.

backend/api/gpt_image_frames.py

@@ -5,6 +5,7 @@ Feeds Veo a keyframe that matches 3–4 product reference photos.
 
 from __future__ import annotations
 
+import asyncio
 import io
 import os
 from typing import Any, List
@@ -201,29 +202,19 @@ async def generate_segment_first_frame(body: SegmentFirstFrameRequest):
     size = _aspect_to_size(body.aspect_ratio)
 
     file_tuples: list[tuple[str, io.BytesIO, str]] = []
-    urls_in_order = body.reference_image_urls[:4]
-    async with httpx.AsyncClient() as dl:
-        downloaded: list[tuple[bytes, str]] = []
-        for idx, url in enumerate(urls_in_order):
-            try:
-                downloaded.append(await _download_image(dl, url))
-            except httpx.HTTPError as e:
-                raise HTTPException(
-                    status_code=502,
-                    detail=(
-                        f"Could not fetch reference image {idx + 1} of {len(urls_in_order)} ({url[:160]}…): HTTP error: {e}"
-                    ),
-                ) from None
-            except ValueError as e:
-                raise HTTPException(
-                    status_code=400,
-                    detail=(
-                        f"Invalid reference image {idx + 1} of {len(urls_in_order)} ({url[:160]}…): {e}"
-                    ),
-                ) from None
-    for i, (raw, ctype) in enumerate(downloaded):
-        ext = "png" if "png" in ctype else "jpeg"
-        file_tuples.append((f"ref_{i}.{ext}", io.BytesIO(raw), ctype))
+    try:
+        urls_in_order = body.reference_image_urls[:4]
+        async with httpx.AsyncClient() as dl:
+            downloaded = await asyncio.gather(
+                *(_download_image(dl, url) for url in urls_in_order)
+            )
+        for i, (raw, ctype) in enumerate(downloaded):
+            ext = "png" if "png" in ctype else "jpeg"
+            file_tuples.append((f"ref_{i}.{ext}", io.BytesIO(raw), ctype))
+    except httpx.HTTPError as e:
+        raise HTTPException(status_code=502, detail=f"Could not download reference image: {e}")
+    except ValueError as e:
+        raise HTTPException(status_code=400, detail=str(e))
 
     client = AsyncOpenAI(api_key=api_key)
     result = None

backend/api/seedance_generation.py

@@ -129,8 +129,10 @@ def _replicate_input_from_seedance_urls(
     }
     if n == 1:
         inp["image"] = mapped[0]
+    elif n == 2:
+        inp["image"] = mapped[0]
+        inp["last_frame_image"] = mapped[1]
     else:
-        # Multiple URLs = product reference angles, not first+last-frame (see _seedance_input_from_urls).
         inp["reference_images"] = mapped[:9]
     return inp
 
@@ -253,11 +255,9 @@ def _seedance_input_from_urls(
     """
     Seedance modes are mutually exclusive (OpenAPI):
     - 1 URL → image-to-video (first frame only): first_frame_url
-    - 2+ URLs → multimodal reference-to-video: reference_image_urls (max 9)
-
-    We intentionally do not map 2 URLs to first_frame + last_frame: this app passes
-    multiple *product reference* angles (hero + gallery). That old 2-URL branch made
-    KIE treat the second image as an end frame and fail fetches (content[1].image_url).
+    - 2 URLs → first & last frames: first_frame_url + last_frame_url
+    - 3+ URLs → multimodal reference-to-video: reference_image_urls only (max 9)
+    Do not mix first/last/reference in one request.
     """
     n = len(urls)
     base = {
@@ -270,6 +270,9 @@ def _seedance_input_from_urls(
     }
     if n == 1:
         base["first_frame_url"] = urls[0]
+    elif n == 2:
+        base["first_frame_url"] = urls[0]
+        base["last_frame_url"] = urls[1]
     else:
         base["reference_image_urls"] = urls[:9]
     return base
@@ -357,7 +360,7 @@ def _extract_seedance_callback_state(payload: dict) -> dict:
 
 @router.post("/seedance/create", response_model=SeedanceCreateResponse)
 async def seedance_create(body: SeedanceCreateBody):
-    """Start a Seedance 2 video task. One URL = first-frame I2V; multiple = reference-to-video."""
+    """Start a Seedance 2 video task. Maps 1–2 image URLs to first/last frame I2V; 3+ to reference_image_urls."""
     dur = _clamp_seedance_duration(body.duration)
 
     allowed_ratio = {"16:9", "4:3", "1:1", "3:4", "9:16", "21:9", "adaptive"}