vargai 0.4.0-alpha104 → 0.4.0-alpha105

package/package.json

@@ -104,7 +104,7 @@
   "license": "Apache-2.0",
   "author": "varg.ai <hello@varg.ai> (https://varg.ai)",
   "sideEffects": false,
-  "version": "0.4.0-alpha104",
+  "version": "0.4.0-alpha105",
   "exports": {
     ".": "./src/index.ts",
     "./ai": "./src/ai-sdk/index.ts",

package/src/ai-sdk/providers/elevenlabs.ts

@@ -89,7 +89,7 @@ class ElevenLabsMusicModel implements MusicModelV3 {
     const elevenLabsOptions = providerOptions?.elevenlabs ?? {};
     const audio = await this.client.music.compose({
       prompt,
-      musicLengthMs: duration ? duration * 1000 : undefined,
+      musicLengthMs: duration ? Math.round(duration * 1000) : undefined,
       modelId: this.modelId,
       ...elevenLabsOptions,
     } as Parameters<typeof this.client.music.compose>[0]);

package/src/ai-sdk/providers/fal.ts

@@ -196,7 +196,7 @@ const IMAGE_MODELS: Record<string, string> = {
   "recraft-v3": "fal-ai/recraft/v3/text-to-image",
   "nano-banana-pro": "fal-ai/nano-banana-pro",
   "nano-banana-pro/edit": "fal-ai/nano-banana-pro/edit",
-  "nano-banana-2": "fal-ai/nano-banana-2/edit",
+  "nano-banana-2": "fal-ai/nano-banana-2",
   "nano-banana-2/edit": "fal-ai/nano-banana-2/edit",
   "seedream-v4.5/edit": "fal-ai/bytedance/seedream/v4.5/edit",
   // Qwen Image 2 - text-to-image and image-to-image editing (standard + pro)
@@ -924,13 +924,21 @@ class FalImageModel implements ImageModelV3 {
     }
 
     const hasFiles = files && files.length > 0;
-    const finalEndpoint = this.resolveEndpoint();
 
     let stableKey: string | undefined;
     if (hasFiles && files) {
       const fileHashes = await computeFileHashes(files);
+      const imageUrls = await pMap(files, fileToUrl, { concurrency: 2 });
+      // Reve uses singular image_url instead of image_urls array
+      if (SINGULAR_IMAGE_URL_MODELS.has(this.modelId)) {
+        input.image_url = imageUrls[0];
+      } else {
+        input.image_urls = imageUrls;
+      }
+      // Compute stable key after files are resolved
+      const finalEndpointForKey = this.resolveEndpoint(hasFiles);
       stableKey = JSON.stringify({
-        endpoint: finalEndpoint,
+        endpoint: finalEndpointForKey,
         prompt,
         n,
         size,
@@ -940,13 +948,6 @@ class FalImageModel implements ImageModelV3 {
         modelId: this.modelId,
         fileHashes,
       });
-      const imageUrls = await pMap(files, fileToUrl, { concurrency: 2 });
-      // Reve uses singular image_url instead of image_urls array
-      if (SINGULAR_IMAGE_URL_MODELS.has(this.modelId)) {
-        input.image_url = imageUrls[0];
-      } else {
-        input.image_urls = imageUrls;
-      }
     }
 
     if (isQwenAngles && !input.image_urls) {
@@ -962,6 +963,10 @@ class FalImageModel implements ImageModelV3 {
       }
     }
 
+    // Resolve endpoint after file processing so dual-endpoint models
+    // (e.g. nano-banana-2 vs nano-banana-2/edit) route correctly
+    const finalEndpoint = this.resolveEndpoint(hasFiles);
+
     const result = await executeWithQueueRecovery<{ data: unknown }>(
       finalEndpoint,
       input,
@@ -998,11 +1003,16 @@ class FalImageModel implements ImageModelV3 {
     };
   }
 
-  private resolveEndpoint(): string {
+  private resolveEndpoint(hasFiles?: boolean): string {
     if (this.modelId.startsWith("raw:")) {
       return this.modelId.slice(4);
     }
 
+    // Nano Banana 2: route to /edit when images are provided, base endpoint for t2i
+    if (this.modelId === "nano-banana-2" && hasFiles) {
+      return "fal-ai/nano-banana-2/edit";
+    }
+
     return IMAGE_MODELS[this.modelId] ?? this.modelId;
   }
 }

package/src/definitions/models/nano-banana-2.ts

@@ -1,6 +1,6 @@
 /**
- * Nano Banana 2 image editing model (Google's next-gen image generation/editing)
- * Edit-only model requiring image_urls input
+ * Nano Banana 2 image model (Google's next-gen image generation/editing)
+ * Supports both text-to-image (no images) and image editing (with image_urls)
  */
 
 import { z } from "zod";
@@ -35,8 +35,9 @@ const nanoBanana2InputSchema = z.object({
   prompt: z.string().describe("Text description for image editing"),
   image_urls: z
     .array(z.string().url())
+    .optional()
     .describe(
-      "Input image URLs for image-to-image editing. Required for this model.",
+      "Input image URLs for image editing. When provided, routes to the /edit endpoint. Omit for text-to-image generation.",
     ),
   resolution: nanoBanana2ResolutionSchema
     .default("1K")
@@ -103,11 +104,11 @@ export const definition: ModelDefinition<typeof schema> = {
   type: "model",
   name: "nano-banana-2",
   description:
-    "Google Nano Banana 2 - next-gen image editing model. Requires image_urls for all operations.",
+    "Google Nano Banana 2 - next-gen image generation and editing model. Supports text-to-image and image editing (with image_urls).",
   providers: ["fal"],
   defaultProvider: "fal",
   providerModels: {
-    fal: "fal-ai/nano-banana-2/edit",
+    fal: "fal-ai/nano-banana-2",
   },
   schema,
 };

