TanStack
diff --git a/‎.changeset/gemini-veo-video-adapter.md‎
Lines changed: 42 additions & 0 deletions b/‎.changeset/gemini-veo-video-adapter.md‎
Lines changed: 42 additions & 0 deletions
diff --git a/‎.changeset/image-and-video-inputs.md‎
Lines changed: 30 additions & 0 deletions b/‎.changeset/image-and-video-inputs.md‎
Lines changed: 30 additions & 0 deletions
diff --git a/‎.gitignore‎
Lines changed: 1 addition & 0 deletions b/‎.gitignore‎
Lines changed: 1 addition & 0 deletions
diff --git a/‎.prettierignore‎
Lines changed: 1 addition & 0 deletions b/‎.prettierignore‎
Lines changed: 1 addition & 0 deletions
diff --git a/‎docs/adapters/grok.md‎
Lines changed: 45 additions & 0 deletions b/‎docs/adapters/grok.md‎
Lines changed: 45 additions & 0 deletions
diff --git a/‎docs/media/image-generation.md‎
Lines changed: 165 additions & 2 deletions b/‎docs/media/image-generation.md‎
Lines changed: 165 additions & 2 deletions
@@ -0,0 +1,42 @@
+---
+'@tanstack/ai': minor
+'@tanstack/ai-gemini': minor
+---
+
+Add a Google Veo video adapter (`geminiVideo` / `createGeminiVideo`) and the
+per-model typed-duration video contract it is built on (#534, #634).
+
+**`@tanstack/ai`** (additive, non-breaking): `VideoAdapter` /
+`BaseVideoAdapter` gain a `TModelDurationByName` generic (defaulting to
+`Record<string, number>`, preserving today's `duration?: number` typing for
+adapters without a map) plus two introspection methods with safe defaults:
+
+- `availableDurations()` — a `DurationOptions` tagged union
+  (`discrete | range | mixed | none`) describing the durations the current
+  model accepts. Default: `{ kind: 'none' }`.
+- `snapDuration(seconds)` — coerce raw seconds to the closest valid duration
+  (`snapToDurationOption` is exported for adapter authors). Default:
+  `undefined`.
+
+`generateVideo({ duration })` is now typed per model via
+`VideoDurationForAdapter<TAdapter>`.
+
+**`@tanstack/ai-gemini`**: new Veo adapter over the long-running
+`:predictLongRunning` operation, supporting `veo-3.1-generate-preview`,
+`veo-3.1-fast-generate-preview`, `veo-3.0-generate-001`,
+`veo-3.0-fast-generate-001`, and `veo-2.0-generate-001`:
+
+- `geminiVideo('veo-3.0-generate-001')` → `duration?: 4 | 6 | 8`
+  (Veo 2: `5 | 6 | 8`); `adapter.snapDuration(7)` → `6`.
+- Multimodal prompts: the first un-roled / `'start_frame'` image part
+  becomes the input image, `'end_frame'` → `lastFrame`, `'reference'` /
+  `'character'` → `referenceImages`.
+- `size` takes Veo aspect ratios (`'16:9' | '9:16'`); everything else from
+  the SDK's `GenerateVideosConfig` (e.g. `resolution`, `generateAudio`,
+  `negativePrompt`) is available through `modelOptions`.
+- Responsible-AI filtering is surfaced as a failed job with the filter
+  reasons.
+
+Note: Veo result URLs are served by the Gemini Files API and require the
+Google API key to download (`x-goog-api-key` header or `key` query
+parameter).
