadata=None, lm_model="qwen3_2b"):

comfy_api_nodes/nodes_gemini.py

@@ -789,8 +789,6 @@ class GeminiImage2(IO.ComfyNode):
         validate_string(prompt, strip_whitespace=True, min_length=1)
         if model == "Nano Banana 2 (Gemini 3.1 Flash Image)":
             model = "gemini-3.1-flash-image-preview"
-            if response_modalities == "IMAGE+TEXT":
-                raise ValueError("IMAGE+TEXT is not currently available for the Nano Banana 2 model.")
 
         parts: list[GeminiPart] = [GeminiPart(text=prompt)]
         if images is not None:
@@ -895,7 +893,7 @@ class GeminiNanoBanana2(IO.ComfyNode):
                 ),
                 IO.Combo.Input(
                     "response_modalities",
-                    options=["IMAGE"],
+                    options=["IMAGE", "IMAGE+TEXT"],
                     advanced=True,
                 ),
                 IO.Combo.Input(
@@ -925,6 +923,7 @@ class GeminiNanoBanana2(IO.ComfyNode):
             ],
             outputs=[
                 IO.Image.Output(),
+                IO.String.Output(),
             ],
             hidden=[
                 IO.Hidden.auth_token_comfy_org,

comfy_execution/jobs.py

@@ -20,7 +20,7 @@ class JobStatus:
 
 
 # Media types that can be previewed in the frontend
-PREVIEWABLE_MEDIA_TYPES = frozenset({'images', 'video', 'audio', '3d'})
+PREVIEWABLE_MEDIA_TYPES = frozenset({'images', 'video', 'audio', '3d', 'text'})
 
 # 3D file extensions for preview fallback (no dedicated media_type exists)
 THREE_D_EXTENSIONS = frozenset({'.obj', '.fbx', '.gltf', '.glb', '.usdz'})
@@ -75,6 +75,23 @@ def normalize_outputs(outputs: dict) -> dict:
         normalized[node_id] = normalized_node
     return normalized
 
+# Text preview truncation limit (1024 characters) to prevent preview_output bloat
+TEXT_PREVIEW_MAX_LENGTH = 1024
+
+
+def _create_text_preview(value: str) -> dict:
+    """Create a text preview dict with optional truncation.
+
+    Returns:
+        dict with 'content' and optionally 'truncated' flag
+    """
+    if len(value) <= TEXT_PREVIEW_MAX_LENGTH:
+        return {'content': value}
+    return {
+        'content': value[:TEXT_PREVIEW_MAX_LENGTH],
+        'truncated': True
+    }
+
 
 def _extract_job_metadata(extra_data: dict) -> tuple[Optional[int], Optional[str]]:
     """Extract create_time and workflow_id from extra_data.
@@ -221,23 +238,43 @@ def get_outputs_summary(outputs: dict) -> tuple[int, Optional[dict]]:
                 continue
 
             for item in items:
-                normalized = normalize_output_item(item)
-                if normalized is None:
-                    continue
+                if not isinstance(item, dict):
+                    # Handle text outputs (non-dict items like strings or tuples)
+                    normalized = normalize_output_item(item)
+                    if normalized is None:
+                        # Not a 3D file string — check for text preview
+                        if media_type == 'text':
+                            count += 1
+                            if preview_output is None:
+                                if isinstance(item, tuple):
+                                    text_value = item[0] if item else ''
+                                else:
+                                    text_value = str(item)
+                                text_preview = _create_text_preview(text_value)
+                                enriched = {
+                                    **text_preview,
+                                    'nodeId': node_id,
+                                    'mediaType': media_type
+                                }
+                                if fallback_preview is None:
+                                    fallback_preview = enriched
+                        continue
+                    # normalize_output_item returned a dict (e.g. 3D file)
+                    item = normalized
 
                 count += 1
 
                 if preview_output is not None:
                     continue
 
-                if isinstance(normalized, dict) and is_previewable(media_type, normalized):
+                if is_previewable(media_type, item):
                     enriched = {
-                        **normalized,
+                        **item,
                         'nodeId': node_id,
                     }
-                    if 'mediaType' not in normalized:
+                    if 'mediaType' not in item:
                         enriched['mediaType'] = media_type
-                    if normalized.get('type') == 'output':
+                    if item.get('type') == 'output':
                         preview_output = enriched
                     elif fallback_preview is None:
                         fallback_preview = enriched

