Generate Image

The Generate Image node turns a text prompt into an image. It is the main text-to-image node in Nodaro. You write what you want to see and choose one of the image models. The node returns a picture you can send to any other node, for example to animate it with Generate Video.

When to use it

• You need a still image from a description: a storyboard frame, a product shot, a thumbnail, a character portrait.
• You want the first frame of a video. Wire the result into the start frame of Generate Video.
• You want the same character, location or product in many images. Wire the asset nodes into the Assets input.
• You want to change one part of a result. Paint a mask and run the node again, without adding another node.

Quick start

Example: a prompt and two pickers

This run wires three nodes into Generate Image: a Text node with the prompt, a Mood picker set to Serene, and a Lighting picker set to Golden Hour with the Back / Rim direction. The model is GPT Image 2 at 1K in 16:9.

The model received the typed prompt first, then the words of each picker in Full mode:

Mood describes a face and a posture, so it suits pictures of people more than landscapes. For the feel of a place, such as mist or rain, use Atmosphere instead. The same image became the start frame of the clips in Generate Video.

Inputs

Generate Image has seven inputs on its left edge. Each input accepts only the kinds of nodes listed, and the editor rejects a connection that does not fit. Click an input to see what is connected to it, jump to a connected node, disconnect it, or add a new compatible node.

The output, image, is the URL of the generated picture. It can feed any number of nodes at once.

Typed prompts and connected prompts

A prompt connected to Prompt is added after the prompt you type in the node. If the typed prompt names the connected node as a variable, such as {Subject}, the connected text appears only where the variable is. To stop adding connected prompts, turn off Inject Prompt under Prompt Injection in the settings panel.

A connection to Prompt turns grey and dashed only when the node does not use the connected text. Point at a grey connection to read the reason:

• Inject Prompt is off, and the typed prompt does not name the connected node as a variable. The connection says "Not used: Inject Prompt is off".
• Several text nodes are connected to Prompt. The node uses the one connected last, and the earlier connections turn grey. They say "Not used: a text node connected after this one takes its place". A text node that the typed prompt names as a variable is always used.

Settings input

The Settings input takes the Aspect Ratio and Provider nodes. When the node runs, each connected node replaces one of the node's own settings:

• Several nodes of one kind. When two nodes of the same kind are connected, such as two Provider nodes, the one connected last is used.
• A node set to several models. When you choose more than one model with Add another model, the node runs every model you chose. A connected Provider node replaces that list, so the node runs only the Provider node's model. While the Provider node is connected, the settings panel shows its model instead of the list.
• A model the node cannot run. A Provider node set to a video model is refused. The model's chip on the node turns red, and Run is disabled with the reason, for example "This node can't run VEO 3.1 Quality. Choose one of its models in the Provider node." A run that reaches the node, for example from a workflow run, stops there before anything is charged, and the error names the Provider node.
• What the node shows. Chips at the top left of the node list the values it runs with, for example 9:16 and Nano Banana. When the node shows a result, the chips hide while you point at the result. In the strip under the node, the model and the ratio cannot be changed. In the settings panel, Provider and Aspect Ratio show the connected node and the value the node runs with, in place of their controls.
• The price. The Run button, the workflow total and the credits held when the run starts all use the connected model. A node set to several models is priced for the connected model only. While a setting is connected, a note under Run says "Price may change with connected settings".
• No Duration or Motion. A still image has no length, so the Settings input takes no Duration node. It takes no Motion node either, because Motion works only on video nodes. The canvas refuses both connections.

Settings

Aspect Ratio and Provider nodes connected to the Settings input replace Aspect Ratio and Provider when the node runs. See Settings input.

Models

Generate Image can run every image model below. Click a model for its settings, credit prices and prompt tips.

Which model to choose

• GPT Image 2 — the default. Strong at characters and identity, and at exact copy, labels and non-Latin text in the image.
• GPT Image 2.6 Flare — choose it for speed: posters, product shots and quick iteration, with strong layouts.
• GPT Image 2.6 Sunburst — choose it when every detail counts. It gives the finest detail and the strongest consistency across edits. It also draws the wide and in-between ratios that GPT Image 2 cannot draw, such as 21:9, 4:5 and 5:4.
• GPT Image 2.5 Flare — fast and high quality. Use it to iterate and for high-volume work such as social posts and ad variants.
• GPT Image 2.5 Sunburst — slower and more precise. Use it for the final render of brand-sensitive work: packaging, diagrams, product retouching.
• Nano Banana 2.1 and Z-Image — low cost and fast. Use them for drafts and storyboards. Nano Banana 2.1 takes up to 10 reference images and renders up to 4K.
• Seedream 5 Flash — fast and low cost, at 1K or 2K for the same price.
• Seedream 5 Pro — strong instruction following at a low price.
• Seedream 5.5 — pictures that should look photographed, with natural light, materials and skin, and posters with dense small text. It draws a transparent background in one step. See Seedream 5.5.
• MiniMax H3 Image — low cost, with readable text inside the picture, up to about 16 MP at 4K. See MiniMax H3 Image.
• Nano Banana Pro — detail, diagrams and complex compositions.

