hokusai/builder.py

@@ -91,9 +91,9 @@ def _collect_extra_params(target_name: str, config: ProjectConfig) -> dict[str,
     if target_cfg.height is not None:
         params["height"] = target_cfg.height
     if target_cfg.reference_images:
-        params["reference_images"] = list(target_cfg.reference_images)
+        params["reference_images"] = tuple(target_cfg.reference_images)
     if target_cfg.control_images:
-        params["control_images"] = list(target_cfg.control_images)
+        params["control_images"] = tuple(target_cfg.control_images)
     return params
 
 

hokusai/providers/openai_image.py

@@ -3,7 +3,6 @@
 from __future__ import annotations
 
 import base64
-import io
 from pathlib import Path
 from typing import Literal, override
 
@@ -129,6 +128,7 @@ class OpenAIImageProvider(Provider):
                     target_config.reference_images,
                     project_dir,
                     size,
+                    output_format,
                 )
             else:
                 response = await _generate_new(
@@ -183,30 +183,17 @@ async def _generate_edit(
     reference_images: list[str],
     project_dir: Path,
     size: _SIZE | None,
+    output_format: str | None = None,
 ) -> ImagesResponse:
     """Generate an image using reference images via the edits endpoint.
 
     gpt-image-* models accept up to 16 images and return b64 by default
-    (they reject ``response_format`` and ``output_format``).
-    DALL-E 2 accepts only one image.
+    (they reject ``response_format``).  DALL-E 2 accepts only one image.
     """
-    raw_images = [(project_dir / name).read_bytes() for name in reference_images]
+    images = [(project_dir / name).read_bytes() for name in reference_images]
+    image: bytes | list[bytes] = images[0] if len(images) == 1 else images
 
     if model.startswith("gpt-image-"):
-        # gpt-image-* models require file-like objects with a name attribute;
-        # raw bytes trigger the legacy multipart path that only accepts dall-e-2.
-        def _to_named_buf(data: bytes, name: str) -> io.BytesIO:
-            buf = io.BytesIO(data)
-            buf.name = name
-            return buf
-
-        file_images = [
-            _to_named_buf(data, name)
-            for data, name in zip(raw_images, reference_images, strict=True)
-        ]
-        image: io.BytesIO | list[io.BytesIO] = (
-            file_images[0] if len(file_images) == 1 else file_images
-        )
         kwargs: dict[str, object] = {
             "image": image,
             "prompt": prompt,
@@ -215,13 +202,12 @@ async def _generate_edit(
         }
         if size is not None:
             kwargs["size"] = size
+        if output_format is not None:
+            kwargs["output_format"] = output_format
         return await client.images.edit(**kwargs)  # pyright: ignore[reportCallIssue,reportArgumentType,reportUnknownVariableType]
 
-    dalle_image: bytes | list[bytes] = (
-        raw_images[0] if len(raw_images) == 1 else raw_images
-    )
     kwargs = {
-        "image": dalle_image,
+        "image": image,
         "prompt": prompt,
         "model": model,
         "n": 1,

pyproject.toml

@@ -9,7 +9,7 @@ dependencies = [
     "httpx>=0.27.0",
     "mistralai>=1.0.0",
     "networkx>=3.6.1",
-    "openai>=2.25.0",
+    "openai>=2.21.0",
     "pillow>=11.0.0",
     "pydantic>=2.12.5",
     "pyyaml>=6.0",

tests/test_builder.py

@@ -152,7 +152,7 @@ class TestCollectHelpers:
         params = _collect_extra_params("out.png", config)
         assert params["width"] == 512
         assert params["height"] == 256
-        assert params["reference_images"] == ["ref.png"]
+        assert params["reference_images"] == ("ref.png",)
 
     def test_collect_extra_params_empty(self, write_config: WriteConfig) -> None:
         config = write_config({"targets": {"out.txt": {"prompt": "x"}}})

tests/test_providers.py

@@ -7,7 +7,6 @@ Mock-heavy tests produce many Any-typed expressions from MagicMock.
 from __future__ import annotations
 
 import base64
-import io
 from pathlib import Path
 from unittest.mock import AsyncMock, MagicMock, patch
 
