fix(ai-gemini): stop dropping modelOptions on the native image path

.changeset/gemini-native-image-model-options.md

@@ -0,0 +1,11 @@
+---
+'@tanstack/ai-gemini': minor
+---
+
+Type Gemini-native image models with their own provider options. `GeminiImageModelProviderOptionsByName` mapped **every** image model to the Imagen-shaped `GeminiImageProviderOptions`, so `modelOptions: { safetySettings, thinkingConfig, imageConfig, systemInstruction }` was a compile error on `gemini-3.1-flash-image-preview`, `gemini-3.1-flash-lite-image`, `gemini-3-pro-image-preview`, and `gemini-2.5-flash-image` — even though those models are served by `generateContent`, whose `GenerateContentConfig` accepts all of them. The adapter compensated by forwarding only `seed`, silently dropping anything else.
+
+The map now splits native vs Imagen, mirroring the split already used by `GeminiImageModelSizeByName` and `GeminiImageModelInputModalitiesByName`: native models get the new `GeminiNativeImageProviderOptions` (`seed`, `safetySettings`, `thinkingConfig`, `imageConfig`, `systemInstruction`), Imagen models keep `GeminiImageProviderOptions`. Both API paths now pick their config fields by name — never a wholesale spread — so neither shape's fields can reach the other's endpoint. Runtime routing moves the same way, off a `gemini-` prefix test onto membership in `GEMINI_NATIVE_IMAGE_MODELS`: a `gemini-*` image model not present in that list now routes to `generateImages` instead of `generateContent`, so it fails against that endpoint rather than silently taking the native path.
+
+`responseModalities` stays a protected adapter default (`['TEXT', 'IMAGE']`) and is deliberately absent from the new type. `modelOptions.imageConfig` merges **over** the `imageConfig` derived from the portable `size` option, per field — passing only `imageConfig.imageSize` keeps the `aspectRatio` that `size` implied. `HarmCategory` and `HarmBlockThreshold` are now re-exported so `safetySettings` can be written without adding `@google/genai` to your own dependencies.
+
+**BREAKING (types only):** Imagen fields no longer compile on the four Gemini-native image models — `aspectRatio`, `negativePrompt`, `personGeneration`, `safetyFilterLevel`, `addWatermark`, `language`, `outputMimeType`, `outputCompressionQuality`, `guidanceScale`, `enhancePrompt`, `includeSafetyAttributes`, `includeRaiReason`, `outputGcsUri`, `labels`. They previously type-checked but were already dropped at runtime (only `seed` was ever forwarded to `generateContent`), so no request behaviour changes — the compiler now reports what was already happening. Migrate `aspectRatio` to the portable `size` option (`'16:9_4K'`) or to `modelOptions.imageConfig`, and drop the rest. `GeminiImageAdapter.generateImages` (and its `~types.providerOptions`) also widens from `ImageGenerationOptions<GeminiImageProviderOptions>` to `ImageGenerationOptions<GeminiAnyImageProviderOptions>`, which affects code structurally annotated against the old signature.

docs/adapters/gemini.md

@@ -414,7 +414,7 @@ The Gemini adapter supports two types of image generation:
 - **Gemini native image models** (NanoBanana) — Use the `generateContent` API with models like `gemini-3.1-flash-image-preview`. These support extended resolution tiers (1K, 2K, 4K) and aspect ratio control.
 - **Imagen models** — Use the `generateImages` API with models like `imagen-4.0-generate-001`. These are dedicated image generation models with WIDTHxHEIGHT sizing.
 
-The adapter automatically routes to the correct API based on the model name — models starting with `gemini-` use `generateContent`, while `imagen-` models use `generateImages`.
+The adapter automatically routes to the correct API based on membership in the known list of Gemini native models (`GEMINI_NATIVE_IMAGE_MODELS`) — the Gemini native models listed above use `generateContent`, while Imagen models, and any model id this package doesn't know about, use `generateImages`.
 
