SupportGet started

Batch

Generate AI images from up to 20,000 prompts in a table or CSV, see the price before you run, and keep every earlier run. Requires the Max plan.

Last verified

Overview

Batch Banana lets you generate many AI images at once from a table or CSV, up to 20,000 rows. Add one prompt per row, choose each row's model, style, aspect ratio and image count, then select Run all. You see the price before anything is charged. Results appear in a new column for each run, so earlier images stay available. Batch requires the Max plan.

Create game assets, product variations or a set of social posts in one run. Adjust the prompts or settings and run again; earlier results stay available.

Start a batch

Open Batch in the sidebar and choose New batch. A new batch opens in the project you have open, with a box for your prompts: paste or type them, one per line, and choose Add to batch. Each line becomes a row. Open the menu on the batch's name to rename or delete it.

Every row uses the batch's Model, Style, References, Ratio and Images unless you set its own. To change the batch's setting, open the column's header; every row that uses the batch's setting follows. To give one row its own setting, click its cell. A row's own setting shows in full color, and one it takes from the batch shows grey. The × beside a row's own setting puts it back on the batch's. A style adds its example images to the row as references, so the row needs a model that takes reference images.

Use references

Your references keep a look, a character or a product the same across every row. Open the References header to pick the references every row uses. A new batch starts with the references marked Auto-attach in its project and workspace, and you can take any of them off there.

Each References cell shows a small pile of pictures for every reference the row sends. The batch's references show faded, and a row's own show in full color. To add references to one row, click its cell and check them under This row. A row's references add to the batch's rather than replacing them, so the batch can hold the style while each row adds the product or character it shows. The batch's own are listed under From the batch; uncheck one there to leave it off this row only. The list holds the references of the batch's project and the workspace.

To send only some images of a reference on one row, open the row's References cell and click the count beside the reference, All until you change it. Its pictures open below it; click one to leave it out of this row, and the cell then shows, for example, "2/5" beside the pile. When a row's references hold more images than its model takes, each reference and the style send their first images in turn until the model is full, and the list says how many go. If a reference is deleted or moved to another project after you add it, the batch still works and you can keep editing it, but rows that use the reference fail with the reason when you run them, at no cost. It shows as Deleted reference under No longer available in the References list; uncheck it there to take it off.

Edit rows

Use Add row for one more row. To change many rows at once, select them and use the bar that appears; a row's number turns into its checkbox when you point at it. Once every row on screen is selected, Select all in the bar adds the rest of the rows that match the current filter.

Rows load as you scroll, so a long batch opens quickly. Search prompts finds rows by their prompt across the whole batch, up to 50 at a time, best match first. Columns hides the columns you don't use.

Import a CSV

For a long list, choose Import CSV and drop in a .csv or .tsv file, up to 25 MB and 20,000 rows. Before anything is added, you see the columns that were read, the price to run the rows and every row with a problem, by its line in the file, with how to fix it.

The file needs a prompt column. These columns are read, and any others are ignored:

ColumnHeader names it acceptsWhat goes in it
Promptprompt, text or descriptionThe prompt. Required.
Modelmodel or model nameAn image model's name or ID, in any case
Aspect ratioaspect_ratio, aspect ratio, ar or ratioA ratio such as 3:4 or 16:9
Images per promptimages, num images or countA whole number from 1 to 8
Stylestyle or style slugA style's slug, the last part of its page address in the style library
Referencesreferences, reference or reference imagesReference images separated by spaces: public https image links or file IDs. Each link is added to your library once.
Reference setsreference_sets, reference sets, reference set or setsNames of your references, separated by semicolons, such as Brand; Mia. They add to the batch's references.
Optionsoptions, or a column named after one option, such as resolutionModel options as JSON, such as {"resolution":"2K"}

Headers can be in any case and can have spaces. An empty cell uses the batch's default.

A value the batch can't use makes its row a problem: a model that doesn't exist, a video model, a ratio that isn't offered, a style the row's model can't use, or a reference name that matches none of the project's or workspace's references, or matches two of them. Fix the file and drop it in again, or skip those rows: the import button then reads, for example, Skip 2 and import 98 rows. Blank lines are skipped.

The rows import in the background, in the order of the file. The batch shows Importing with progress, and you can close the dialog. Run all and Import CSV wait until the import finishes. If an import stops partway, the rows it wrote stay in the batch.

Choose Download template in the import dialog for an example CSV with every column.

A batch holds up to 20,000 rows.