package/src/providers/elevenlabs.ts

@@ -117,7 +117,8 @@ export class ElevenLabsProvider extends BaseProvider {
 
     const audio = await this.client.music.compose({
       prompt,
-      musicLengthMs,
+      musicLengthMs:
+        musicLengthMs != null ? Math.round(musicLengthMs) : undefined,
       modelId: "music_v1",
     });
 

package/src/providers/fal.ts

@@ -54,9 +54,13 @@ export class FalProvider extends BaseProvider {
         return "fal-ai/nano-banana-pro/edit";
       }
     }
-    // Nano Banana 2: always route to /edit endpoint (edit-only model)
+    // Nano Banana 2: route to /edit when image_urls are provided, otherwise use base t2i endpoint
     if (model === "fal-ai/nano-banana-2") {
-      return "fal-ai/nano-banana-2/edit";
+      const imageUrls = inputs.image_urls as string[] | undefined;
+      if (imageUrls && imageUrls.length > 0) {
+        return "fal-ai/nano-banana-2/edit";
+      }
+      return "fal-ai/nano-banana-2";
     }
     // Qwen Image 2: route to /edit endpoint when image_urls are provided
     if (model === "fal-ai/qwen-image-2/text-to-image") {

package/src/react/renderers/captions.ts

@@ -156,12 +156,17 @@ function parseSrt(content: string): SrtEntry[] {
   return entries;
 }
 
+/**
+ * Format seconds to ASS timestamp `H:MM:SS.CC`.
+ * Computes from total centiseconds to avoid overflow when rounding
+ * lands on 100 cs (e.g. 1.999s would otherwise produce `0:00:01.100`).
+ */
 function formatAssTime(seconds: number): string {
-  const h = Math.floor(seconds / 3600);
-  const m = Math.floor((seconds % 3600) / 60);
-  const s = Math.floor(seconds % 60);
-  const cs = Math.floor((seconds % 1) * 100);
-
+  const totalCs = Math.max(0, Math.round(seconds * 100));
+  const h = Math.floor(totalCs / 360000);
+  const m = Math.floor((totalCs % 360000) / 6000);
+  const s = Math.floor((totalCs % 6000) / 100);
+  const cs = totalCs % 100;
   return `${h}:${String(m).padStart(2, "0")}:${String(s).padStart(2, "0")}.${String(cs).padStart(2, "0")}`;
 }
 
@@ -190,9 +195,15 @@ Format: Layer, Start, End, Style, Name, MarginL, MarginR, MarginV, Effect, Text
 
   const entries = parseSrt(srtContent);
   const assDialogues = entries
-    .map((entry) => {
+    .map((entry, i) => {
       const start = formatAssTime(entry.start);
-      const end = formatAssTime(entry.end);
+      // Clamp end to next entry's start to prevent overlapping subtitles
+      // (transcription engines often produce overlapping word timestamps)
+      const nextStart =
+        i < entries.length - 1 ? entries[i + 1]!.start : undefined;
+      const clampedEnd =
+        nextStart !== undefined ? Math.min(entry.end, nextStart) : entry.end;
+      const end = formatAssTime(clampedEnd);
       const text = entry.text.replace(/\n/g, "\\N");
       return `Dialogue: 0,${start},${end},Default,,0,0,0,,${text}`;
     })

package/src/react/renderers/image.ts

@@ -37,9 +37,9 @@ async function resolvePrompt(
   if (typeof prompt === "string") {
     return prompt;
   }
-  const resolvedImages = await Promise.all(
-    prompt.images.map((img) => resolveImageInput(img, ctx)),
-  );
+  const resolvedImages = prompt.images
+    ? await Promise.all(prompt.images.map((img) => resolveImageInput(img, ctx)))
+    : [];
   return { text: prompt.text, images: resolvedImages };
 }
 

package/src/react/resolve.ts

@@ -577,9 +577,11 @@ async function resolveImagePrompt(
   prompt: ImagePrompt,
 ): Promise<string | { text?: string; images: Uint8Array[] }> {
   if (typeof prompt === "string") return prompt;
-  const resolvedImages = await Promise.all(
-    prompt.images.map((img) => resolveImageInputForStandalone(img)),
-  );
+  const resolvedImages = prompt.images
+    ? await Promise.all(
+        prompt.images.map((img) => resolveImageInputForStandalone(img)),
+      )
+    : [];
   return { text: prompt.text, images: resolvedImages };
 }
 

package/src/react/types.ts

@@ -129,7 +129,7 @@ export interface OverlayProps extends BaseProps, PositionProps, AudioProps {
 }
 
 export type ImageInput = Uint8Array | string | VargElement<"image">;
-export type ImagePrompt = string | { text?: string; images: ImageInput[] };
+export type ImagePrompt = string | { text?: string; images?: ImageInput[] };
 
 export interface ImageProps extends BaseProps, PositionProps {
   prompt?: ImagePrompt;