Generate Image

Generate AI Image

Generate an image from a text prompt using Simplified, across many leading AI providers, and return a viewable image URL (plus an asset id you can reuse).

What it can do

• Text-to-image (capability: "prompt") — make an image from a description.
• Image editing / image-to-image (capability: "reference_image") — transform

or edit using one reference image.

• Multi-reference composition (capability: "multiple_images") — guide with

several reference images (supported on some models).

Good for: product shots, hero/banner images, social graphics, illustrations, 3D-style renders, icons/logo concepts, photoreal scenes, and text rendered inside the image (posters, quote cards, ad headlines).

How to use it

1. Discover — call api_getModelFields(type: "image") to get the current list of

models, capabilities, and credit costs. Filter out models that cannot satisfy the requested capability; don't choose on model name alone.

1. Choose the model — use [Model selection](#model-selection). For the selected

model, call api_getModelFields(type: "image", model_id, capability) and use its exact parameters schema. This is the source of truth; don't guess model ids, capabilities, field names, or costs.

1. Choose storage — transient for a one-off, asset to reuse the image (e.g.

post it via the simplified-social skill).

1. Explain the choice when it matters — before a costly or ambiguous request,

name the selected model, why it fits, and the discovered credit cost. If the user explicitly chose a model, honor it when it supports the requested capability.

1. Generate — call api_generateImage with parameters matching the discovered

schema. This spends credits.

1. Present the result — show the returned URL as a link, never embedded (see

[Presenting the result](#presenting-the-result)).

For an ordinary prompt-only request, use the quality-first default below after confirming it is still available. Always inspect live fields for reference-image, multi-image, exact-size, quality, or resolution requests.

The request

The tools

• api_getModelFields — discover available models and the per-(model, capability)

field schema. Read-only, spends no credits. Call it first.

• api_generateImage — consumes paid AI credits.

Fields

Top-level fields for api_generateImage:

• model — a model id from api_getModelFields (e.g. google.gemini-3.1-flash-image-preview).
• capability — prompt | reference_image | multiple_images.
• storage — see [Storage](#storage) (default transient).
• parameters — a required nested object; never flatten its fields to the top

level, and put the prompt text in parameters.prompt (not in capability).

The exact keys inside parameters vary by model — get them from api_getModelFields(type: "image", model_id, capability), don't assume. They differ in real ways: most models take aspect_ratio, but OpenAI GPT Image uses size + quality + count, Gemini adds image_size, Flux 2 uses resolution, and the reference-image field is variously named input_image, image_prompt, reference_images, source_image, or style_reference_images.

Resolving Simplified asset references

Treat a Simplified asset_id as the canonical reference, but follow the live model schema at the generation boundary. When a model field is a URL or URL list (for example Gemini reference_images):

1. Call api_getAsset with the permanent asset UUID.
2. Require status: 4 (DONE) and the expected asset_type before generating.
3. Pass the current file_url returned by api_getAsset into the model-specific

reference field. If the URL is signed, preserve its complete query string and use it before expiry.

1. Do not trust a cached URL copied from a brand-kit record when an asset_id is

available. Brand records can contain stale or malformed derived URLs; resolve the ID immediately before generation instead.

In short: IDs at rest, URLs at the model boundary, IDs downstream. Do not pass a client-local path to the hosted connector.

Storage

| storage | Behavior | |---|---| | transient | Default. Temporary URL, not saved, expires. Best for one-off images. | | asset | Persistent — no expiry, returns an asset_id. Use when you want to reuse the image, e.g. attach it to a post via the simplified-social skill (pass the asset_id in media). | | default | Saved to your AiImageArt gallery. |

Examples

Text-to-image (default, transient):

{ "model": "google.gemini-3.1-flash-image-preview", "capability": "prompt", "storage": "transient",
  "parameters": { "prompt": "A white ceramic coffee cup on a clean white background", "aspect_ratio": "1:1" } }

Keep it to reuse / post to social (asset):

{ "model": "google.gemini-3.1-flash-image-preview", "capability": "prompt", "storage": "asset",
  "parameters": { "prompt": "product hero shot of sneakers", "aspect_ratio": "4:5" } }

Edit / reference-guided — the reference field name is model-specific; take it from api_getModelFields (here input_image for a Flux Kontext model, not a guessed name):

{ "model": "flux.flux-kontext-pro", "capability": "reference_image", "storage": "asset",
  "parameters": { "prompt": "put this logo on a t-shirt", "input_image": "" } }

Model selection

Choose for the requested outcome, not provider popularity. These routes are maintained defaults, but model availability, capabilities, parameters, and credits can change; api_getModelFields(type: "image") remains authoritative.