comfy_extras/nodes_hooks.py

@@ -248,7 +248,7 @@ class SetClipHooks:
 
     def apply_hooks(self, clip: CLIP, schedule_clip: bool, apply_to_conds: bool, hooks: comfy.hooks.HookGroup=None):
         if hooks is not None:
-            clip = clip.clone()
+            clip = clip.clone(disable_dynamic=True)
             if apply_to_conds:
                 clip.apply_hooks_to_conds = hooks
             clip.patcher.forced_hooks = hooks.clone()

comfy_extras/nodes_mahiro.py

@@ -10,7 +10,7 @@ class Mahiro(io.ComfyNode):
     def define_schema(cls):
         return io.Schema(
             node_id="Mahiro",
-            display_name="Mahiro CFG",
+            display_name="Positive-Biased Guidance",
             category="_for_testing",
             description="Modify the guidance to scale more on the 'direction' of the positive prompt rather than the difference between the negative prompt.",
             inputs=[
@@ -20,27 +20,35 @@ class Mahiro(io.ComfyNode):
                 io.Model.Output(display_name="patched_model"),
             ],
             is_experimental=True,
+            search_aliases=[
+                "mahiro",
+                "mahiro cfg",
+                "similarity-adaptive guidance",
+                "positive-biased cfg",
+            ],
         )
 
     @classmethod
     def execute(cls, model) -> io.NodeOutput:
         m = model.clone()
+
         def mahiro_normd(args):
-            scale: float = args['cond_scale']
-            cond_p: torch.Tensor = args['cond_denoised']
-            uncond_p: torch.Tensor = args['uncond_denoised']
-            #naive leap
+            scale: float = args["cond_scale"]
+            cond_p: torch.Tensor = args["cond_denoised"]
+            uncond_p: torch.Tensor = args["uncond_denoised"]
+            # naive leap
             leap = cond_p * scale
-            #sim with uncond leap
+            # sim with uncond leap
             u_leap = uncond_p * scale
             cfg = args["denoised"]
             merge = (leap + cfg) / 2
             normu = torch.sqrt(u_leap.abs()) * u_leap.sign()
             normm = torch.sqrt(merge.abs()) * merge.sign()
             sim = F.cosine_similarity(normu, normm).mean()
-            simsc = 2 * (sim+1)
-            wm = (simsc*cfg + (4-simsc)*leap) / 4
+            simsc = 2 * (sim + 1)
+            wm = (simsc * cfg + (4 - simsc) * leap) / 4
             return wm
+
         m.set_model_sampler_post_cfg_function(mahiro_normd)
         return io.NodeOutput(m)
 

comfy_extras/nodes_wan.py

@@ -1456,6 +1456,63 @@ class WanInfiniteTalkToVideo(io.ComfyNode):
         return io.NodeOutput(model_patched, positive, negative, out_latent, trim_image)
 
