Merge fb8726244c into cf9cbec596

Co-authored-by: Talmaj Marinc <talmaj@comfy.org>

CODEOWNERS

@@ -1,2 +1,2 @@
 # Admins
-* @comfyanonymous @kosinkadink @guill
+* @comfyanonymous @kosinkadink @guill @alexisrolland @rattus128

comfy/latent_formats.py

@@ -224,6 +224,7 @@ class Flux2(LatentFormat):
 
         self.latent_rgb_factors_bias = [-0.0329, -0.0718, -0.0851]
         self.latent_rgb_factors_reshape = lambda t: t.reshape(t.shape[0], 32, 2, 2, t.shape[-2], t.shape[-1]).permute(0, 1, 4, 2, 5, 3).reshape(t.shape[0], 32, t.shape[-2] * 2, t.shape[-1] * 2)
+        self.taesd_decoder_name = "taef2_decoder"
 
     def process_in(self, latent):
         return latent

comfy/lora.py

@@ -342,6 +342,12 @@ def model_lora_keys_unet(model, key_map={}):
                 key_map["base_model.model.{}".format(key_lora)] = k  # Official base model loras
                 key_map["lycoris_{}".format(key_lora.replace(".", "_"))] = k  # LyCORIS/LoKR format
 
+    if isinstance(model, comfy.model_base.ErnieImage):
+        for k in sdk:
+            if k.startswith("diffusion_model.") and k.endswith(".weight"):
+                key_lora = k[len("diffusion_model."):-len(".weight")]
+                key_map["transformer.{}".format(key_lora)] = k
+
     return key_map
 
 

comfy/sd.py

@@ -479,7 +479,10 @@ class VAE:
                                                             encoder_config={'target': "comfy.ldm.modules.diffusionmodules.model.Encoder", 'params': encoder_config},
                                                             decoder_config={'target': "comfy.ldm.modules.temporal_ae.VideoDecoder", 'params': decoder_config})
             elif "taesd_decoder.1.weight" in sd:
-                self.latent_channels = sd["taesd_decoder.1.weight"].shape[1]
+                if isinstance(metadata, dict) and "tae_latent_channels" in metadata:
+                    self.latent_channels = metadata["tae_latent_channels"]
+                else:
+                    self.latent_channels = sd["taesd_decoder.1.weight"].shape[1]
                 self.first_stage_model = comfy.taesd.taesd.TAESD(latent_channels=self.latent_channels)
             elif "vquantizer.codebook.weight" in sd: #VQGan: stage a of stable cascade
                 self.first_stage_model = StageA()

comfy/supported_models.py

@@ -1879,6 +1879,86 @@ class CogVideoX_I2V(CogVideoX_T2V):
         out = model_base.CogVideoX(self, image_to_video=True, device=device)
         return out
 
-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, ZImagePixelSpace, 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, RT_DETR_v4, ErnieImage, SAM3, SAM31, CogVideoX_I2V, CogVideoX_T2V]
 
-models += [SVD_img2vid]
+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,
+    ZImagePixelSpace,
+    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,
+    RT_DETR_v4,
+    ErnieImage,
+    SAM3,
+    SAM31,
+    CogVideoX_I2V,
+    CogVideoX_T2V,
+    SVD_img2vid,
+]

comfy/taesd/taesd.py

@@ -17,32 +17,79 @@ class Clamp(nn.Module):
         return torch.tanh(x / 3) * 3
 
 class Block(nn.Module):
-    def __init__(self, n_in, n_out):
+    def __init__(self, n_in: int, n_out: int, use_midblock_gn: bool = False):
         super().__init__()
         self.conv = nn.Sequential(conv(n_in, n_out), nn.ReLU(), conv(n_out, n_out), nn.ReLU(), conv(n_out, n_out))
         self.skip = comfy.ops.disable_weight_init.Conv2d(n_in, n_out, 1, bias=False) if n_in != n_out else nn.Identity()
         self.fuse = nn.ReLU()
-    def forward(self, x):
+        if not use_midblock_gn:
+            self.pool = None
+            return
+        n_gn = n_in * 4
+        self.pool = nn.Sequential(
+            comfy.ops.disable_weight_init.Conv2d(n_in, n_gn, 1, bias=False),
+            comfy.ops.disable_weight_init.GroupNorm(4, n_gn),
+            nn.ReLU(inplace=True),
+            comfy.ops.disable_weight_init.Conv2d(n_gn, n_in, 1, bias=False),
+        )
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        if self.pool is not None:
+            x = x + self.pool(x)
         return self.fuse(self.conv(x) + self.skip(x))
 