| User need | Preferred model | Why / tradeoff | |---|---|---| | Normal social image, product shot, illustration, character continuity, or general edit | google.gemini-3.1-flash-image-preview | Quality-first default. Strong all-around prompt following and reference fidelity. Do not interpret “Flash” as the cheapest option. | | Complex professional design, dense typography/layout, menu, invitation, high-fidelity product mockup, factual visualization, or explicit 4K | google.gemini-3-pro-image-preview | Premium quality and instruction handling; slower and typically costs more. Use only when the request benefits from it. | | Budget-sensitive generation or explicit GPT Image request | openai.imgen-2 | The catalog's credits_per_image is a baseline, not the final charge. Cost varies with size, quality, and count. Use the live API field quality: "auto" (the operational “effort auto” setting) unless the user requests a different quality. It uses size rather than aspect_ratio. | | Short headline or typography-first poster/banner | ideogram.ideogram-v3-turbo | Specialized text rendering. Prefer Gemini Pro when the design also requires a dense or complex professional layout. | | Targeted edit with a single source image | google.gemini-3.1-flash-image-preview; flux.flux-kontext-pro when explicitly requested or better suited by live metadata | Default to Gemini for fidelity. Flux Kontext is a specialized alternative; inspect its input_image contract first. | | Many reference images or exact reference limits | Best compatible model returned live | Filter by multiple_images and the discovered reference limit. Never assume every model accepts the same number or field name. | | User names Flux, Seedream, Qwen, Stable Diffusion, or another available model | The requested model, if compatible | Respect an explicit preference. Otherwise do not automatically route to an unvalidated specialist merely because it is available or cheaper. |

Routing rules

1. Infer the hard constraints: capability, reference count, aspect ratio/size,

resolution, text/layout complexity, budget, and any explicit provider choice.

1. Filter the live catalog by those constraints.
2. Use Gemini 3.1 Flash when no stronger constraint applies. Upgrade to Gemini 3 Pro

only for the professional-design cases above. Consider GPT Image 2 when minimizing credits is explicit or as the first fallback, but state that its live catalog rate is only a baseline and the final charge varies with size, quality, and count. Default to size: "auto", quality: "auto", and count: 1 unless the request requires different values.

1. For a typography-first graphic, choose Ideogram Turbo; for dense layout or 4K,

choose Gemini Pro instead.

1. Never silently change models after an error. Report the failure and proposed

fallback with its live credit cost, then regenerate only when the user's existing intent clearly authorizes the additional spend.

When a request is ambiguous and the choice materially changes cost or output, offer the most relevant two choices, leading with the recommended model. Do not dump the entire catalog on the user.

Response

The response shape depends on storage:

• transient (default) — result is a list of URL strings:

``json { "status": "SUCCESS", "detail": { "result": ["https://replicate.delivery/…/out-0.webp"], "transient": true } } ` Read detail.result[0] (a URL string). No asset_id` — the URL is temporary.

• asset — result is a list of objects with a reusable id:

``json { "status": "SUCCESS", "detail": { "result": [{ "asset_id": "", "url": "https://…/image.webp?Expires=…" }], "transient": false, "storage": "asset" } } ` Read detail.result[0].url (the image; **signed URL — expires**) and detail.result[0].asset_id (permanent — hand off to simplified-social's media`).

Output format varies by model and provider. Inspect the returned asset or response metadata instead of assuming WebP; for example, Gemini may return JPEG.

Presenting the result

Never embed the returned image URL with Markdown image syntax (``), and never do anything that makes the client fetch/render the image inline. Always present the result as a plain URL or a Markdown link the user can click:

• ✅ Here's your image: https://…/out-0.webp
• ✅ [View generated image](https://…/out-0.webp)
• ❌ ``

Reasons: these URLs are signed and expire, inline rendering fails or shows a broken image, and clients like Codex otherwise try to display the asset instead of handing the user a usable link — poor UX. When storage:"asset", also surface the permanent asset_id (as text) so it can be reused with simplified-social.

Gotchas

• Discover before generating. Call api_getModelFields to confirm the model id

and parameters schema — it eliminates 400 errors on invalid/missing keys and prevents routing from stale model or credit assumptions.

• Resolve asset-backed references before generating. Use api_getAsset, require

status: 4, and pass its current file_url when the live model field expects a URL. Keep the source asset_id for future runs.

• Generation spends credits. If the request is ambiguous, restate what you'll

generate and confirm once. If it's explicit, proceed.

• Do not overstate GPT Image 2 pricing. Treat credits_per_image as baseline

metadata. Final usage varies with size, quality, and count. The live API calls the effort control quality; use quality: "auto" for the usual “effort auto” behavior and never describe the baseline as the guaranteed charge.

• 429 = AI credits exhausted; tell the user plainly and don't retry.
• On error, report it; don't silently retry.

Example prompts to try

• "A minimalist product photo of a white ceramic coffee cup on a clean white background, soft studio lighting"
• "A vibrant 3D render of a friendly robot mascot, pastel colors, studio lighting, 1:1"
• "A cinematic 16:9 landscape of snowy mountains at golden hour"
• "A flat vector app icon of a paper plane, rounded corners, blue gradient"
• "A bold quote card that says 'Ship it' in modern type" (use ideogram.ideogram-v3-turbo for crisp text)

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: celeryhq
• Source: celeryhq/simplified-ai
• License: MIT

Install and usage instructions live in the source repository linked above.