 ### Example: Gemini Native Image Generation (NanoBanana)
 
@@ -494,6 +494,10 @@ const result = await generateImage({
 
 ### Image Model Options
 
+`modelOptions` is typed per model family, because the two families hit different APIs.
+
+Imagen models (`generateImages`) take `GenerateImagesConfig` fields:
+
 ```typescript ignore
 import { generateImage } from "@tanstack/ai";
 import { geminiImage } from "@tanstack/ai-gemini";
@@ -509,6 +513,26 @@ const result = await generateImage({
 });
 ```
 
+Gemini native models (`generateContent`) take `GenerateContentConfig` fields instead — `seed`, `safetySettings`, `thinkingConfig`, `imageConfig`, and `systemInstruction`:
+
+```typescript
+import { generateImage } from "@tanstack/ai";
+import { geminiImage } from "@tanstack/ai-gemini";
+
+const result = await generateImage({
+  adapter: geminiImage("gemini-3.1-flash-image-preview"),
+  prompt: "...",
+  size: "16:9_4K",
+  modelOptions: {
+    thinkingConfig: { thinkingBudget: 512 },
+    // Merged over the imageConfig derived from `size`, per field.
+    imageConfig: { imageSize: "2K" },
+  },
+});
+```
+
+See [Image Generation](../media/image-generation) for the full native option list.
+
 ## Text-to-Speech (Experimental)
 
 > **Note:** Gemini TTS is experimental and may require the Live API for full functionality.

docs/config.json

@@ -437,7 +437,7 @@
           "label": "Image Generation",
           "to": "media/image-generation",
           "addedAt": "2026-04-15",
-          "updatedAt": "2026-08-13"
+          "updatedAt": "2026-08-14"
         },
         {
           "label": "Video Generation",
@@ -796,7 +796,7 @@
           "label": "Google Gemini",
           "to": "adapters/gemini",
           "addedAt": "2026-04-15",
-          "updatedAt": "2026-07-22"
+          "updatedAt": "2026-08-14"
         },
         {
           "label": "Ollama",

docs/media/image-generation.md

@@ -634,7 +634,7 @@ const result = await generateImage({
 
 #### Gemini Native Model Options (NanoBanana)
 
-Gemini native image models accept `GenerateContentConfig` options directly in `modelOptions`:
+Gemini native image models are served by `generateContent`, so their `modelOptions` are `GenerateContentConfig` fields — a different shape from the Imagen options above:
 
 ```typescript
 import { generateImage } from "@tanstack/ai";
@@ -644,9 +644,43 @@ const result = await generateImage({
   adapter: geminiImage("gemini-3.1-flash-image-preview"),
   prompt: "A beautiful garden",
   size: "16:9_4K",
+  modelOptions: {
+    seed: 42,
+    thinkingConfig: { thinkingBudget: 512 },
+    systemInstruction: "Always render in watercolor.",
+    // Merged over the imageConfig derived from `size`, per field: this keeps
+    // the 16:9 aspect ratio and overrides only the resolution tier.
+    imageConfig: { imageSize: "2K" },
+  },
+});
+```
+
+`safetySettings` takes the SDK's `HarmCategory` / `HarmBlockThreshold` enums, so plain strings won't type-check. Both are re-exported from `@tanstack/ai-gemini` — you don't need `@google/genai` in your own dependencies:
+
+```typescript
+import { generateImage } from "@tanstack/ai";
+import {
+  HarmBlockThreshold,
+  HarmCategory,
+  geminiImage,
+} from "@tanstack/ai-gemini";
+
+const result = await generateImage({
+  adapter: geminiImage("gemini-3.1-flash-image-preview"),
+  prompt: "A beautiful garden",
+  modelOptions: {
+    safetySettings: [
+      {
+        category: HarmCategory.HARM_CATEGORY_HATE_SPEECH,
+        threshold: HarmBlockThreshold.BLOCK_ONLY_HIGH,
+      },
+    ],
+  },
 });
 ```
 
+`responseModalities` is not accepted — the adapter always requests `['TEXT', 'IMAGE']`, so nothing can silently disable image output.
+
 ### Response Format
 
 The image generation result includes:

packages/ai-gemini/src/adapters/image.ts

@@ -7,6 +7,7 @@ import {
 } from '../utils'
 import { buildGeminiUsage } from '../usage'
 import {
+  isGeminiNativeImageModel,
   parseNativeImageSize,
   sizeToAspectRatio,
   validateImageSize,
@@ -15,10 +16,11 @@ import {
 } from '../image/image-provider-options'
 import type { GEMINI_IMAGE_MODELS } from '../model-meta'
 import type {
+  GeminiAnyImageProviderOptions,
   GeminiImageModelInputModalitiesByName,
   GeminiImageModelProviderOptionsByName,
   GeminiImageModelSizeByName,
-  GeminiImageProviderOptions,
+  GeminiNativeImageProviderOptions,
 } from '../image/image-provider-options'
 import type {
   GeneratedImage,
@@ -35,6 +37,7 @@ import type {
   GenerateImagesConfig,
   GenerateImagesResponse,
   GoogleGenAI,
+  ImageConfig,
   Part,
 } from '@google/genai'
 import type { GeminiClientConfig } from '../utils/client'
@@ -65,7 +68,7 @@ export class GeminiImageAdapter<
   TModel extends GeminiImageModel,
 > extends BaseImageAdapter<
   TModel,
-  GeminiImageProviderOptions,
+  GeminiAnyImageProviderOptions,
   GeminiImageModelProviderOptionsByName,
   GeminiImageModelSizeByName,
   GeminiImageModelInputModalitiesByName
@@ -75,7 +78,7 @@ export class GeminiImageAdapter<
 
   // Type-only property - never assigned at runtime
   declare '~types': {
-    providerOptions: GeminiImageProviderOptions
+    providerOptions: GeminiAnyImageProviderOptions
     modelProviderOptionsByName: GeminiImageModelProviderOptionsByName
     modelSizeByName: GeminiImageModelSizeByName
     modelInputModalitiesByName: GeminiImageModelInputModalitiesByName
@@ -89,7 +92,7 @@ export class GeminiImageAdapter<
   }
 
   async generateImages(
-    options: ImageGenerationOptions<GeminiImageProviderOptions>,
+    options: ImageGenerationOptions<GeminiAnyImageProviderOptions>,
   ): Promise<ImageGenerationResult> {
     const { model, logger } = options
 
@@ -121,7 +124,7 @@ export class GeminiImageAdapter<
         )
       }
 
-      if (this.isGeminiImageModel(model)) {
+      if (isGeminiNativeImageModel(model)) {
         return await this.generateWithGeminiApi(options, resolved)
       }
 
@@ -155,29 +158,40 @@ export class GeminiImageAdapter<
     }
   }
 
-  private isGeminiImageModel(model: string): boolean {
-    return model.startsWith('gemini-')
-  }
-
   private async generateWithGeminiApi(
-    options: ImageGenerationOptions<GeminiImageProviderOptions>,
+    options: ImageGenerationOptions<GeminiNativeImageProviderOptions>,
     resolved: ResolvedMediaPrompt,
   ): Promise<ImageGenerationResult> {
     const { model, size, numberOfImages, modelOptions } = options
 
     const parsedSize = size ? parseNativeImageSize(size) : undefined
 
-    // GeminiImageProviderOptions is Imagen-shaped — most fields
-    // (personGeneration, safetyFilterLevel, addWatermark, outputMimeType,
-    // outputCompressionQuality, guidanceScale, enhancePrompt,
-    // includeSafetyAttributes, includeRaiReason, outputGcsUri, labels,
-    // negativePrompt, language) are only valid on GenerateImagesConfig and
-    // would be rejected by the Gemini-native generateContent path. Pick only
-    // the fields that are valid on GenerateContentConfig instead of spreading
-    // the whole options object.
-    const nativeConfig: GenerateContentConfig = {}
-    if (modelOptions?.seed !== undefined) {
-      nativeConfig.seed = modelOptions.seed
+    // The portable `size` option is the baseline; modelOptions.imageConfig is
+    // the provider escape hatch and wins per field, so a caller passing only
+    // `imageConfig.imageSize` keeps the aspectRatio derived from `size`.
+    const imageConfig: ImageConfig = {
+      ...(parsedSize?.aspectRatio && { aspectRatio: parsedSize.aspectRatio }),
+      ...(parsedSize?.resolution && { imageSize: parsedSize.resolution }),
+      ...modelOptions?.imageConfig,
+    }
+
+    // Named picks, never a wholesale spread: the Imagen-shaped fields of
+    // GeminiImageProviderOptions (personGeneration, safetyFilterLevel,
+    // addWatermark, outputMimeType, …) are only valid on GenerateImagesConfig
+    // and would be rejected by generateContent. Picking by name means no
+    // Imagen field can reach this path even if one slips past the per-model
+    // provider-options map.
+    const nativeConfig: GenerateContentConfig = {
+      ...(modelOptions?.seed !== undefined && { seed: modelOptions.seed }),
+      ...(modelOptions?.safetySettings !== undefined && {
+        safetySettings: modelOptions.safetySettings,
+      }),
+      ...(modelOptions?.thinkingConfig !== undefined && {
+        thinkingConfig: modelOptions.thinkingConfig,
+      }),
+      ...(modelOptions?.systemInstruction !== undefined && {
+        systemInstruction: modelOptions.systemInstruction,
+      }),
     }
 