If you are unsure, draft on GPT Image 2.6 Flare and finish on GPT Image 2.6 Sunburst. For a comparison of every model, read Choosing a model.

GPT Image 2.6 quality and price

GPT Image 2.6 Flare and GPT Image 2.6 Sunburst cost the same. Both draw 15 aspect ratios, from 3:1 to 1:3, at 1K, 2K or 4K.

• Six qualities. Auto, Low, Medium (the default), High, Extra High and Max. Each quality has its own price at each resolution, and the Run button shows it before the run.
• Draft low, finish higher. Draft at Low, and finish at Medium or above.
• Auto. The model chooses the quality. The price shown before the run is the most the run can cost, the Max price. The run is then charged for the quality the model used.
• Reference images. Both models take up to 16 input images. Each image adds to the price, so a run with fewer images costs less.

MiniMax H3 Image

MiniMax H3 Image draws detailed pictures with readable text inside them, up to about 16 MP at 4K.

• Sizes. 1K, 2K (the default) or 4K. 1K and 2K cost the same, and 4K costs more.
• Shapes. Ten aspect ratios plus Auto: 1:1, 16:9, 9:16, 4:3, 3:4, 3:2, 2:3, 5:4, 4:5 and 21:9. With no photo connected, Auto lets the model choose the shape from the prompt.
• Reference images. Connect up to 16 reference images, and the node runs the MiniMax H3 Image edit model. The edit is priced by size and by each image, and Auto keeps the first image's shape. The Run button shows the exact price before the run.
• Other settings. It takes a Seed, writes PNG or JPEG, and accepts prompts up to 5,000 characters.

Seedream 5.5

Seedream 5.5 is a model of its own, not a newer Seedream 5 Lite, Pro or Flash. Choose it for pictures that should look photographed, with natural light, materials and skin. It also keeps dense small text readable, for example on posters and packaging.

• Transparent background. Seedream 5.5 draws a transparent background in one step, at every size, as a PNG. With JPEG the background becomes Auto, and the node says so.
• Sizes. 1K, 1.5K and 2K (the default). 1K and 1.5K cost the same, so 1.5K is the sharper choice at that price.
• Edits. With reference images or a base image, the node runs the Seedream 5.5 edit model. The edit changes what you ask, round after round, and keeps the rest of the picture as it was. It takes up to 10 input images, and each image adds to the price. The Run button shows the exact price before the run.
• Prompt length. Prompts need 5 to 5,000 characters. A shorter prompt is refused before the run, and nothing is charged.
• Auto with no photo. With no photo connected, Auto is sent as 1:1, because the text-to-image model has no automatic shape. With a photo, choose a ratio when the shape matters.

Formats: PNG, JPEG, WebP, GIF and SVG

GPT Image 2.6 Flare and Sunburst write five kinds of file. Choose one under Format. Sunburst costs the same as Flare in every format.

• The price before the run. The Run button shows the price of the format you chose. At quality Auto, it shows the most the run can cost, the Max price of that format, and the run is charged for what the model used.
• Reference images. With reference images or a base image, a GIF or an SVG adds the same amount per image as a PNG does.
• Background. A transparent background is kept in PNG and WebP. With JPEG or GIF it becomes Auto, and an opaque background is kept. An SVG takes no background setting.
• Masks. A GIF or an SVG cannot be inpainted. A run with a mask and either format is refused before anything is charged. Choose PNG, JPEG or WebP to edit a masked area.
• Safe SVG files. Every SVG is checked and cleaned before it is saved. Scripts, event handlers, embedded pages and links to other files are removed. An SVG that cannot be read safely is not saved: the run fails, and you are not charged. On the free tier, the drawing carries a small white mark with a dark outline.
• The result. The file keeps its own extension, such as .jpg, .webp, .gif or .svg. A GIF plays on the canvas and in the Library, and an SVG is drawn there. A GIF or an SVG also gets a PNG still beside it: the first frame of a GIF, or the drawn SVG.
• Wired into another image node. Image models take still pictures, so a GIF reaches the next model as its first frame, and an SVG as a drawn PNG.

Use variables in the prompt

You can put the value of another node in the prompt by writing its label in curly braces. {Mood} becomes whatever the Mood picker is set to, and {Maya} becomes the text of a node labeled Maya.

• A default value. {person || man} uses the connected Person picker when there is one, and the word "man" when there is not.
• Highlighting. In the prompt editor, a variable is cyan when a matching node exists and amber when nothing provides it yet. Amber is only a warning: a variable with a default still works.
• A missing node stops the run. If a variable names a node that does not exist anywhere in the workflow, the run is refused before any credits are spent, so a typo never produces a wrong image.

Read more in Prompt variables.

Edit a result in place

Once the node has a result, you can change it without adding another node.

Change one area with a mask

1. Open the settings panel and scroll to the Inpainting Mask painter. Click Edit Mask.
2. Paint over the area to change. White means "change this", black means "keep this".
3. Write a prompt that describes the change, and run the node again.

Only the painted area is regenerated. Everything outside the mask stays pixel-identical, on every model. A mask works only with a still format: on GPT Image 2.6, set Format to PNG, JPEG or WebP, because a GIF or an SVG cannot be inpainted.

Refine the whole image

Click Refine from this result to use the current result as the starting point for a new run over the whole frame. Use it for changes such as "warmer colors" or "more cinematic". On models that support it, Strength controls how far the new image moves away from the current one.

Edit regions with Grok Imagine 2

With the Grok Imagine 2 model, the settings panel adds Refine Regions:

1. Click Detect regions. This is free. The node finds named regions, such as sky or person, and outlines each one on the result when you point at it.
2. Tick the regions to change, write what should change, and click Apply. Leave every region unticked to edit the whole image.
3. The edit appears as a new version in the result strip, and you can detect and edit its regions again.

A masked edit, a refine and a region edit each cost the same as a new generation on that model. Detecting regions is free.

Keep a character consistent

Wire a Character Asset node into the Assets input, or mention the character in the prompt with @ and its number, for example @maya:1 walks through the market. Nodaro sends the character's approved pictures and description with the prompt, on every model that accepts references.

You can also name an Upload Image node and mention it: a node labeled "Town" becomes @town:1, and @town:1:background says what to take from it. Read Reference roles for every role you can use, and Consistent characters for the full method.

When a model blocks the prompt

A model's safety filter sometimes blocks a harmless prompt.

• Your credits come back. A blocked generation is always refunded.
• Some models retry once. GPT Image 2 and the GPT Image 2.5 and GPT Image 2.6 models are retried once automatically, at no extra cost.
• A one-click alternative. When the model has a recommended fallback, the node shows the block in amber with a button such as Try on Nano Banana Pro. Nodaro never switches models on its own.
• Copyright and likeness blocks are final. A match against protected characters or a real person's face is not retried.

Tips

• Draft cheap, finish sharp. Iterate at 1K on a draft model, then switch the model or the resolution for the final image.
• Put the look in pickers, not in the prompt. Pickers such as Lighting and Lens add tested wording, and you can change them without rewriting the prompt.
• Choose a model for text. For a sign, a label or a poster with words, use GPT Image 2 or GPT Image 2.6 Flare.
• Keep several image nodes on one model. Connect one Provider node to the Settings input of each of them, and switch the model in one place. Do the same with an Aspect Ratio node to give them one shape.

Troubleshooting

Run is disabled and a chip on the node is red. The Provider node on the Settings input holds a model this node cannot run, such as a video model. A new Provider node starts on the Video category. Set its Category to Image and choose an image model. A workflow run stops at this node before anything is charged, and its error names the Provider node.

The image's shape differs from the connected Aspect Ratio node. The model does not make that ratio, so the node runs the nearest one. The chips on the node show the ratio it runs with. See Settings input.

The node is set to several models, but only one runs. A Provider node on the Settings input replaces the node's list of models. Disconnect the Provider node to run every model again.

From the API

Every setting above is also available to code and to AI assistants. POST /v1/generate-image accepts a direction object of picker ids — for example { "shotSize": "wide-shot", "timeOfDay": "golden-hour" } — and Nodaro writes the matching wording into the prompt. See Run a single node and the MCP tools.

File format

On GPT Image 2.6 Flare and Sunburst, outputFormat sets the file: png (the default), jpeg, webp, gif or svg. With gif, gifFrames (2 to 16, 9 by default) and gifFrameDurationMs (20 to 2,000, 200 by default) set the frames and their speed. With svg, svgLevel is standard (the default), detailed or max. A GIF request that names an aspect ratio or a resolution runs square at 512 × 512, and the answer says which settings it ignored. In MCP, the same fields are output_format, gif_frames, gif_frame_duration_ms and svg_level. A request with a mask and gif or svg answers 400 before anything is charged.

Your own palettes and moodboards

The direction object also accepts your own saved palettes and moodboards. "colorLook": "palette:<id>" names the palette's colours in the prompt. "style": "moodboard:<id>" adds the moodboard's description, and sends its images as style references after your own, up to the model's reference limit. A malformed id answers 400, and an id that is not yours answers 404. A call that also applies a preset (presetId) cannot use them, and answers 400. See Palettes and moodboards.

Several images in one call

A call from the API, the SDK or MCP can ask for several images at once:

• count asks for 1 to 4 images of one model. Each image is its own job.
• models sends the same request to 2 to 4 image models, instead of provider. With count, the count applies per model, up to 8 images in total.

Each model's settings are fitted to that model, as on a single call. The batch is charged as a whole or not at all, and each image costs what a single call for it costs. With a seed, image number i gets seed + i. A preset (presetId) works with both, and with models it is applied once per model. The answer lists every job id, grouped by model. On the canvas, one Generate Image node still makes one image. See Several images in one call.