@@ -0,0 +1,30 @@
+---
+'@tanstack/ai': minor
+'@tanstack/ai-openai': minor
+'@tanstack/ai-gemini': minor
+'@tanstack/ai-fal': minor
+'@tanstack/ai-grok': minor
+'@tanstack/ai-openrouter': minor
+'@tanstack/ai-client': minor
+'@tanstack/ai-event-client': patch
+---
+
+`generateImage()` and `generateVideo()` now accept a multimodal `prompt`: a plain string, or an ordered array of content parts (`TextPart` / `ImagePart` / `VideoPart` / `AudioPart`) for image-conditioned generation, image-to-image, multi-reference, image-to-video, and edit / inpaint flows. Part order is meaningful — "not like this _(image)_, more like this _(image)_" — and each media part may carry a `metadata.role` hint (`'reference' | 'mask' | 'control' | 'start_frame' | 'end_frame' | 'character'`) that adapters use to route to the provider-specific field, plus an informational `metadata.tag` label for your own bookkeeping. The accepted part types are narrowed per model at compile time via each adapter's input-modality map, so passing an image part to a text-only model is a type error (with a clear runtime throw as backstop).
+
+Prompt text is always sent **verbatim** — the SDK never injects or rewrites in-prompt referencing markers. To reference inputs from your prompt, write the provider's own convention (fal Kling / Seedance `@Image1`, OpenAI / FLUX.2 `"image 1"` prose, Gemini content descriptions); see the image-generation docs for the per-provider table.
+
+Provider behavior in this release:
+
+- **OpenAI image** — Prompts with image parts route `gpt-image-2` / `gpt-image-1` / `gpt-image-1-mini` to `images.edit()` (up to 16 source images plus optional mask); `dall-e-2` routes to `images.edit()` with one source image; `dall-e-3` rejects image parts at compile time and at runtime.
+- **OpenAI video** — Sora-2 / Sora-2-Pro accept a single image part as `input_reference`; passing more than one throws.
+- **Gemini image** — Native models (`gemini-*-flash-image`, "nano-banana") map prompt parts 1:1 onto multimodal `contents`, preserving interleaved order. Imagen is text-only (compile-time + runtime rejection).
+- **fal.ai** — Field names resolve per endpoint from a map generated from the fal SDK's endpoint types (362 endpoints with nonstandard fields, e.g. nano-banana edit → `image_urls`, Kling i2v start frame → `image_url`, Veo first-last-frame → `first_frame_url` / `last_frame_url`). Defaults for endpoints not in the map: single → `image_url`, multiple → `image_urls`; `role: 'mask'` → `mask_url`; `role: 'control'` → `control_image_url`; `role: 'reference'` / `'character'` → `reference_image_urls`; video `role: 'start_frame'` / `'end_frame'` → `start_image_url` / `end_image_url`. Per-model prompt modalities are derived at the type level from the SDK's endpoint input types. Regenerate the map after a fal SDK bump with `pnpm generate:fal-image-fields` (a unit test fails when it goes stale). In `FalImageProviderOptions` / `FalVideoProviderOptions`, media-conditioning fields the mappers can populate (`image_url`, `start_image_url`, `video_url`, `audio_url`, …) are demoted from required to optional — supply them as prompt parts, or keep passing them explicitly via `modelOptions`.
+- **Grok** — New `grok-imagine-image` / `grok-imagine-image-quality` models. Prompts with image parts route to xAI's JSON `/v1/images/edits` endpoint (up to 3 source images, addressed by xAI in request order; the prompt is sent verbatim). `role: 'mask'` / `'control'` throw. Their `size` uses an `aspectRatio_resolution` template (`'16:9_2k'`, suffix optional) mirroring Gemini's native image models. `grok-2-image-1212` remains text-to-image only.
+- **OpenRouter** — Prompt parts map 1:1 onto multimodal `text` / `image_url` chat content parts, preserving interleaved order, and are forwarded to the underlying image model. URL sources pass through verbatim (no fetching or re-encoding in your process); `data` sources become data URIs.
+- **Anthropic** — Unchanged (no image generation API).
+
+A new `resolveMediaPrompt()` utility (exported from `@tanstack/ai`) is the single downrev point from the canonical interleaved prompt shape to flattened text + per-modality part buckets, for adapter authors.