-def Encoder(latent_channels=4):
-    return nn.Sequential(
-        conv(3, 64), Block(64, 64),
-        conv(64, 64, stride=2, bias=False), Block(64, 64), Block(64, 64), Block(64, 64),
-        conv(64, 64, stride=2, bias=False), Block(64, 64), Block(64, 64), Block(64, 64),
-        conv(64, 64, stride=2, bias=False), Block(64, 64), Block(64, 64), Block(64, 64),
-        conv(64, latent_channels),
-    )
+class Encoder(nn.Sequential):
+    def __init__(self, latent_channels: int = 4, use_gn: bool = False):
+        super().__init__(
+            conv(3, 64), Block(64, 64),
+            conv(64, 64, stride=2, bias=False), Block(64, 64), Block(64, 64), Block(64, 64),
+            conv(64, 64, stride=2, bias=False), Block(64, 64), Block(64, 64), Block(64, 64),
+            conv(64, 64, stride=2, bias=False), Block(64, 64, use_gn), Block(64, 64, use_gn), Block(64, 64, use_gn),
+            conv(64, latent_channels),
+        )
 
+class Decoder(nn.Sequential):
+    def __init__(self, latent_channels: int = 4, use_gn: bool = False):
+        super().__init__(
+            Clamp(), conv(latent_channels, 64), nn.ReLU(),
+            Block(64, 64, use_gn), Block(64, 64, use_gn), Block(64, 64, use_gn), nn.Upsample(scale_factor=2), conv(64, 64, bias=False),
+            Block(64, 64), Block(64, 64), Block(64, 64), nn.Upsample(scale_factor=2), conv(64, 64, bias=False),
+            Block(64, 64), Block(64, 64), Block(64, 64), nn.Upsample(scale_factor=2), conv(64, 64, bias=False),
+            Block(64, 64), conv(64, 3),
+        )
+
+class DecoderFlux2(Decoder):
+    def __init__(self, latent_channels: int = 128, use_gn: bool = True):
+        if latent_channels != 128 or not use_gn:
+            raise ValueError("Unexpected parameters for Flux2 TAE module")
+        super().__init__(latent_channels=32, use_gn=True)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        B, C, H, W = x.shape
+        x = (
+            x
+            .reshape(B, 32, 2, 2, H, W)
+            .permute(0, 1, 4, 2, 5, 3)
+            .reshape(B, 32, H * 2, W * 2)
+        )
+        return super().forward(x)
+
+class EncoderFlux2(Encoder):
+    def __init__(self, latent_channels: int = 128, use_gn: bool = True):
+        if latent_channels != 128 or not use_gn:
+            raise ValueError("Unexpected parameters for Flux2 TAE module")
+        super().__init__(latent_channels=32, use_gn=True)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        result = super().forward(x)
+        B, C, H, W = result.shape
+        return (
+            result
+            .reshape(B, C, H // 2, 2, W // 2, 2)
+            .permute(0, 1, 3, 5, 2, 4)
+            .reshape(B, 128, H // 2, W // 2)
+        )
 