 
+class WanSCAILToVideo(io.ComfyNode):
+    @classmethod
+    def define_schema(cls):
+        return io.Schema(
+            node_id="WanSCAILToVideo",
+            category="conditioning/video_models",
+            inputs=[
+                io.Conditioning.Input("positive"),
+                io.Conditioning.Input("negative"),
+                io.Vae.Input("vae"),
+                io.Int.Input("width", default=512, min=32, max=nodes.MAX_RESOLUTION, step=32),
+                io.Int.Input("height", default=896, min=32, max=nodes.MAX_RESOLUTION, step=32),
+                io.Int.Input("length", default=81, min=1, max=nodes.MAX_RESOLUTION, step=4),
+                io.Int.Input("batch_size", default=1, min=1, max=4096),
+                io.ClipVisionOutput.Input("clip_vision_output", optional=True),
+                io.Image.Input("reference_image", optional=True),
+                io.Image.Input("pose_video", optional=True, tooltip="Video used for pose conditioning. Will be downscaled to half the resolution of the main video."),
+                io.Float.Input("pose_strength", default=1.0, min=0.0, max=10.0, step=0.01, tooltip="Strength of the pose latent."),
+                io.Float.Input("pose_start", default=0.0, min=0.0, max=1.0, step=0.01, tooltip="Start step to use pose conditioning."),
+                io.Float.Input("pose_end", default=1.0, min=0.0, max=1.0, step=0.01, tooltip="End step to use pose conditioning."),
+            ],
+            outputs=[
+                io.Conditioning.Output(display_name="positive"),
+                io.Conditioning.Output(display_name="negative"),
+                io.Latent.Output(display_name="latent", tooltip="Empty latent of the generation size."),
+            ],
+            is_experimental=True,
+        )
+
+    @classmethod
+    def execute(cls, positive, negative, vae, width, height, length, batch_size, pose_strength, pose_start, pose_end, reference_image=None, clip_vision_output=None, pose_video=None) -> io.NodeOutput:
+        latent = torch.zeros([batch_size, 16, ((length - 1) // 4) + 1, height // 8, width // 8], device=comfy.model_management.intermediate_device())
+
+        ref_latent = None
+        if reference_image is not None:
+            reference_image = comfy.utils.common_upscale(reference_image[:1].movedim(-1, 1), width, height, "bilinear", "center").movedim(1, -1)
+            ref_latent = vae.encode(reference_image[:, :, :, :3])
+
+        if ref_latent is not None:
+            positive = node_helpers.conditioning_set_values(positive, {"reference_latents": [ref_latent]}, append=True)
+            negative = node_helpers.conditioning_set_values(negative, {"reference_latents": [torch.zeros_like(ref_latent)]}, append=True)
+
+        if clip_vision_output is not None:
+            positive = node_helpers.conditioning_set_values(positive, {"clip_vision_output": clip_vision_output})
+            negative = node_helpers.conditioning_set_values(negative, {"clip_vision_output": clip_vision_output})
+
+        if pose_video is not None:
+            pose_video = comfy.utils.common_upscale(pose_video[:length].movedim(-1, 1), width // 2, height // 2, "area", "center").movedim(1, -1)
+            pose_video_latent = vae.encode(pose_video[:, :, :, :3]) * pose_strength
+            positive = node_helpers.conditioning_set_values_with_timestep_range(positive, {"pose_video_latent": pose_video_latent}, pose_start, pose_end)
+            negative = node_helpers.conditioning_set_values_with_timestep_range(negative, {"pose_video_latent": pose_video_latent}, pose_start, pose_end)
+
+        out_latent = {}
+        out_latent["samples"] = latent
+        return io.NodeOutput(positive, negative, out_latent)
+
+
 class WanExtension(ComfyExtension):
     @override
     async def get_node_list(self) -> list[type[io.ComfyNode]]:
@@ -1476,6 +1533,7 @@ class WanExtension(ComfyExtension):
             WanAnimateToVideo,
             Wan22ImageToVideoLatent,
             WanInfiniteTalkToVideo,
+            WanSCAILToVideo,
         ]
 
 async def comfy_entrypoint() -> WanExtension:

main.py

@@ -192,7 +192,7 @@ import hook_breaker_ac10a0
 import comfy.memory_management
 import comfy.model_patcher
 
-if enables_dynamic_vram():
+if enables_dynamic_vram() and comfy.model_management.is_nvidia() and not comfy.model_management.is_wsl():
     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")
     elif comfy_aimdo.control.init_device(comfy.model_management.get_torch_device().index):

node_helpers.py

@@ -1,5 +1,6 @@
 import hashlib
 import torch
+import logging
 
 from comfy.cli_args import args
 
@@ -21,6 +22,36 @@ def conditioning_set_values(conditioning, values={}, append=False):
 