+
+On the client side, `ImageGenerateInput.prompt` and `VideoGenerateInput.prompt` (`@tanstack/ai-client`, and the `useGenerateImage` / `useGenerateVideo` hooks built on them) are widened from `string` to the same `MediaPrompt` shape, so prompt parts can be sent from the browser through your server route to `generateImage()` / `generateVideo()`.
+
+Closes #618.
@@ -78,3 +78,4 @@ solo.yml
 # Agent scratch output (gap-analysis reports, triage notes — generated locally)
 .agent/gap-analysis/
 .agent/triage/
+.agent/research/
@@ -5,6 +5,7 @@
 **/coverage
 **/dist
 **/docs
+packages/ai-fal/src/image/generated/
 pnpm-lock.yaml
 
 .angular
 
@@ -160,6 +160,51 @@ const result = await generateImage({
 console.log(result.images);
 ```
 
+The grok-imagine models (`grok-imagine-image`, `grok-imagine-image-quality`)
+are aspect-ratio sized — `size` takes an `aspectRatio_resolution` template
+like `"16:9_2k"` (the `_2k` suffix is optional):
+
+```typescript
+const result = await generateImage({
+  adapter: grokImage("grok-imagine-image"),
+  prompt: "A futuristic cityscape at sunset",
+  size: "16:9_2k",
+});
+```
+
+### Image Editing (image-to-image)
+
+The grok-imagine models accept image prompt parts for image-conditioned
+generation via xAI's `/v1/images/edits` endpoint — up to 3 source images,
+addressed by xAI in the order they appear in the prompt. Per xAI's docs
+there is no in-prompt referencing syntax; write the prompt naturally and
+your text is sent verbatim:
+
+```typescript
+const result = await generateImage({
+  adapter: grokImage("grok-imagine-image"),
+  prompt: [
+    {
+      type: "text",
+      content: "Render the product in the style of the second image",
+    },
+    {
+      type: "image",
+      source: { type: "url", value: "https://example.com/product.png" },
+    },
+    {
+      type: "image",
+      source: { type: "url", value: "https://example.com/style.png" },
+    },
+  ],
+});
+```
+
+URL sources are fetched by xAI's servers, so they must be publicly
+reachable; use a `data` source for private images. `grok-2-image-1212` is
+text-to-image only — image prompt parts are a compile-time type error and
+throw at runtime.
+
 ## Text-to-Speech
 
 Generate speech with Grok TTS:
 
@@ -22,7 +22,7 @@ TanStack AI provides support for image generation through dedicated image adapte
 
 Image generation is handled by image adapters that follow the same tree-shakeable architecture as other adapters in TanStack AI. The image adapters support:
 
-- **OpenAI**: DALL-E 2, DALL-E 3, GPT-Image-1, and GPT-Image-1-Mini models
+- **OpenAI**: DALL-E 2, DALL-E 3, GPT-Image-1, GPT-Image-1-Mini, and GPT-Image-2 models
 - **Gemini**: Gemini native image models (NanoBanana) and Imagen 3/4 models
 - **fal.ai**: 600+ models including Nano Banana Pro, FLUX, and more
 
@@ -76,7 +76,7 @@ All image adapters support these common options:
 | Option | Type | Description |
 |--------|------|-------------|
 | `adapter` | `ImageAdapter` | Image adapter instance with model (required) |
-| `prompt` | `string` | Text description of the image to generate (required) |
+| `prompt` | `string \| MediaPromptPart[]` | Description of the image to generate (required). A plain string, or — on models that support image-conditioned generation — an ordered array of content parts interleaving text with image inputs. See [Image-Conditioned Generation](#image-conditioned-generation) below. |
 | `numberOfImages` | `number` | Number of images to generate |
 | `size` | `string` | Size of the generated image in WIDTHxHEIGHT format |
 | `modelOptions?` | `object` | Model-specific options (renamed from `providerOptions`) |
@@ -130,6 +130,169 @@ const result = await generateImage({
 })
 ```
 
+## Image-Conditioned Generation
+
