Add separate VAE loader for SDXL/Illustrious/Pony models

- Use sdxl_vae.safetensors by default instead of checkpoint VAE
- Add vae parameter to generate_image and API endpoint
- Better quality for modern SDXL-based models

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

tensors/comfyui.py

@@ -661,8 +661,8 @@ LORA_LOADER_NODE: dict[str, Any] = {
     },
 }
 
-# Default SDXL/Flux compatible workflow template
+# Default SDXL/Illustrious/Pony compatible workflow template
-# This is a minimal text-to-image workflow that works with most models
+# Uses separate VAE loader for better quality with modern models
 DEFAULT_WORKFLOW_TEMPLATE: dict[str, Any] = {
     "3": {
         "class_type": "KSampler",
@@ -697,14 +697,21 @@ DEFAULT_WORKFLOW_TEMPLATE: dict[str, Any] = {
     },
     "8": {
         "class_type": "VAEDecode",
-        "inputs": {"samples": ["3", 0], "vae": ["4", 2]},
+        "inputs": {"samples": ["3", 0], "vae": ["11", 0]},
     },
     "9": {
         "class_type": "SaveImage",
         "inputs": {"filename_prefix": "comfy", "images": ["8", 0]},
     },
+    "11": {
+        "class_type": "VAELoader",
+        "inputs": {"vae_name": "sdxl_vae.safetensors"},
+    },
 }
 
+# Default VAE for SDXL/Illustrious/Pony models
+DEFAULT_VAE = "sdxl_vae.safetensors"
+
 
 def _build_workflow(
     prompt: str,
@@ -720,6 +727,7 @@ def _build_workflow(
     lora_name: str | None = None,
     lora_strength: float = 1.0,
     batch_size: int = 1,
+    vae: str | None = None,
 ) -> dict[str, Any]:
     """Build a text-to-image workflow from parameters.
 
@@ -737,6 +745,7 @@ def _build_workflow(
         lora_name: LoRA model filename (optional)
         lora_strength: LoRA strength (default 1.0)
         batch_size: Number of images to generate in one workflow (default 1)
+        vae: VAE filename (defaults to sdxl_vae.safetensors)
 
     Returns:
         ComfyUI workflow dict
@@ -766,6 +775,9 @@ def _build_workflow(
     workflow["6"]["inputs"]["text"] = prompt
     workflow["7"]["inputs"]["text"] = negative_prompt
 
+    # Set VAE
+    workflow["11"]["inputs"]["vae_name"] = vae or DEFAULT_VAE
+
     # Inject LoRA loader if specified
     if lora_name:
         # Add LoRA loader node (node 10)
@@ -806,6 +818,7 @@ def generate_image(
     lora_name: str | None = None,
     lora_strength: float = 1.0,
     batch_size: int = 1,
+    vae: str | None = None,
 ) -> GenerationResult | None:
     """Generate an image using a simple text-to-image workflow.
 
@@ -827,6 +840,7 @@ def generate_image(
         lora_name: LoRA model filename (optional)
         lora_strength: LoRA strength (default 1.0)
         batch_size: Number of images to generate in one workflow (default 1)
+        vae: VAE filename (defaults to sdxl_vae.safetensors)
 
     Returns:
         GenerationResult with image paths, or None if generation failed
@@ -860,6 +874,7 @@ def generate_image(
         lora_name=lora_name,
         lora_strength=lora_strength,
         batch_size=batch_size,
+        vae=vae,
     )
 
     # Run workflow

tensors/server/comfyui_api_routes.py

@@ -44,6 +44,7 @@ class GenerateRequest(PydanticBaseModel):
     seed: int = Field(default=-1, description="Random seed (-1 for random)")
     sampler: str = Field(default="euler", description="Sampler name")
     scheduler: str = Field(default="normal", description="Scheduler name")
+    vae: str | None = Field(default=None, description="VAE model name (defaults to sdxl_vae.safetensors)")
 
 
 class GenerateResponse(PydanticBaseModel):
@@ -233,6 +234,7 @@ def comfyui_generate(request: GenerateRequest) -> dict[str, Any]:
         seed=request.seed,
         sampler=request.sampler,
         scheduler=request.scheduler,
+        vae=request.vae,
     )
 
     if not result: