Lists and batching

A list is a set of items that a workflow processes together: prompts, images, videos, audio files or the rows of a table. Fan-out runs a node once for every item, so ten prompts become ten images. Fan-in turns many results back into one, for example by stitching clips into one video or by letting an AI judge choose the best image.

How it works

The connection mode decides whether a list is processed item by item or all at once. A connection in Each mode runs the next node once per item. A connection in Bundle mode hands the whole list over in one run. Nodes and connections explains every mode.

Where lists come from

Fan-out: one run per item

When a connection in Each mode carries a list into a node, the node runs once for every item and keeps every result. Connections from a List and from Split Text use Each by default.

• Items replace the typed prompt. Each item becomes the prompt of its run. It replaces the prompt typed on the node instead of being added to it. For a row whose prompt is empty, the typed prompt is used.
• Items run one after another. A node processes its items in order, from the first row to the last. More items mean a longer run.
• Chains stay paired. When the next node also receives the results through an Each connection, it runs once per result, in the same order. Five prompts give five images and then five videos, not twenty-five.
• Repeat multiplies. With Repeat set to 2 on a node, every item runs twice.
• Credits multiply. The node costs its price times the number of items, times the Repeat count. When five items reach a node that costs 4 credits, the node costs 20 credits. The Run button shows the total before you run.

Badges show the batch. Before a run, the header of a node that will fan out shows the number of runs, such as x5. During the run it shows the progress, such as 3/5, and a Processing list... bar that reads Complete at the end.

Failures stay local. When one item fails, the items that succeeded keep their results, and the failed item's credits are refunded.

Tables: several values per run

A List starts as one column of text. When one run needs more than one value, for example a prompt and a reference image, the List becomes a table.

1. Connect a node to the + input at the bottom left of the List. A new column appears.
2. Set the type of each column: text, image, video, audio or JSON. Each column has its own input and its own output.
3. Connect each column's output to the input of the node that needs it, for example the prompt column to Prompt and the image column to References.

The rows pair up as follows:

• Row 3 stays row 3. Every input of a run gets its value from the same row. An empty cell gives nothing to that input for that run, and the rows below do not move up.
• A row runs when any of its cells has a value. A row that is empty in every column is ignored, so the blank row at the bottom never adds a run.
• The order of the connections does not matter. The column connected to the prompt input supplies the prompt, whichever connection you drew first. A column connected to Negative is never used as the prompt.
• Lists of different lengths repeat. When two lists with a different number of rows feed one node, the longer list sets the number of runs. The shorter list starts again from its first row when it runs out.
• Only Each fans out. A column connected in Bundle mode hands its whole list over in a single run.

On the canvas, a List shows either a compact summary, such as "4 rows × 2 cols", or the full table or gallery. A small button under the node switches between the two views. A table with images, videos or audio opens in the gallery view by default.

Fan-in: from many results to one

Choose Best

The Choose Best node turns a list of candidates into one result. Choose what it does under What to do with the candidates:

For AI picks the best, set three things:

• AI Model: the judge. An Economy model costs 3 credits, a Standard model 10 credits and a Premium model 25 credits. Standard is the default.
• Judge by: what a winner looks like, for example "the most eye-catching cover for a dark editorial feed, readable as a thumbnail".
• The candidates are: Texts or Images. Choose Images when the candidates are pictures, because on Texts the judge compares only the links of the pictures.

The node shows the chosen result, its position among the candidates and the judge's reason. The Candidates tab of the settings panel lists every candidate and marks the chosen one. Five images judged by a Standard model cost the five images plus 10 credits.

Empty candidates are ignored. A failed item leaves an empty candidate, and Choose Best removes empty candidates first. Count them therefore counts only the items that succeeded. When every candidate is empty, Join into one text returns an empty text and Count them returns 0. The other options fail, and the nodes after Choose Best do not run.

Collect

The Collect node gathers the results of any number of nodes into four lists, one for each kind: text, image, video and audio. Its four outputs are always there, so you can connect the next nodes before anything has run. Drag the rows in its settings panel to change the order of the items. Collect is free.

Connect Collect to a Choose Best node to compare results that came from different nodes, for example three covers made by three different models.

Selector

The Selector node picks items from a list and has two outputs: picked, the items it chose, and rest, every other item. Use it instead of an Item or range connection when several nodes need the same selection, or when you need a mode that connections do not have.

A Selector never stops the workflow because of a setting: an index that is out of range, or an expression it cannot read, falls back to a sensible pick. Selector is free.

Other list tools

The Lists & Batching family of the node picker also has Filter List, to keep the items that match conditions, Sort List, Remove Duplicates and Merge Lists, to join two lists end to end or item by item. All four are free.

Tips

• One idea per item. Each item should be a complete prompt or value on its own.
• Test with a short list. Run a list of two or three items first, check the results, then add the rest.
• Split large batches. A very large list makes one long run. Several smaller lists are easier to check.
• Close every fan-out. When a batch should end in one result, finish it with Choose Best, a Bundle connection or an Item connection.
• Use variables in items. Items can contain prompt variables, such as {Style}, so that every run shares the same wording.