Prompt recipes

Short patterns built from keywords, answers and the model pills. Each one is a prompt and what to tick.

One character, four places

[style], [character], [scene]

• Tick one answer for [style] and one for [character].
• Tick four for [scene].

The summary reads 4 jobs total. Each job is labelled with the scene that made it, so the results tell you which is which.

A grid, then only the good ones

Three characters, three scenes and three kinds of lighting is 27 jobs. Usually you want a handful.

• Tick all nine answers and open the Prompt Replacements panel.
• Read the 27 prompts in its preview and untick the ones you do not want.

The total follows the ticks. A prompt you unticked is never queued, which is not the same as queueing it and cancelling.

Two different figures in one picture

A keyword used twice gives the same answer twice. Address the answers by position instead:

[style], [character][0] and [character][1] sharing an umbrella, [scene]

Tick two characters. The first one ticked is [0] and the second [1]; the menu shows each one's number. An indexed keyword does not multiply the run, so this is one job.

Two finished descriptions side by side

When each figure is itself built from keywords (a character, hair, an outfit), build it once and capture it.

1. Write the first figure with its keywords and tick its answers.
2. In the preview, press ＋ on the prompt you like. It is saved as [prompt1].
3. Change the ticks for the second figure and capture it as [prompt2].
4. Write the scene around them: [prompt1] stands beside [prompt2], [scene].

A capture is finished text. It does not vary, and both figures keep their own hair and outfit where two uses of one keyword would share an answer.

Change the world, keep the prompt

[style], [character], [expression], [action], [scene], [lighting]

Run it, change the Collection drop-down, run it again. Nothing else is edited. Starter collections shows this prompt in all four.

The same picture on three models

• In the model field, search for and pick three checkpoints. Each becomes a pill.
• Leave the seed empty and set the run count to 2.

That is three jobs of two runs. The models share seeds: run 1 uses one seed on all three, run 2 another. ComfyUI loads each checkpoint once and renders both of its runs before moving to the next.

That picture again, with one change

• Open the file and raise Remix with ⤷ Inherit (this image).
• Press ↺ this file's seed. The seed is pinned.
• Change one word of the prompt, or one LoRA's strength, and run.

With the seed pinned, what changes in the result is what you changed.

A kept picture through another workflow

• Open a picture you like and raise Remix.
• Choose an image-to-video or image-to-image workflow from the drop-down. The picture becomes its input image.
• When asked which prompt to keep, keep the picture's own.

With Match Input Image ticked, the run is sized to the picture.