+For image-to-image, reference-guided, multi-reference, and edit / inpaint
+flows, pass the `prompt` as an ordered array of content parts — the same
+`TextPart` / `ImagePart` shapes used elsewhere for multimodal content:
+
+```typescript
+import { generateImage } from '@tanstack/ai'
+import { openaiImage } from '@tanstack/ai-openai'
+
+await generateImage({
+  adapter: openaiImage('gpt-image-2'),
+  prompt: [
+    { type: 'text', content: 'Turn this into a cinematic product photo' },
+    {
+      type: 'image',
+      source: { type: 'url', value: 'https://example.com/product.png' },
+    },
+  ],
+})
+```
+
+Part order is meaningful. Providers with natively multimodal prompts
+(Gemini image models, OpenRouter) receive the parts exactly as written, so
+text can refer to its neighbouring images:
+
+```typescript
+await generateImage({
+  adapter: geminiImage('gemini-3.1-flash-image-preview'),
+  prompt: [
+    { type: 'text', content: 'Not like this' },
+    { type: 'image', source: { type: 'url', value: badExampleUrl } },
+    { type: 'text', content: 'more like this' },
+    { type: 'image', source: { type: 'url', value: goodExampleUrl } },
+  ],
+})
+```
+
+Providers with named request fields (OpenAI, fal, xAI) extract the image
+parts and flatten the text (text parts are joined verbatim, paragraph
+separated).
+
+The accepted part types are narrowed **per model at compile time**: passing
+an image part to a text-only model (e.g. `dall-e-3`, Imagen) is a type
+error, not just a runtime throw.
+
+### Referencing images from your prompt
+
+**Your prompt text is always sent verbatim — the SDK never injects or
+rewrites referencing markers.** When you want the text to refer to specific
+input images, write the provider's own convention yourself:
+
+| Provider | Convention | Example |
+| -------- | ---------- | ------- |
+| **OpenAI** (gpt-image) | Indexed prose, per OpenAI's prompting guide | `"apply the style of image 2 to image 1"` |
+| **FLUX.2 on fal / BFL** | Indexed prose (BFL's docs parse `image N`) | `"subject from image 1, style from image 2"` |
+| **Gemini** (native image models) | Describe the reference by content/role | `"using the attached fabric sample as the texture"` |
+| **fal Kling / Seedance endpoints** | `@`-tags, 1-indexed by input order | `"Put @Image1 in the style of @Image2"` |
+| **xAI grok-imagine** | No in-prompt syntax — images addressed in request order | `"render the product in the style of the second image"` |
+
+To keep track of which part you meant by "image 2" or `@Image2`, you can
+label parts with the informational `metadata.tag` field — the SDK ignores
+it, but it keeps your code self-documenting:
+
+```typescript
+prompt: [
+  { type: 'text', content: 'Put @Image1 in the style of @Image2' },
+  { type: 'image', source: { type: 'url', value: productUrl },
+    metadata: { tag: 'product' } },
+  { type: 'image', source: { type: 'url', value: styleUrl },
+    metadata: { tag: 'style' } },
+]
+```
+
+### Source format
+
+`ImagePart.source` is a discriminated union supporting both URLs and inline
+base64 data — pass whichever you have:
+
+```typescript
+// URL source
+{ type: 'image', source: { type: 'url', value: 'https://example.com/img.png' } }
+
+// Inline base64 data (mimeType required)
+{ type: 'image', source: { type: 'data', value: base64String, mimeType: 'image/png' } }
+```
+
+OpenAI's edit endpoint requires file uploads; the adapter fetches URL sources
+and converts base64 to a `File` automatically.
+
+### Role hints via `metadata.role`
+
+When a generation has multiple inputs with different roles (mask vs reference
+vs start/end frame), set `metadata.role` on each part. Adapters route by role
+to the provider-specific field; parts without a role fall back to positional
+mapping.
