Add ByteDance Seedance 4.5 image generation support

CLAUDE.md

@@ -117,11 +117,11 @@ Model name prefixes determine routing:
 - **Anthropic**: claude-sonnet-4-0/4-5/4-6, claude-haiku-4-5, claude-opus-4-5/4-6/4-7/4-8, claude-fable-5, claude-3-7-sonnet, claude-3-5-haiku
 - **Google**: gemini-2.5-flash, gemini-2.5-flash-lite, gemini-2.5-pro, gemini-3-pro-preview, gemini-3-flash-preview, gemini-3.1-pro-preview, gemini-3.1-flash-lite-preview, gemini-3.5-flash; image generation: gemini-2.5-flash-image, gemini-3.1-flash-image
 - **xAI**: grok-2, grok-3, grok-3-mini, grok-4, grok-4-fast, grok-4-1-fast; image generation: grok-2-image
-- **ByteDance** (BytePlus ModelArk, OpenAI-compatible, ap-southeast): seed-1.6, seed-1.8, seed-2.0-lite, deepseek-v4-flash, deepseek-v4-pro; image generation: seedream-4.0
+- **ByteDance** (BytePlus ModelArk, OpenAI-compatible, ap-southeast): seed-1.6, seed-1.8, seed-2.0-lite, deepseek-v4-flash, deepseek-v4-pro; image generation: seedream-4.0, seedance-4.5
 - **Nous Research** (Nous Portal, OpenAI-compatible): hermes-4-405b, hermes-4-70b
 - **Z.ai** (Model API, OpenAI-compatible): glm-5.2; image generation: glm-image
 
-Image generation via xAI (grok-2-image), ByteDance (seedream-4.0), and Z.ai
+Image generation via xAI (grok-2-image), ByteDance (seedream-4.0, seedance-4.5), and Z.ai
 (glm-image) is served through a provider `/images/generations` endpoint rather
 than the chat path, but is surfaced on `/v1/chat/completions` exactly like
 Gemini's inline-image models (images returned out-of-band under the message

tee_gateway/llm_backend.py

@@ -364,16 +364,22 @@ def generate_images(model: str, prompt: str, n: int = 1) -> tuple[list[str], int
         raise RuntimeError(f"{provider} HTTP client has not been initialized")
 
     # n is clamped to the OpenAI-compatible providers' documented 1..10 range.
-    # Z.ai's GLM-Image endpoint currently returns exactly one image and does not
-    # document n/response_format support, so keep its payload to the documented
-    # fields.
+    # Z.ai's GLM-Image and ByteDance Seedance endpoints don't document n/
+    # response_format support, so keep their payloads to documented fields.
     count = max(1, min(int(n), 10))
     payload: dict[str, Any] = {
         "model": cfg.api_name,
         "prompt": prompt,
     }
     if provider == "zai":
         payload["size"] = "1280x1280"
+    elif provider == "bytedance" and cfg.api_name.startswith("ep-"):
+        # Seedance deployment endpoints use URL format and require extra params.
+        payload["response_format"] = "url"
+        payload["sequential_image_generation"] = "disabled"
+        payload["watermark"] = False
+        payload["size"] = "2K"
+        payload["stream"] = False
     else:
         payload["n"] = count
         payload["response_format"] = "b64_json"

tee_gateway/model_registry.py

@@ -376,6 +376,17 @@ class SupportedModel(Enum):
         image_generation=True,
         per_image_price_usd=Decimal("0.03"),
     )
+    # Seedance 4.5 image generation via a ModelArk deployment endpoint.
+    # Uses URL response format and seedance-specific request params
+    # (sequential_image_generation, watermark, size). Billed per image.
+    SEEDANCE_4_5 = ModelConfig(
+        provider="bytedance",
+        api_name="ep-20260624042612-7dxcv",
+        input_price_usd=Decimal("0"),
+        output_price_usd=Decimal("0"),
+        image_generation=True,
+        per_image_price_usd=Decimal("0.05"),
+    )
 
     # ── Nous Research (Nous Portal, OpenAI-compatible) ──────────────────
     # Hermes 4 family, served via Nous's OpenAI-compatible inference API.
@@ -497,6 +508,9 @@ class SupportedModel(Enum):
     "seedream-4-0-250828": SupportedModel.SEEDREAM_4_0,
     "seedream-4.0": SupportedModel.SEEDREAM_4_0,
     "seedream-4-0": SupportedModel.SEEDREAM_4_0,
+    "ep-20260624042612-7dxcv": SupportedModel.SEEDANCE_4_5,
+    "seedance-4.5": SupportedModel.SEEDANCE_4_5,
+    "seedance-4-5": SupportedModel.SEEDANCE_4_5,
     # Nous Research
     "hermes-4-405b": SupportedModel.HERMES_4_405B,
     "hermes-4-70b": SupportedModel.HERMES_4_70B,

tee_gateway/test/test_image_generation.py

@@ -1,5 +1,5 @@
 """Tests for endpoint-based image generation (xAI Grok, ByteDance Seedream,
-Z.ai GLM-Image).
+ByteDance Seedance, Z.ai GLM-Image).
 
 Unlike Gemini's inline-image chat models (see test_image_billing.py), these
 models are served via a dedicated OpenAI-compatible ``/images/generations``
@@ -25,6 +25,7 @@
 
 GROK_IMAGE = "grok-2-image"
 SEEDREAM = "seedream-4.0"
+SEEDANCE = "seedance-4.5"
 GLM_IMAGE = "glm-image"
 
 
@@ -89,6 +90,26 @@ def test_zai_glm_image_uses_documented_payload_and_url_response(self):
         self.assertNotIn("n", payload)
         self.assertNotIn("response_format", payload)
 
+    def test_seedance_uses_url_format_and_extra_params(self):
+        client = MagicMock()
+        client.post.return_value = _mock_response([{"url": "https://cdn/img.jpg"}])
+        with patch.object(llm_backend, "bytedance_http_client", client):
+            images, count = generate_images(SEEDANCE, "a black hole", n=1)
+
+        self.assertEqual(count, 1)
+        self.assertEqual(images, ["https://cdn/img.jpg"])
+
+        _, kwargs = client.post.call_args
+        payload = kwargs["json"]
+        self.assertEqual(payload["model"], get_model_config(SEEDANCE).api_name)
+        self.assertEqual(payload["prompt"], "a black hole")
+        self.assertEqual(payload["response_format"], "url")
+        self.assertEqual(payload["sequential_image_generation"], "disabled")
+        self.assertFalse(payload["watermark"])
+        self.assertEqual(payload["size"], "2K")
+        self.assertFalse(payload["stream"])
+        self.assertNotIn("n", payload)
+
     def test_n_is_clamped_to_provider_range(self):
         client = MagicMock()
         client.post.return_value = _mock_response([{"b64_json": "x"}])
@@ -128,7 +149,7 @@ def _zero_usage() -> dict:
         return {"prompt_tokens": 0, "completion_tokens": 0, "total_tokens": 0}
 
     def test_single_image_charged_flat_price(self):
-        for model in (GROK_IMAGE, SEEDREAM, GLM_IMAGE):
+        for model in (GROK_IMAGE, SEEDREAM, SEEDANCE, GLM_IMAGE):
             with self.subTest(model=model):
                 cfg = get_model_config(model)
                 cost = compute_session_cost(model, self._zero_usage(), image_count=1)