Generating and editing with a prompt

Most AI image work in Madoo happens in nodes built for one job — a product on a surface, a garment on a model, a background swapped — which carry their own expertise in their prompt. The nodes on this page are the general ones: they do what the words say. Text to Image makes a picture from a description. Image Transform changes a picture, or combines several, following an instruction. Image Fusion composes several products into one scene with a layout. And three vision nodes go the other way: they look at pictures and answer in words — a description, a category, a prompt.

Use them when no specialised node fits: a campaign illustration, a mood picture for a newsletter, a vector icon, a scene with three products, a photo moved to a new setting, a listing text written from the product itself. The common machinery — models, settings, presets, variants, price — is in How AI image steps work; this page is about what each node does and how to talk to it.

Text to Image

Text to Image (ai/text_to_image) makes a picture from a prompt, with an optional negative_prompt — both can come from inputs, so a prompt written by an earlier step (an AI, a template, a spreadsheet row) drives the picture.

Setting	What it does
aspect_ratio (preset)	1_1, 4_5, 3_4, 16_9, 9_16: the frame. A model with frames of its own (MAI Image, Nano Banana, Kling) gets the nearest one it makes — 4_5 on MAI Image becomes 3:4 — and GPT Image the nearest image_size; set the model’s own aspect_ratio for a frame outside the five, such as 3:2
image_style (preset)	31 styles, from photorealistic, product_photo and cinematic to watercolor, isometric, pixel_art, pop_art
mood (preset)	16 atmospheres: fresh, luxury, cozy, dramatic, minimalist, summer…
quality (preset)	draft, standard, high, ultra: steps and guidance for the models that use them
variants	1–4 proposals in one step
seed	the same seed with the same settings repeats a picture on the models that honour it

The presets add their words after yours: “your prompt, in the style, quality, mood atmosphere, highly detailed”. They are a quick way to keep a series consistent; your prompt remains the instruction.

Which model. The default, Nano Banana 2, is a good all-rounder at 5 credits and follows long instructions well. Other models, chosen on the node, cover other needs:

Need	Model	Credits
Many quick drafts	FLUX Schnell	1 per megapixel (a 16:9 picture is 2)
Sharp pictures at any frame, text and QR codes included, at the lowest price	Muse Image (Meta), and Muse Image Edit for edits from up to 10 pictures	1 per image, about 2.5 megapixels
Photorealistic pictures, about 1 megapixel	MAI Image 2.5 (Microsoft); MAI Image 2.5 Pro for the finest detail	3–4; Pro 9–11
The update of the default — better quality, prompt adherence and text — with frames from 8:1 to 1:8, a thinking level and web search for current facts	Nano Banana 2.1 (Google), and Nano Banana 2.1 Edit for edits from up to 14 pictures	3 per image at 1K, 4 at 2K, 9 at 4K; thinking_level: high adds 2, web search 1 to 2; the edit up to 2 more for the pictures it reads
Wide or tall frames — 2:1 to 1:2, phone-screen 19.5:9 and 20:9 — at 1K or 2K	Grok Imagine Image 2.0 (xAI), and Grok Imagine Image 2.0 Edit for edits from up to 5 pictures	4 per image at 1K, 5 at 2K (3 and 4 at low quality); the edit adds 1 to 4 for the pictures it reads
A vector file — an icon, a logo sketch, an illustration for print	Recraft V3 with style: vector_illustration	6 (3 × 2), delivered as SVG
Text inside the picture — posters, logos, packaging, signs	Ideogram v4.5, up to 2560 × 1440 at the same price; Ideogram v4.5 Edit for edits guided by up to 4 reference pictures	2 low, 4 medium, 14 high per image; the edit also 1 at very low
The largest pictures	Nano Banana 2.1 or Nano Banana 2 at resolution: 4K, FLUX.2 by megapixel	9 and 10; from 2
Text in the picture, transparent backgrounds, sizes up to 4K	GPT Image 2.5 Flare (Sunburst for the finest detail, slower)	3 at high quality, 1 at low, 7 at 4K
The top quality of a family	Nano Banana Pro, GPT Image 2, Recraft V4 Pro	10–15

The whole list, with each model’s settings, is on the node’s model picker and in GET /api/v1/node-types/ai/text_to_image/models.

Image Transform

Image Transform (ai/image_transform) takes 1 to 14 pictures on images (images_0, images_1, …) and a prompt, and returns one picture. The prompt is sent as you write it: no template, no added words. It is the node for everything the specialised ones do not name — move a product to another setting, change the season of a photo, put two products together, turn a sketch into a render, apply the look of one picture to another.