     return c
 
+def conditioning_set_values_with_timestep_range(conditioning, values={}, start_percent=0.0, end_percent=1.0):
+    """
+    Apply values to conditioning only during [start_percent, end_percent], keeping the
+    original conditioning active outside that range. Respects existing per-entry ranges.
+    """
+    if start_percent > end_percent:
+        logging.warning(f"start_percent ({start_percent}) must be <= end_percent ({end_percent})")
+        return conditioning
+
+    EPS = 1e-5 # the sampler gates entries with strict > / <, shift boundaries slightly to ensure only one conditioning is active per timestep
+    c = []
+    for t in conditioning:
+        cond_start = t[1].get("start_percent", 0.0)
+        cond_end   = t[1].get("end_percent",   1.0)
+        intersect_start = max(start_percent, cond_start)
+        intersect_end   = min(end_percent,   cond_end)
+
+        if intersect_start >= intersect_end: # no overlap: emit unchanged
+            c.append(t)
+            continue
+
+        if intersect_start > cond_start: # part before the requested range
+            c.extend(conditioning_set_values([t], {"start_percent": cond_start, "end_percent": intersect_start - EPS}))
+
+        c.extend(conditioning_set_values([t], {**values, "start_percent": intersect_start, "end_percent": intersect_end}))
+
+        if intersect_end < cond_end: # part after the requested range
+            c.extend(conditioning_set_values([t], {"start_percent": intersect_end + EPS, "end_percent": cond_end}))
+    return c
+
 def pillow(fn, arg):
     prev_value = None
     try:

requirements.txt

@@ -22,7 +22,7 @@ alembic
 SQLAlchemy
 av>=14.2.0
 comfy-kitchen>=0.2.7
-comfy-aimdo>=0.2.2
+comfy-aimdo>=0.2.4
 requests
 
 #non essential dependencies:

tests-unit/app_test/frontend_manager_test.py

@@ -49,6 +49,12 @@ def mock_provider(mock_releases):
     return provider
 
 
+@pytest.fixture(autouse=True)
+def clear_cache():
+    import utils.install_util
+    utils.install_util.PACKAGE_VERSIONS = {}
+
+
 def test_get_release(mock_provider, mock_releases):
     version = "1.0.0"
     release = mock_provider.get_release(version)

tests/execution/test_jobs.py

@@ -38,13 +38,13 @@ class TestIsPreviewable:
     """Unit tests for is_previewable()"""
 
     def test_previewable_media_types(self):
-        """Images, video, audio, 3d media types should be previewable."""
-        for media_type in ['images', 'video', 'audio', '3d']:
+        """Images, video, audio, 3d, text media types should be previewable."""
+        for media_type in ['images', 'video', 'audio', '3d', 'text']:
             assert is_previewable(media_type, {}) is True
 
     def test_non_previewable_media_types(self):
         """Other media types should not be previewable."""
-        for media_type in ['latents', 'text', 'metadata', 'files']:
+        for media_type in ['latents', 'metadata', 'files']:
             assert is_previewable(media_type, {}) is False
 
     def test_3d_extensions_previewable(self):

utils/install_util.py

@@ -1,5 +1,7 @@
 from pathlib import Path
 import sys
+import logging
+import re
 
 # The path to the requirements.txt file
 requirements_path = Path(__file__).parents[1] / "requirements.txt"
@@ -16,3 +18,34 @@ Please install the updated requirements.txt file by running:
 {sys.executable} {extra}-m pip install -r {requirements_path}
 If you are on the portable package you can run: update\\update_comfyui.bat to solve this problem.
 """.strip()
+
+
+def is_valid_version(version: str) -> bool:
+    """Validate if a string is a valid semantic version (X.Y.Z format)."""
+    pattern = r"^(\d+)\.(\d+)\.(\d+)$"
+    return bool(re.match(pattern, version))
+
+
+PACKAGE_VERSIONS = {}
+def get_required_packages_versions():
+    if len(PACKAGE_VERSIONS) > 0:
+        return PACKAGE_VERSIONS.copy()
+    out = PACKAGE_VERSIONS
+    try:
+        with open(requirements_path, "r", encoding="utf-8") as f:
+            for line in f:
+                line = line.strip().replace(">=", "==")
+                s = line.split("==")
+                if len(s) == 2:
+                    version_str = s[-1]
+                    if not is_valid_version(version_str):
+                        logging.error(f"Invalid version format in requirements.txt: {version_str}")
+                        continue
+                    out[s[0]] = version_str
+        return out.copy()
+    except FileNotFoundError:
+        logging.error("requirements.txt not found.")
+        return None
+    except Exception as e:
+        logging.error(f"Error reading requirements.txt: {e}")
+        return None