+
+| Role            | Maps to                                                                                |
+| --------------- | -------------------------------------------------------------------------------------- |
+| `'reference'`   | fal `reference_image_urls`; Gemini multimodal part; positional fallback                |
+| `'character'`   | Same as `'reference'`; Veo `referenceImages` slot (planned — no Veo adapter yet)       |
+| `'mask'`        | OpenAI `mask` (gpt-image-2, gpt-image-1, dall-e-2); fal `mask_url`                     |
+| `'control'`     | fal `control_image_url` (ControlNet / depth / pose conditioning)                       |
+| `'start_frame'` | fal `start_image_url`; Veo `image` (planned) (used by `generateVideo`)                 |
+| `'end_frame'`   | fal `end_image_url`; Veo `lastFrame` (planned) (used by `generateVideo`)               |
+
+#### Inpaint / edit with a mask
+
+```typescript
+await generateImage({
+  adapter: openaiImage('gpt-image-2'),
+  prompt: [
+    { type: 'text', content: 'Replace the masked region with a tree' },
+    {
+      type: 'image',
+      source: { type: 'url', value: photoUrl },
+    },
+    {
+      type: 'image',
+      source: { type: 'url', value: maskUrl },
+      metadata: { role: 'mask' },
+    },
+  ],
+})
+```
+
+#### Multi-reference composition
+
+```typescript
+await generateImage({
+  adapter: geminiImage('gemini-3.1-flash-image-preview'),
+  prompt: [
+    {
+      type: 'text',
+      content:
+        'Generate a new image of the product using the style of the second reference',
+    },
+    {
+      type: 'image',
+      source: { type: 'url', value: 'https://example.com/product.png' },
+    },
+    {
+      type: 'image',
+      source: { type: 'url', value: 'https://example.com/style.png' },
+    },
+  ],
+})
+```
+
+### Provider support
+
+| Provider     | Behavior                                                                                                  |
+| ------------ | --------------------------------------------------------------------------------------------------------- |
+| **OpenAI**   | `gpt-image-2` / `gpt-image-1` / `gpt-image-1-mini` → routes to `images.edit()`, up to 16 source images plus optional mask.<br>`dall-e-2` → `images.edit()` with 1 source image only.<br>`dall-e-3` → throws (no edit support). |
+| **Gemini**   | Native models (`gemini-*-flash-image`, "nano-banana", etc.) → prompt parts map 1:1 onto multimodal `contents`, preserving interleaved order. Up to ~14 input images (provider limit, not enforced by the SDK).<br>Imagen models → throws (text-to-image only). |
+| **fal.ai**   | Field names resolve per endpoint from a map generated from the fal SDK's endpoint types (e.g. nano-banana edit gets `image_urls`, Fooocus masks get `mask_image_url`). Defaults for unknown endpoints: 1 input → `image_url`; multiple → `image_urls`; `role: 'mask'` → `mask_url`; `role: 'control'` → `control_image_url`; `role: 'reference'` / `'character'` → `reference_image_urls`. Override with `modelOptions` for endpoint-specific fields. |
+| **Grok**     | grok-imagine models → xAI's `/v1/images/edits` (up to 3 source images, addressed by xAI in request order; prompt sent verbatim). `role: 'mask'` / `'control'` throw (no Imagine API equivalent). `grok-2-image-1212` throws (text-to-image only). |
+| **OpenRouter** | Prompt parts map 1:1 onto multimodal `image_url` / `text` content parts, preserving interleaved order, and are forwarded to the underlying image model.                                                                                    |
+| **Anthropic** | n/a — no image generation API.                                                                                                                                                                          |
+
+Adapters that don't support image-conditioned generation throw a clear
+runtime error so calls fail fast rather than silently dropping the inputs.
+
 ## Model Options
 
 ### OpenAI Model Options