@juspay/neurolink 9.86.5 → 9.87.1

package/CHANGELOG.md

@@ -1,3 +1,15 @@
+## [9.87.1](https://github.com/juspay/neurolink/compare/v9.87.0...v9.87.1) (2026-07-14)
+
+### Bug Fixes
+
+- **(utils):** safeFetch pinned-agent lookup honors options.all (autoSelectFamily) ([a18af4a](https://github.com/juspay/neurolink/commit/a18af4a688a90c7f085dd878235b6d7ee4a573e5))
+
+## [9.87.0](https://github.com/juspay/neurolink/compare/v9.86.5...v9.87.0) (2026-07-14)
+
+### Features
+
+- **(video):** imageInputKey option for Replicate image-to-video models ([a66cab9](https://github.com/juspay/neurolink/commit/a66cab95749ad664d5e14824e0ee731ba9c3de8a))
+
 ## [9.86.5](https://github.com/juspay/neurolink/compare/v9.86.4...v9.86.5) (2026-07-14)
 
 ### Bug Fixes

package/dist/adapters/video/replicateVideoHandler.js

@@ -51,27 +51,37 @@ export class ReplicateVideoHandler {
         const startTime = Date.now();
         const model = options.model ?? DEFAULT_MODEL;
         const dataUri = `data:image/${this.detectImageType(image)};base64,${image.toString("base64")}`;
-        // Wan-Alpha + most image-to-video models accept this shape; specific
-        // models may require provider-specific extras passed through
-        // VideoOutputOptions.[unknown key].
+        // Replicate image-to-video models do not share an input schema. The
+        // legacy shape below (`image` + `num_frames`/`fps`/`aspect_ratio`) fits
+        // Wan-Alpha-era models; current models (minimax/hailuo-2.3-*,
+        // wan-video/wan-2.7-i2v) key the image as `first_frame_image` /
+        // `first_frame` and take `duration` instead of frame counts — a missing
+        // required image key fails the prediction on submit. `imageInputKey`
+        // selects the image key AND the modern `duration`/`resolution` shape;
+        // omitting it preserves the legacy payload exactly.
         //
-        // `resolution` is forwarded as the `resolution` input parameter.
-        // Wan-Alpha and several other Replicate image-to-video models accept it
-        // (e.g. "720p", "1080p"). Models that do not recognise it will silently
-        // ignore the field — the Replicate API does not reject unknown input keys.
         // `calculateDimensions` still populates the metadata `dimensions` field
         // so downstream consumers always receive correct width/height regardless
         // of whether the model honoured the resolution hint.
-        const inputPayload = {
-            image: dataUri,
-            prompt,
-            num_frames: (options.length ?? 4) * 24, // Assume 24 fps
-            fps: 24,
-            aspect_ratio: options.aspectRatio,
-            ...(options.resolution !== undefined
-                ? { resolution: options.resolution }
-                : {}),
-        };
+        const inputPayload = options.imageInputKey !== undefined
+            ? {
+                [options.imageInputKey]: dataUri,
+                prompt,
+                duration: options.length ?? 4,
+                ...(options.resolution !== undefined
+                    ? { resolution: options.resolution }
+                    : {}),
+            }
+            : {
+                image: dataUri,
+                prompt,
+                num_frames: (options.length ?? 4) * 24, // Assume 24 fps
+                fps: 24,
+                aspect_ratio: options.aspectRatio,
+                ...(options.resolution !== undefined
+                    ? { resolution: options.resolution }
+                    : {}),
+            };
         let prediction;
         try {
             prediction = await predict(auth, { model, input: inputPayload }, { abortSignal: options.abortSignal });