-def Decoder(latent_channels=4):
-    return nn.Sequential(
-        Clamp(), conv(latent_channels, 64), nn.ReLU(),
-        Block(64, 64), Block(64, 64), Block(64, 64), nn.Upsample(scale_factor=2), conv(64, 64, bias=False),
-        Block(64, 64), Block(64, 64), Block(64, 64), nn.Upsample(scale_factor=2), conv(64, 64, bias=False),
-        Block(64, 64), Block(64, 64), Block(64, 64), nn.Upsample(scale_factor=2), conv(64, 64, bias=False),
-        Block(64, 64), conv(64, 3),
-    )
 
 class TAESD(nn.Module):
     latent_magnitude = 3
@@ -51,8 +98,15 @@ class TAESD(nn.Module):
     def __init__(self, encoder_path=None, decoder_path=None, latent_channels=4):
         """Initialize pretrained TAESD on the given device from the given checkpoints."""
         super().__init__()
-        self.taesd_encoder = Encoder(latent_channels=latent_channels)
-        self.taesd_decoder = Decoder(latent_channels=latent_channels)
+        if latent_channels == 128:
+            encoder_class = EncoderFlux2
+            decoder_class = DecoderFlux2
+        else:
+            encoder_class = Encoder
+            decoder_class = Decoder
+        self.taesd_encoder = encoder_class(latent_channels=latent_channels)
+        self.taesd_decoder = decoder_class(latent_channels=latent_channels)
+
         self.vae_scale = torch.nn.Parameter(torch.tensor(1.0))
         self.vae_shift = torch.nn.Parameter(torch.tensor(0.0))
         if encoder_path is not None:
@@ -61,19 +115,19 @@ class TAESD(nn.Module):
             self.taesd_decoder.load_state_dict(comfy.utils.load_torch_file(decoder_path, safe_load=True))
 
     @staticmethod
-    def scale_latents(x):
+    def scale_latents(x: torch.Tensor) -> torch.Tensor:
         """raw latents -> [0, 1]"""
         return x.div(2 * TAESD.latent_magnitude).add(TAESD.latent_shift).clamp(0, 1)
 
     @staticmethod
-    def unscale_latents(x):
+    def unscale_latents(x: torch.Tensor) -> torch.Tensor:
         """[0, 1] -> raw latents"""
         return x.sub(TAESD.latent_shift).mul(2 * TAESD.latent_magnitude)
 
-    def decode(self, x):
+    def decode(self, x: torch.Tensor) -> torch.Tensor:
         x_sample = self.taesd_decoder((x - self.vae_shift) * self.vae_scale)
         x_sample = x_sample.sub(0.5).mul(2)
         return x_sample
 
-    def encode(self, x):
+    def encode(self, x: torch.Tensor) -> torch.Tensor:
         return (self.taesd_encoder(x * 0.5 + 0.5) / self.vae_scale) + self.vae_shift

comfy_api/latest/_input_impl/video_types.py

@@ -251,6 +251,7 @@ class VideoFromFile(VideoInput):
             container.seek(start_pts, stream=video_stream)
 
         image_format = 'gbrpf32le'
+        process_image_format = lambda a: a
         audio = None
 
         streams = [video_stream]
@@ -283,11 +284,25 @@ class VideoFromFile(VideoInput):
                             break
 
                         if not checked_alpha:
+                            alpha_channel = False
                             for comp in frame.format.components:
                                 if comp.is_alpha or frame.format.name == "pal8":
                                     alphas = []
-                                    image_format = 'gbrapf32le'
+                                    alpha_channel = True
                                     break
+                            if frame.format.name in ("yuvj420p", "yuvj422p", "yuvj444p", "rgb24", "rgba", "pal8"):
+                                process_image_format = lambda a: a.float() / 255.0
+                                if alpha_channel:
+                                    image_format = 'rgba'
+                                else:
+                                    image_format = 'rgb24'
+                            else:
+                                process_image_format = lambda a: a
+                                if alpha_channel:
+                                    image_format = 'gbrapf32le'
+                                else:
+                                    image_format = 'gbrpf32le'
+
                             checked_alpha = True
 
                         img = frame.to_ndarray(format=image_format)  # shape: (H, W, 4)
@@ -323,9 +338,9 @@ class VideoFromFile(VideoInput):
                     else:
                         audio_frames.append(frame.to_ndarray())
 
-        images = torch.stack(frames) if len(frames) > 0 else torch.zeros(0, 0, 0, 3)
+        images = process_image_format(torch.stack(frames)) if len(frames) > 0 else torch.zeros(0, 0, 0, 3)
         if alphas is not None:
-            alphas = torch.stack(alphas) if len(alphas) > 0 else torch.zeros(0, 0, 0, 1)
+            alphas = process_image_format(torch.stack(alphas)) if len(alphas) > 0 else torch.zeros(0, 0, 0, 1)
 
         # Get frame rate
         frame_rate = Fraction(video_stream.average_rate) if video_stream.average_rate else Fraction(1)

comfy_api_nodes/apis/bytedance.py

@@ -157,6 +157,11 @@ class SeedanceCreateAssetResponse(BaseModel):
     asset_id: str = Field(...)
 
