Generate Image
AI image generation, 20+ models
Try asking
About
Generate Image is a full AI image studio with 20+ models in a single tool. Create from text prompts, edit existing images with natural language, compose new shots from multiple references, and upscale to 4K — all without switching between different apps.
Models range from fast drafts to photorealistic renders, with specialists for typography, editorial photography, product shots, and artistic styles. The default model handles most tasks well; call list_models when you need something specific. Models approximate fonts, so use a text-focused model for drafts, then typeset pixel-exact copy in a design tool. Checked 2026-08-26 against this manifest.
What you can do
- text_to_image — generate images from text prompts with full control over size, style, aspect ratio, and model
- edit_image — modify an existing image using natural language: change backgrounds, add objects, remove elements, apply style transfers
- image_to_image — compose a new image from up to 4 reference photos, combining scenes, personas, products, and outfits
- check_image — poll for a pending result when a model is still processing
- upscale_image — increase resolution up to 10x from inside the same workflow
- list_models — browse all available models with capabilities and supported parameters
Who it's for
Marketers producing campaign assets, product teams visualizing concepts, content creators building visual content, and developers adding image generation to AI workflows.
How to use it
- Use text_to_image with a descriptive prompt — the default model works well without setting a model parameter
- Use edit_image to modify an existing photo with a natural language instruction
- Use image_to_image to combine reference images from your library (scenes, personas, products, outfits) into one new composition
- Use list_models to pick a specialist model for typography, photorealism, or artistic styles
- Use check_image before delivery and upscale_image only after approving the composition
Getting started
Try text_to_image with a wireless headphone on marble under soft studio light. Representative manifest-schema shape: images, model, seed [redacted]. Visual accuracy is not guaranteed; timing varies.
Generate an image from a text prompt
Edit, transform, or inpaint an existing image using AI
Compose a new image from up to 4 reference images plus a prompt
AI-upscale an image up to 10x resolution
Generate an image from a text prompt. 20+ models with different quality, speed, style, and cost tradeoffs. Returns image URL(s) and downloadable assets.
Edit, transform, or inpaint an existing image using AI. Provide an image URL and a text prompt describing the edit. Supports style transfer, object removal/addition, background changes, and targeted edits with masks.
Compose a new image from up to 4 reference images plus a prompt. Combine scenes, personas, outfits, products into one shot. Pass scene_file_id, persona_file_id, product_file_ids[], outfit_file_ids[], and/or raw image_urls[]. Default model nano-banana-2.
Check on a pending fal-backed image request and retrieve it if ready. Use after text_to_image or edit_image returns status: pending with a fal_request_id.
AI-upscale an image up to 10x resolution. Enhances detail and sharpness while maintaining quality. Choose a target resolution (720p-4K) or a scale factor (1-10x).
List all available image generation models with capabilities, pricing, supported parameters, and descriptions. Filter by capability type. Models not in the registry can still be used by passing their model endpoint ID directly.
- Added Sume as an image provider — 25 models including GPT Image 2.5, Nano Banana Pro, Seedream 5.0, Ideogram 4.5, Recraft 4.1, Grok Imagine 2 and Higgsfield Soul. Use a sume- prefixed model key.
- Image editing works again with Qwen Image, Seedream, FireRed, and Grok Imagine edit models
- Saved personas, scenes, outfits, and products now describe themselves to the model — every reference is numbered and labelled automatically, so prompts no longer need to say which image is which.
- Reference limits now follow the model: up to 16 references on GPT Image 2 and 2.5, 9 on FLUX.2, 10 on Seedream.
- image_urls accepts saved file IDs (ast_…) alongside public URLs, and singular product_file_id / outfit_file_id are accepted too.
- Every problem with a reference is reported in one message with what to do about it.
- Added Qwen Image 3, Recraft V4.1, Kling Image O3, Grok Imagine 2.0, and FLUX.2 Flex/Max.
- Higgsfield Soul is now Soul 2.0. Phota now uses its latest base model by default.
- Imagen 4 on Google direct is retired — use Nano Banana 2 or Nano Banana Pro instead.
- Higgsfield is now a full image provider — Soul, Soul 2, Popcorn, Reve, Ideogram 4, Recraft 4.1, Grok Imagine 2.0, and the rest of the official Higgsfield image catalog. Use a higgsfield- prefixed model key or backend.order to route through Higgsfield.
- Added OpenAI GPT Image 2.5 — "gpt-image-2.5-flare" (fast default, up to 50% faster than GPT Image 2) and "gpt-image-2.5-sunburst" (precision tier for campaign and product work). Both have edit variants ("-edit") that take up to 16 reference images plus an optional mask.
- Quality now accepts xhigh and max (plus auto) for GPT Image models — available on edit_image and image_to_image as well as text_to_image.
- Fixed edit_image and image_to_image with GPT Image edit models — reference images were not reaching the model.
- Default look when no style reference is given is now a candid iPhone camera-roll photo instead of an editorial lifestyle shoot. Pass style_reference_prompt or a saved style to override.
- Added request-time same-model provider switching and retryable failover through backend.order, backend.only, backend.ignore, and backend.allow_fallbacks.
- Added Seedream 5.0 Pro and Lite for higher-quality image generation and editing. Nano Banana 2 remains the default.
- Added typography guidance to model descriptions — ideogram-v3 and gpt-image-2 are now clearly recommended for text-heavy designs; gallery previews now show the generation prompt instead of a generic label.
- Quality-tier billing now exact across every model — low/medium/high calls on gpt-image-2 bill at $0.01 / $0.04 / $0.15 instead of a flat rate. Applies universally: every provider now returns the true per-request cost.
- Added OpenAI GPT Image 2 — excellent prompt adherence and typography. Model key "gpt-image-2" works for text-to-image, editing, and mask-based inpainting; add image_urls to trigger edit mode. Also available via OpenRouter as "openrouter-gpt-image-2".
- Accept personas, scenes, products, and outfits from your file library
- Added image_to_image skill — compose new images from up to 4 reference URLs or saved file IDs (scenes, personas, products, outfits) plus a prompt. Pass scene_file_id, persona_file_id, product_file_ids, outfit_file_ids, or raw image_urls and the gateway resolves and merges them automatically.
- Style references now supported — pass a style reference when generating or editing images for consistent visual direction
- Ideogram style values are now case-insensitive — pass "realistic" or "REALISTIC", both work (valid: REALISTIC, GENERAL, DESIGN, AUTO)
- Added Google as a direct provider — 6 new models: Nano Banana, Nano Banana 2, Nano Banana Pro, Imagen 4 Fast, Imagen 4, Imagen 4 Ultra. Direct Google API access with token-based billing for Nano Banana and flat-rate for Imagen 4.
- Added Photalabs as a direct provider — phota, phota-edit, and phota-enhance models now route via the Photalabs API instead of fal.ai
- Default model changed to nano-banana-2
- Removed hardcoded model names from descriptions — agents discover models dynamically via list_models
- Added Higgsfield Soul model — realistic human-centric UGC photos (model key: "soul")
- Added Prodia as an alternative provider — 10 new models (prodia-flux-2-pro, prodia-sdxl, prodia-recraft-v4, prodia-gemini-3-pro, etc.)
- upscale_image now delegates to image-upscale tool via composition
- Fuzzy model resolution — natural names like "nano banana 2" now resolve correctly
- Broadened live model discovery to catch more fal.ai categories and tags
- Added subtitle, expanded description, and agent instructions
- Initial release
- Expanded Nano Banana aspect-ratio support to include auto, ultrawide, and extreme ratios
- fal-backed image generation now returns recoverable pending results with check_image instead of silently degrading on long queue waits
- fal live-pricing lookup now falls back to static registry pricing to avoid pricing-endpoint outages breaking image generation
Powered by
Related
Guides
- ChatGPT MCP Not Working: Fixes That Match the ProductAdd Toolrouter from the plugin directory. If custom MCP fails, check Developer mode, the chat toggle, and the server URL.
- Claude Skills vs MCP: When to Use WhichSkills teach Claude how. MCP gives it access. You need both. Google already suggests this search. ToolRouter is the access layer.
- How Much Does MCP Cost for a Small Team?Claude and ChatGPT seats are separate. ToolRouter team is $29/seat with shared credits. Free catalog tools stay free. Paid skills start at $0.01.
- How to Add MCP on Claude TeamOn Claude Team, only an Owner adds a custom connector. Members then Connect. URL for ToolRouter is https://api.toolrouter.com/mcp/anthropic.
Questions
Can I create an image from a text prompt?
Yes. `text_to_image` turns a prompt into a new image and returns downloadable image assets.
Can I edit an image I already have?
Yes. `edit_image` can change backgrounds, add or remove objects, and make targeted visual edits.
Can I upscale an image for higher resolution?
Yes. `upscale_image` can enhance an image up to 10x resolution, including 4K-level output.
How do I know which model to use?
Use `list_models` to compare capabilities, speed, and style strengths before you generate.