     const config: GenerateContentConfig = {
@@ -186,16 +200,7 @@ export class GeminiImageAdapter<
       // IMPORTANT: responseModalities is a protected default — set it AFTER
       // nativeConfig so nothing can silently disable image output.
       responseModalities: ['TEXT', 'IMAGE'],
-      ...(parsedSize && {
-        imageConfig: {
-          ...(parsedSize.aspectRatio && {
-            aspectRatio: parsedSize.aspectRatio,
-          }),
-          ...(parsedSize.resolution && {
-            imageSize: parsedSize.resolution,
-          }),
-        },
-      }),
+      ...(Object.keys(imageConfig).length > 0 && { imageConfig }),
     }
 
     const contents = this.buildContents(resolved, numberOfImages)
@@ -321,19 +326,70 @@ export class GeminiImageAdapter<
   }
 
   private buildImagenConfig(
-    options: ImageGenerationOptions<GeminiImageProviderOptions>,
+    options: ImageGenerationOptions<GeminiAnyImageProviderOptions>,
   ): GenerateImagesConfig {
     const { size, numberOfImages, modelOptions } = options
 
     // Build with conditional spreads — under exactOptionalPropertyTypes the
     // vendor `GenerateImagesConfig` fields are `field?: T` (no `| undefined`),
     // so we can only assign the property when we actually have a value.
     const sizeAspectRatio = size ? sizeToAspectRatio(size) : undefined
+
+    // Named picks, never a wholesale spread — the mirror image of the native
+    // path below. A native-only field (safetySettings, thinkingConfig,
+    // imageConfig, systemInstruction) belongs to GenerateContentConfig and is
+    // rejected by generateImages with 400 INVALID_ARGUMENT, so it must not be
+    // able to reach here even when the caller's `modelOptions` was typed
+    // against both shapes at once (e.g. an adapter inferred from a union of
+    // model names).
     return {
       numberOfImages: numberOfImages ?? 1,
-      // Map size to aspect ratio if provided (modelOptions.aspectRatio will override)
+      // Map size to aspect ratio if provided; modelOptions.aspectRatio,
+      // picked after it, overrides.
       ...(sizeAspectRatio !== undefined && { aspectRatio: sizeAspectRatio }),
-      ...modelOptions,
+      ...(modelOptions?.aspectRatio !== undefined && {
+        aspectRatio: modelOptions.aspectRatio,
+      }),
+      ...(modelOptions?.personGeneration !== undefined && {
+        personGeneration: modelOptions.personGeneration,
+      }),
+      ...(modelOptions?.safetyFilterLevel !== undefined && {
+        safetyFilterLevel: modelOptions.safetyFilterLevel,
+      }),
+      ...(modelOptions?.seed !== undefined && { seed: modelOptions.seed }),
+      ...(modelOptions?.addWatermark !== undefined && {
+        addWatermark: modelOptions.addWatermark,
+      }),
+      ...(modelOptions?.language !== undefined && {
+        language: modelOptions.language,
+      }),
+      ...(modelOptions?.negativePrompt !== undefined && {
+        negativePrompt: modelOptions.negativePrompt,
+      }),
+      ...(modelOptions?.outputMimeType !== undefined && {
+        outputMimeType: modelOptions.outputMimeType,
+      }),
+      ...(modelOptions?.outputCompressionQuality !== undefined && {
+        outputCompressionQuality: modelOptions.outputCompressionQuality,
+      }),
+      ...(modelOptions?.guidanceScale !== undefined && {
+        guidanceScale: modelOptions.guidanceScale,
+      }),
+      ...(modelOptions?.enhancePrompt !== undefined && {
+        enhancePrompt: modelOptions.enhancePrompt,
+      }),
+      ...(modelOptions?.includeSafetyAttributes !== undefined && {
+        includeSafetyAttributes: modelOptions.includeSafetyAttributes,
+      }),
+      ...(modelOptions?.includeRaiReason !== undefined && {
+        includeRaiReason: modelOptions.includeRaiReason,
+      }),
+      ...(modelOptions?.outputGcsUri !== undefined && {
+        outputGcsUri: modelOptions.outputGcsUri,
+      }),
+      ...(modelOptions?.labels !== undefined && {
+        labels: modelOptions.labels,
+      }),
     }
   }