 
+class SeedanceVirtualLibraryCreateAssetRequest(BaseModel):
+    url: str = Field(..., description="Publicly accessible URL of the image asset to upload.")
+    hash: str = Field(..., description="Dedup key. Re-submitting the same hash returns the existing asset id.")
+
+
 # Dollars per 1K tokens, keyed by (model_id, has_video_input).
 SEEDANCE2_PRICE_PER_1K_TOKENS = {
     ("dreamina-seedance-2-0-260128", False): 0.007,

comfy_api_nodes/nodes_bytedance.py

@@ -1,3 +1,4 @@
+import hashlib
 import logging
 import math
 import re
@@ -20,6 +21,7 @@ from comfy_api_nodes.apis.bytedance import (
     SeedanceCreateAssetResponse,
     SeedanceCreateVisualValidateSessionResponse,
     SeedanceGetVisualValidateSessionResponse,
+    SeedanceVirtualLibraryCreateAssetRequest,
     Seedream4Options,
     Seedream4TaskCreationRequest,
     TaskAudioContent,
@@ -271,6 +273,30 @@ async def _wait_for_asset_active(cls: type[IO.ComfyNode], asset_id: str, group_i
     )
 
 
+async def _seedance_virtual_library_upload_image_asset(
+    cls: type[IO.ComfyNode],
+    image: torch.Tensor,
+    *,
+    wait_label: str = "Uploading image",
+) -> str:
+    """Upload an image into the caller's per-customer Seedance virtual library."""
+    public_url = await upload_image_to_comfyapi(cls, image, wait_label=wait_label)
+    normalized = image.detach().cpu().contiguous().to(torch.float32)
+    digest = hashlib.sha256()
+    digest.update(str(tuple(normalized.shape)).encode("utf-8"))
+    digest.update(b"\0")
+    digest.update(normalized.numpy().tobytes())
+    image_hash = digest.hexdigest()
+    create_resp = await sync_op(
+        cls,
+        ApiEndpoint(path="/proxy/seedance/virtual-library/assets", method="POST"),
+        response_model=SeedanceCreateAssetResponse,
+        data=SeedanceVirtualLibraryCreateAssetRequest(url=public_url, hash=image_hash),
+    )
+    await _wait_for_asset_active(cls, create_resp.asset_id, group_id="virtual-library")
+    return f"asset://{create_resp.asset_id}"
+
+
 def _seedance2_price_extractor(model_id: str, has_video_input: bool):
     """Returns a price_extractor closure for Seedance 2.0 poll_op."""
     rate = SEEDANCE2_PRICE_PER_1K_TOKENS.get((model_id, has_video_input))
@@ -1507,7 +1533,9 @@ class ByteDance2FirstLastFrameNode(IO.ComfyNode):
         if first_frame_asset_id:
             first_frame_url = image_assets[first_frame_asset_id]
         else:
-            first_frame_url = await upload_image_to_comfyapi(cls, first_frame, wait_label="Uploading first frame.")
+            first_frame_url = await _seedance_virtual_library_upload_image_asset(
+                cls, first_frame, wait_label="Uploading first frame."
+            )
 