@@ -481,10 +480,8 @@ class TestOpenAIImageProvider:
             )
 
             call_args = mock_client.images.edit.call_args
-            # gpt-image-* models pass a BytesIO with a name attribute
-            img_arg = call_args.kwargs["image"]
-            assert img_arg.read() == b"reference data"
-            assert hasattr(img_arg, "name")
+            # Single reference image should be passed as raw bytes
+            assert call_args.kwargs["image"] == b"reference data"
 
         output = project_dir / "out.png"
         assert output.exists()
@@ -517,9 +514,9 @@ class TestOpenAIImageProvider:
             )
 
             call_args = mock_client.images.edit.call_args
-            # gpt-image-* models pass a list of BytesIO with name attributes
-            image_arg: list[io.BytesIO] = call_args.kwargs["image"]
+            # Multiple reference images should be passed as a list of bytes
+            image_arg: list[bytes] = call_args.kwargs["image"]
             assert isinstance(image_arg, list)
             assert len(image_arg) == 2
-            assert image_arg[0].read() == b"ref1 data"
-            assert image_arg[1].read() == b"ref2 data"
+            assert image_arg[0] == b"ref1 data"
+            assert image_arg[1] == b"ref2 data"

uv.lock

@@ -184,7 +184,7 @@ requires-dist = [
     { name = "httpx", specifier = ">=0.27.0" },
     { name = "mistralai", specifier = ">=1.0.0" },
     { name = "networkx", specifier = ">=3.6.1" },
-    { name = "openai", specifier = ">=2.25.0" },
+    { name = "openai", specifier = ">=2.21.0" },
     { name = "pillow", specifier = ">=11.0.0" },
     { name = "pydantic", specifier = ">=2.12.5" },
     { name = "pyyaml", specifier = ">=6.0" },
@@ -386,7 +386,7 @@ wheels = [
 
 [[package]]
 name = "openai"
-version = "2.25.0"
+version = "2.21.0"
 source = { registry = "https://pypi.org/simple" }
 dependencies = [
     { name = "anyio" },
@@ -398,9 +398,9 @@ dependencies = [
     { name = "tqdm" },
     { name = "typing-extensions" },
 ]
-sdist = { url = "https://files.pythonhosted.org/packages/43/2b/a442b206ed74908dd79e2ad1ef3ffaeae66422b1fb506af981f0ef671ba0/openai-2.25.0.tar.gz", hash = "sha256:c3e1965d83c333dbd341eb2c7c8aceb783c272cd57fc57353404d9a443634e29", size = 666577, upload-time = "2026-03-05T18:35:38.292Z" }
+sdist = { url = "https://files.pythonhosted.org/packages/92/e5/3d197a0947a166649f566706d7a4c8f7fe38f1fa7b24c9bcffe4c7591d44/openai-2.21.0.tar.gz", hash = "sha256:81b48ce4b8bbb2cc3af02047ceb19561f7b1dc0d4e52d1de7f02abfd15aa59b7", size = 644374, upload-time = "2026-02-14T00:12:01.577Z" }
 wheels = [
-    { url = "https://files.pythonhosted.org/packages/5b/e6/ca496976bacdd7f86d4fa69359c2d9b5a547d7acbf0ae5cac7bff107ff50/openai-2.25.0-py3-none-any.whl", hash = "sha256:7c7c01a0a4b69771a29913e28d06bc1e9cee8781b4d206bb56cb946d8d2fcb23", size = 1136347, upload-time = "2026-03-05T18:35:36.497Z" },
+    { url = "https://files.pythonhosted.org/packages/cc/56/0a89092a453bb2c676d66abee44f863e742b2110d4dbb1dbcca3f7e5fc33/openai-2.21.0-py3-none-any.whl", hash = "sha256:0bc1c775e5b1536c294eded39ee08f8407656537ccc71b1004104fe1602e267c", size = 1103065, upload-time = "2026-02-14T00:11:59.603Z" },
 ]
 
 [[package]]