• The first picture is the main one; refer to the others by their order: “the vase from the second image”.
• aspect_ratio defaults to match_input, the proportions of the first picture; a ratio can be chosen instead.
• strength and num_inference_steps (advanced) matter only to the models that use them.
• The default model reads up to 14 pictures. A model that reads only one stops the step before any charge when given more (How AI image steps work).

The model changes what you ask and keeps the rest: in the lab of How AI image steps work a museum teapot moved to a marble counter kept its shape and painting exactly. Say what must not change — “keep the product exactly as it is: its shape, its label, its proportions” — and it holds better.

Image Fusion

Image Fusion (ai/image_fusion) composes two to five products (product_image_1 … product_image_5) into one scene, from settings rather than a free prompt: a layout (flatlay, lifestyle, grid, diagonal, scattered, hero, stacked, circular, symmetrical, cascading), a mood, an aspect_ratio, a background described in words and an optional custom_prompt added at the end. It is the quick way to make a “shop the look”, a bundle, a gift set; for more than five pictures, or a composition you want to describe yourself, use Image Transform.

Reading pictures: analysis, classification, prompts

Three vision nodes read pictures and answer in text. They run on a language model — GPT-6 Luna by default, with Claude and Gemini among the others — and are priced by the words read and written: usually a few hundredths of a credit per picture.

Node	Gives	Good for
Image Analysis (ai/image_analysis)	result in prose, or json when response_format is json_object	Listing texts, alt texts, attributes (colour, material, style), a check of what is in a photo. Up to 20 pictures in one question; a system_prompt sets the role
Image Classifier (ai/image_classifier)	category from a list you give (categories, at least two, comma separated) and confidence (high, medium, low)	Routing photos: product type, “usable or not”, which template to use. Add an other category for what fits none; if the model answers something outside your list the step fails with CATEGORY_NOT_IN_LIST rather than guess, and a doubtful case comes back with low confidence — see Reading pictures
Image to Prompt (ai/image_to_prompt)	prompt and negative_prompt describing the picture	Making “one more like this”: connect its outputs to Text to Image; instructions steer what the description stresses

The JSON of Image Analysis goes straight into data steps: a field for a template, a condition, a row of a catalogue. Ask for the keys you need and say how long each value may be.

Writing the instruction

• Say what to keep before what to change. Models change more than asked when nothing tells them what is fixed.
• Name pictures by their order, and say what each one is for: “the first image is the product, the second only a reference for the light”.
• Describe like a photographer: the subject, the surface, the light and where it comes from, the framing, the background. Adjectives of quality (“stunning”, “8k”) add little.
• Put a text you need in the picture in quotes, and prefer a model good at text — or, better, add the text afterwards with Text on Image, in your own fonts.
• Ask for variants rather than rewriting the prompt: two proposals cost two pictures and differ more than two prompts usually do.

The lab

The prompt lab uses four photos of the museum-shop collection: a blue and white teapot, a red teapot, a vase with birds and a Delft shoe.

Step	What it gave	Credits
Image Fusion, symmetrical, luxury, a green velvet shelf	The three objects on a museum-shop shelf, each kept as it is	5
Image Analysis, JSON	object, colours, material, style, condition, alt_text (JSON)	0.03
Image Classifier, teapot / vase / shoe / chair	shoe, high	0.01
Image to Prompt → Text to Image	A detailed prompt, and from it a new red teapot in the same spirit	0.03 + 5
Text to Image, Recraft V3 vector_illustration	A navy line drawing of a teapot, delivered as SVG	6
Run	estimate 18, charged	16.07

The vision steps are estimated generously — the estimate cannot know how long an answer will be — and charged by what was actually read and written, so the run cost less than its estimate. Run it again and the words change: the JSON keys stay, their values are the model’s.

node madoo-install-example.mjs lab-prompt-generation --api <your Madoo API URL> --key <key.json> --run --save generation

Limits

• Image Transform has no variants; run the step twice, or use a specialised node that has them.
• Image Fusion takes at most five products and no free prompt of its own beyond custom_prompt.
• A vector result comes only from Recraft’s vector styles; other models make pixels.
• Vision answers are text from a model: check a classification with low confidence, and validate JSON before relying on a field (Schemas).

Photos: “Teapot”, John Bartlam, about 1765–69; “Teapot”, George E. Ohr, about 1897–1900; “Vase”, Greenwood Pottery Company, about 1884–93; “Shoe”, 1697, Delftware — The Metropolitan Museum of Art, New York (Open Access, public domain). The generated pictures were made by the models named, for this lab.