         content: list[TaskTextContent | TaskImageContent] = [
             TaskTextContent(text=model["prompt"]),
@@ -1527,7 +1555,9 @@ class ByteDance2FirstLastFrameNode(IO.ComfyNode):
             content.append(
                 TaskImageContent(
                     image_url=TaskImageContentUrl(
-                        url=await upload_image_to_comfyapi(cls, last_frame, wait_label="Uploading last frame.")
+                        url=await _seedance_virtual_library_upload_image_asset(
+                            cls, last_frame, wait_label="Uploading last frame."
+                        )
                     ),
                     role="last_frame",
                 ),
@@ -1805,9 +1835,9 @@ class ByteDance2ReferenceNode(IO.ComfyNode):
             content.append(
                 TaskImageContent(
                     image_url=TaskImageContentUrl(
-                        url=await upload_image_to_comfyapi(
+                        url=await _seedance_virtual_library_upload_image_asset(
                             cls,
-                            image=reference_images[key],
+                            reference_images[key],
                             wait_label=f"Uploading image {i}",
                         ),
                     ),

comfy_api_nodes/nodes_openai.py

@@ -415,8 +415,9 @@ class OpenAIGPTImage1(IO.ComfyNode):
                         "1152x2048",
                         "3840x2160",
                         "2160x3840",
+                        "Custom",
                     ],
-                    tooltip="Image size",
+                    tooltip="Image size. Select 'Custom' to use the custom width and height (GPT Image 2 only).",
                     optional=True,
                 ),
                 IO.Int.Input(
@@ -445,6 +446,24 @@ class OpenAIGPTImage1(IO.ComfyNode):
                     default="gpt-image-2",
                     optional=True,
                 ),
+                IO.Int.Input(
+                    "custom_width",
+                    default=1024,
+                    min=1024,
+                    max=3840,
+                    step=16,
+                    tooltip="Used only when `size` is 'Custom'. Must be a multiple of 16 (GPT Image 2 only).",
+                    optional=True,
+                ),
+                IO.Int.Input(
+                    "custom_height",
+                    default=1024,
+                    min=1024,
+                    max=3840,
+                    step=16,
+                    tooltip="Used only when `size` is 'Custom'. Must be a multiple of 16 (GPT Image 2 only).",
+                    optional=True,
+                ),
             ],
             outputs=[
                 IO.Image.Output(),
@@ -471,9 +490,9 @@ class OpenAIGPTImage1(IO.ComfyNode):
                       "high":   [0.133, 0.22]
                     },
                     "gpt-image-2": {
-                      "low":    [0.0048, 0.012],
-                      "medium": [0.041, 0.112],
-                      "high":   [0.165, 0.43]
+                      "low":    [0.0048, 0.019],
+                      "medium": [0.041, 0.168],
+                      "high":   [0.165, 0.67]
                     }
                   };
                   $range := $lookup($lookup($ranges, widgets.model), widgets.quality);
@@ -503,6 +522,8 @@ class OpenAIGPTImage1(IO.ComfyNode):
         mask: Input.Image | None = None,
         n: int = 1,
         size: str = "1024x1024",
+        custom_width: int = 1024,
+        custom_height: int = 1024,
         model: str = "gpt-image-1",
     ) -> IO.NodeOutput:
         validate_string(prompt, strip_whitespace=False)
@@ -510,7 +531,25 @@ class OpenAIGPTImage1(IO.ComfyNode):
         if mask is not None and image is None:
             raise ValueError("Cannot use a mask without an input image")
 
-        if model in ("gpt-image-1", "gpt-image-1.5"):
+        if size == "Custom":
+            if model != "gpt-image-2":
+                raise ValueError("Custom resolution is only supported by GPT Image 2 model")
+            if custom_width % 16 != 0 or custom_height % 16 != 0:
+                raise ValueError(f"Custom width and height must be multiples of 16, got {custom_width}x{custom_height}")
+            if max(custom_width, custom_height) > 3840:
+                raise ValueError(f"Custom resolution max edge must be <= 3840, got {custom_width}x{custom_height}")
+            ratio = max(custom_width, custom_height) / min(custom_width, custom_height)
+            if ratio > 3:
+                raise ValueError(
+                    f"Custom resolution aspect ratio must not exceed 3:1, got {custom_width}x{custom_height}"
+                )
+            total_pixels = custom_width * custom_height
+            if not 655_360 <= total_pixels <= 8_294_400:
+                raise ValueError(
+                    f"Custom resolution total pixels must be between 655,360 and 8,294,400, got {total_pixels}"
+                )
+            size = f"{custom_width}x{custom_height}"
+        elif model in ("gpt-image-1", "gpt-image-1.5"):
             if size not in ("auto", "1024x1024", "1024x1536", "1536x1024"):
                 raise ValueError(f"Resolution {size} is only supported by GPT Image 2 model")
 