Run it

Choose Run all, or select rows and choose Run in the bar. A confirm step shows the price for every row, in total and by model. Rows that can't run as they are, such as a row with an empty prompt, are listed there; they are marked failed with the reason and cost nothing. Choose Run to start.

The run shows Preparing while every row is priced again. Then the credits for the whole run are charged once. If the price changed after you confirmed, nothing is charged and the run stops, so you can run again and see the new price. If the workspace doesn't have enough credits, the run also stops with nothing charged. When a run stops, or you cancel it, before the charge, the table goes back to how it was before the run. When a run stops on its own, a note above the table says why until you close it or start another run. If it stops after the charge, rows that had not started are refunded, rows already running finish and stay charged, and a row that fails after the stop is refunded on its own.

While a run goes, the top of the batch counts the rows done, its column header does too, and Cancel run stops it. The batch's settings can't change until the run ends. Anywhere else in the workspace, a progress notice shows each run with Cancel and tells you how the run ended. You can also cancel from the run's menu. Images that already started still finish. When the run ends, the credits for rows that failed or never ran come back as one refund. Each row sends the reference images it was priced with, even if you change a reference while the run goes. A row whose reference, or one of its images, you delete before it starts fails, and its credits come back with that refund.

If the model is busy, a row waits and tries again on its own, at no extra cost. A row the model refuses for its content rules is not tried again.

Only one run can be active at a time in each batch.

Each run is its own column, newest first, right after Images, so each prompt and its latest images are in view together. The Options column, which a CSV import or an AI app fills in, comes after the runs when a row has options. Above the table, All, Completed, Failed and Running show the rows by their result in the newest run, with a count for each. Open a run's menu to Browse images, Download images or Retry failed rows. Retrying runs only the rows that failed in that column. The menu also shows only an older run's completed or failed rows; the filter then names that run, and its × clears it. A failed row says why it failed and what to try. To retry failed rows in every column at once, use Retry all failed rows in the menu on the batch's name. Retrying shows the price first, like a run.

Download results

The menu on the batch's name has two downloads:

  • Download results (CSV) writes one line per image, with its row number, run, prompt, style, model, status, any error, file ID and a link to the original file. A row that made no image in a run still gets a line, so every row is accounted for.
  • Download all images downloads a ZIP of every image from every run. A run's own menu downloads only that run. A batch past 1,000 rows comes as several ZIP files of up to 1,000 rows each, with one button per file. The links work for an hour. Each file name starts with its row number and run number.

What a batch costs

The table shows credits for one image per prompt. Each row costs the credits per image for its model multiplied by its image count. Some models charge more for each reference image, and a style's example images and a reference's images count as references. Add the row costs to get the total for a run; the confirm step does this for you. For model settings and prices outside Batch, see Credit costs.

ModelCredits per image100 prompts20,000 prompts
GPT Image 2.6 Sunburst303,000600,000
GPT Image 2.6 Flare303,000600,000
Nano Banana 2.1101,000200,000
Nano Banana 2 Lite8800160,000
Nano Banana Pro353,500700,000
Flux 3 Image6600120,000
Seedream 5 Pro111,100220,000
Seedream 5 Lite101,000200,000
Krea 3141,400280,000
Recraft V4.1 Pro595,9001,180,000
Recraft V4.1 Pro SVG595,9001,180,000
Arrow 2595,9001,180,000
Ideogram 4.5141,400280,000
Recraft V4.1101,000200,000
Recraft V4.1 SVG101,000200,000
Flux 3 Image Klein110020,000

From an AI app or the API

Connected AI apps can make and run batches too. Connect the MCP server to Claude, ChatGPT, Cursor or Codex and ask for many images in the chat. The app gets a price first, creates the batch, runs it and checks on it, and in supported apps you can watch the images arrive and cancel the run. See Ask your AI app. Your own code can do the same through the API, which our team turns on per workspace; see API access.

Tips

  • To run one row, open its menu (the dots at the end of the row) and choose Run this row.
  • Write three rows and run them on an inexpensive model from the table, such as Nano Banana 2 Lite, before you fill in the rest. Fixing the prompt pattern is cheaper at three rows than at thirty.
  • Keep style instructions consistent across rows, changing the subject for each asset, or pick one style for every row.
  • Use runs as versions. Refine your prompts or settings and rerun when you are ready to develop the next direction.
  • Download results as CSV before a big edit, so you have a copy of every prompt and its images.