feat(image): add FireRed Image Edit 1.1 support

xinference/model/image/model_spec.json

@@ -1543,6 +1543,81 @@
     "featured": true,
     "updated_at": 1769418379
   },
+  {
+    "version": 2,
+    "model_name": "FireRed-Image-Edit-1.1",
+    "model_family": "stable_diffusion",
+    "model_ability": [
+      "image2image"
+    ],
+    "model_src": {
+      "huggingface": {
+        "model_id": "FireRedTeam/FireRed-Image-Edit-1.1",
+        "model_revision": "main",
+        "gguf_model_id": "vantagewithai/FireRed-Image-Edit-1.1-GGUF",
+        "gguf_quantizations": [
+          "Q3_K_M",
+          "Q3_K_S",
+          "Q4_0",
+          "Q4_1",
+          "Q4_K_M",
+          "Q4_K_S",
+          "Q5_0",
+          "Q5_1",
+          "Q5_K_M",
+          "Q5_K_S",
+          "Q6_K",
+          "Q8_0"
+        ],
+        "gguf_model_file_name_template": "FireRed-Image-Edit-1.1-{quantization}.gguf"
+      },
+      "modelscope": {
+        "model_id": "FireRedTeam/FireRed-Image-Edit-1.1",
+        "model_revision": "master",
+        "gguf_model_id": "vantagewithai/FireRed-Image-Edit-1.1-GGUF",
+        "gguf_quantizations": [
+          "Q3_K_M",
+          "Q3_K_S",
+          "Q4_0",
+          "Q4_1",
+          "Q4_K_M",
+          "Q4_K_S",
+          "Q5_0",
+          "Q5_1",
+          "Q5_K_M",
+          "Q5_K_S",
+          "Q6_K",
+          "Q8_0"
+        ],
+        "gguf_model_file_name_template": "FireRed-Image-Edit-1.1-{quantization}.gguf"
+      }
+    },
+    "default_model_config": {
+      "quantize": true,
+      "quantize_text_encoder": "text_encoder",
+      "torch_dtype": "bfloat16"
+    },
+    "default_generate_config": {
+      "negative_prompt": " ",
+      "num_inference_steps": 40,
+      "true_cfg_scale": 4
+    },
+    "virtualenv": {
+      "packages": [
+        "diffusers>=0.36.0 ; #engine# == \"diffusers\"",
+        "transformers>=4.51.3,<5 ; #engine# == \"diffusers\"",
+        "accelerate>=0.31.0 ; #engine# == \"diffusers\"",
+        "bitsandbytes>=0.48.0 ; #engine# == \"diffusers\" and sys_platform == \"linux\"",
+        "gguf>=0.10.0 ; #engine# == \"diffusers\"",
+        "peft>=0.17.0 ; #engine# == \"diffusers\"",
+        "#system_torch#",
+        "#system_numpy#"
+      ],
+      "no_build_isolation": true
+    },
+    "featured": true,
+    "updated_at": 1786461664
+  },
   {
     "version": 2,
     "model_name": "Qwen-Image-2512",

xinference/model/image/stable_diffusion/core.py

@@ -329,9 +329,9 @@ def load(self):
                     self._model = FluxKontextPipeline.from_pretrained(
                         self._model_path, **self._kwargs
                     )
-                elif "qwen" in model_name_lower:
+                elif "qwen" in model_name_lower or "firered" in model_name_lower:
                     # TODO: remove this branch when auto pipeline supports
-                    # Qwen-Image
+                    # Qwen-Image and models based on its editing pipeline
                     from diffusers import DiffusionPipeline
 
                     self._model = DiffusionPipeline.from_pretrained(
@@ -940,6 +940,12 @@ def image_to_image(
         response_format: str = "url",
         **kwargs,
     ):
+        generate_kwargs = (
+            self._model_spec.default_generate_config or {}  # type: ignore
+        ).copy()
+        generate_kwargs.update({k: v for k, v in kwargs.items() if v is not None})
+        kwargs = generate_kwargs
+
         if self._kwargs.get("controlnet") or self._model_spec.model_ability == [  # type: ignore
             "image2image"
         ]:
@@ -950,6 +956,24 @@ def image_to_image(
                 raise RuntimeError(f"{self._model_uid} does not support image2image")
             model = self._get_model(ability)
 
+        # QwenImageEditPlusPipeline consumes all reference images through its
+        # ``image`` argument.  The OpenAI-compatible endpoint exposes the first
+        # upload as ``image`` and the remaining uploads as ``reference_images``;
+        # combine them here instead of letting the latter be filtered out as an
+        # unsupported pipeline keyword.
+        if (
+            kwargs.get("reference_images")
+            and type(model).__name__ == "QwenImageEditPlusPipeline"
+        ):
+            reference_images = kwargs.pop("reference_images")
+            primary_images = image if isinstance(image, list) else [image]
+            reference_images = (
+                reference_images
+                if isinstance(reference_images, list)
+                else [reference_images]
+            )
+            image = primary_images + reference_images
+
         if padding_image_to_multiple := kwargs.pop("padding_image_to_multiple", None):
             # Model like SD3 image to image requires image's height and width is times of 16
             # padding the image if specified