nodes.py

@@ -728,50 +728,26 @@ class LoraLoaderModelOnly(LoraLoader):
 
 class VAELoader:
     video_taes = ["taehv", "lighttaew2_2", "lighttaew2_1", "lighttaehy1_5", "taeltx_2"]
-    image_taes = ["taesd", "taesdxl", "taesd3", "taef1"]
+    image_taes = ["taesd", "taesdxl", "taesd3", "taef1", "taef2"]
+
     @staticmethod
     def vae_list(s):
         vaes = folder_paths.get_filename_list("vae")
         approx_vaes = folder_paths.get_filename_list("vae_approx")
-        sdxl_taesd_enc = False
-        sdxl_taesd_dec = False
-        sd1_taesd_enc = False
-        sd1_taesd_dec = False
-        sd3_taesd_enc = False
-        sd3_taesd_dec = False
-        f1_taesd_enc = False
-        f1_taesd_dec = False
-
+        have_img_encoder, have_img_decoder = set(), set()
         for v in approx_vaes:
-            if v.startswith("taesd_decoder."):
-                sd1_taesd_dec = True
-            elif v.startswith("taesd_encoder."):
-                sd1_taesd_enc = True
-            elif v.startswith("taesdxl_decoder."):
-                sdxl_taesd_dec = True
-            elif v.startswith("taesdxl_encoder."):
-                sdxl_taesd_enc = True
-            elif v.startswith("taesd3_decoder."):
-                sd3_taesd_dec = True
-            elif v.startswith("taesd3_encoder."):
-                sd3_taesd_enc = True
-            elif v.startswith("taef1_encoder."):
-                f1_taesd_dec = True
-            elif v.startswith("taef1_decoder."):
-                f1_taesd_enc = True
-            else:
+            parts = v.split("_", 1)
+            if len(parts) != 2 or parts[0] not in s.image_taes:
                 for tae in s.video_taes:
                     if v.startswith(tae):
                         vaes.append(v)
-
-        if sd1_taesd_dec and sd1_taesd_enc:
-            vaes.append("taesd")
-        if sdxl_taesd_dec and sdxl_taesd_enc:
-            vaes.append("taesdxl")
-        if sd3_taesd_dec and sd3_taesd_enc:
-            vaes.append("taesd3")
-        if f1_taesd_dec and f1_taesd_enc:
-            vaes.append("taef1")
+                        break
+                continue
+            if parts[1].startswith("encoder."):
+                have_img_encoder.add(parts[0])
+            elif parts[1].startswith("decoder."):
+                have_img_decoder.add(parts[0])
+        vaes += [k for k in have_img_decoder if k in have_img_encoder]
         vaes.append("pixel_space")
         return vaes
 
@@ -827,6 +803,11 @@ class VAELoader:
             else:
                 vae_path = folder_paths.get_full_path_or_raise("vae", vae_name)
             sd, metadata = comfy.utils.load_torch_file(vae_path, return_metadata=True)
+        if vae_name == "taef2":
+            if metadata is None:
+                metadata = {"tae_latent_channels": 128}
+            else:
+                metadata["tae_latent_channels"] = 128
         vae = comfy.sd.VAE(sd=sd, metadata=metadata)
         vae.throw_exception_if_invalid()
         return (vae,)

requirements.txt

@@ -1,5 +1,5 @@
 comfyui-frontend-package==1.42.15
-comfyui-workflow-templates==0.9.63
+comfyui-workflow-templates==0.9.65
 comfyui-embedded-docs==0.4.4
 torch
 torchsde
@@ -19,7 +19,7 @@ scipy
 tqdm
 psutil
 alembic
-SQLAlchemy>=2.0
+SQLAlchemy>=2.0.0
 filelock
 av>=14.2.0
 comfy-kitchen>=0.2.8