The Generate tab

Every Media Gallery makes AI media from the Generate tab on its right-hand side. The tab shows the gallery's recipe: the series of operations that is going to run, in order, each as a form you can open, fill in and run. A recipe with one step is one card; a recipe with several shows them one under another, and nothing else about the panel changes.

The steps

Each step is a card. The title row of the card names the step and its operation -- Step 2: Voice Changer -- and the expand / collapse button is at the right end of the row. Inside, the step is its operation (a dropdown) and the operation's inputs, nothing more. The i button beside the dropdown opens everything about that operation in a dialog -- the visual style in effect, the model build, whether it runs on your machine and on Comfy Cloud, whose likeness it carries, and the operation's description -- with a Close button. When one of a step's inputs is fed by an earlier step, the title says so: Step 2: Voice Changer -- from step 1's output. A step that has already made something carries a tick (Already made); Run all skips it.

The cards start collapsed except the one you are on -- the first step that has not made anything yet. A step that is running, or has just finished, opens by itself so you see the result arrive. The unfold button at the right above the cards (Expand all steps / Collapse all steps) opens or closes every one. Opening a step shows the same form you would see for that operation on its own: the prompt (with Write prompt, which knows the shot's line), the seed, the images and files it takes, and the facts about the model.

Under the operation's name the form lists three facts on separate lines: Model build (the precision, the download size and the video memory the build needs), Local (whether it runs on your machine, and how well) and Comfy Cloud (ready, or not available when the workflow needs nodes Cloud does not have -- such a workflow runs on local ComfyUI only). The workflow's description follows, folded to two lines with Read more.

Steps that feed each other

An input that an earlier step fills says so under the step's description: Source Video: Filled from step 1's output when it finishes. You cannot place a file there by hand until that output exists; afterwards the input shows the file like any other.

When you New recipe step (or swap a step's operation), OA wires it to the step before it by itself: the first input of the new operation that takes the kind of media the previous step makes (a video, an image or audio) is filled from that step's output. If no input fits, nothing is linked, and the step says so under its description: Not filled from step 1's output: step 1 makes a video and this operation takes an image. Pick an operation that takes what the previous step makes (a video upscaler after a video step, not an image one). To change it, open the menu at the right end of that input's line: it lists From step N's output for each earlier step that makes the right kind of media, and Choose a file. Picking a step links the input; picking Choose a file unlinks it and the input takes a file you drop on it, as any other. Inputs on the first step, and on a shipped recipe (choose Edit a copy first), have no menu. Reordering steps drops a link that would point forwards.

Running

Press Run in a step to run just that step. If it cannot run, the reason is written beside the buttons ("Local ComfyUI is not connected.", "This workflow runs on local ComfyUI only.", "A run is already in progress."). Press Run all in the recipe's header row to run the whole recipe: each step that is not already up to date runs in order, waiting for the one before it to finish. While it runs the button reads Stop after this step, the step being run shows a busy spinner where its menu was and its card's border lights up; a render that has already been sent to ComfyUI cannot be recalled and finishes on its own. A slow render is never mistaken for a stuck one: the next step waits as long as the job is still running.

To use a different operation for one step, press the step's operation name. That swaps that step only and leaves the rest of the recipe alone.

A step can name an operation without naming a model -- the ready-made Finish the clip recipe says "upscale" and lets your look choose the upscaler. When your look has not chosen one yet, the step asks you once: its heading is the operation (Upscale), the dropdown lists only the models that do that operation for what the step is given (a video take offers the video upscalers, not the image ones), and the note under it says Your Anime look has no upscaler chosen yet. Pick one; it is remembered as the look's upscaler. Nothing runs until you pick. Your pick is saved on the look, so the next shot's step resolves by itself, and you can change it later on the Appearance page. When you run Finish the clip here, the clip it makes becomes the shot's finished clip: the Finish row in the episode plan shows it beside the rough take, and the sequencer plays it.

Which recipe is this?

Above the steps, when OA knows which recipe the steps belong to, there is one row: the recipe's name (for a shot, Talking shot), a small info button that opens the line saying why it suits the shot, Run all, and a ... menu. The recipe that runs is the one the show has chosen for that step of the production plan, or, if nothing is chosen, the ready-made recipe that fits the shot. See Models, Workflows & Recipes.

The ... menu holds, in this order: Edit a copy, Rename (your own recipes only), Save as the show's recipe for the plan step, Choose another recipe, and Load the show's recipe (only when the steps on screen are not the show's recipe).

Choose another recipe opens a dialog headed Choose a recipe, captioned for what it makes -- for a clip, Other ways to make the shot video -- listing the other ready-made recipes that make the same thing: two or three, never the whole catalogue. Press Use for this run beside one and the steps are replaced with that recipe's, for this run only. OA then offers Remember for the show, Remember for the scene and Remember for the shot: pick one and that recipe becomes what that level uses for this step from now on. The choice shows in the Recipes chosen list on the Project, Scene or Shot properties page, where you can remove it.

Editing a recipe

A ready-made recipe is read-only: you can change what is in its forms and run it, but not add, remove or reorder its steps. Choose Edit a copy from the ... menu, name the copy, and it becomes a recipe of your own in the project. Then you can:

• New recipe step -- choose an operation and it is added at the end.
• Delete a step from the ... menu on its title row; it asks you to confirm first. Anything wired to it is unwired.
• Reorder the steps by dragging the grip at the left of a heading.
• Rename the recipe.

Save as the show's recipe for the step (it names the step of the plan, for example Make the video) makes your recipe the one the whole show uses for that step. There is no separate editor page: the Generate tab is where you see, edit and run it.

Load what made this

On a result's menu (in the Grid, List or Graph view), two actions read its method back:

• Load as Operation puts the one operation that produced it into the Generate tab, with the same prompt and settings and the same images fed in.
• Load what made this loads the whole chain that led to it, as the steps of the recipe. A clip that was made by animating a storyboard picture and then given a character's voice comes back as those two steps, step 2 fed by step 1. Press Run all to make it again, or change something first.

When the chain is a ready-made recipe (a talking shot is Talking shot), the header above the steps names it, says why, and offers Edit a copy, exactly as if you had loaded the recipe itself. When the steps on screen are not the recipe your show uses for that step, Load the show's recipe in the ... menu brings the show's back.

An input that has since been deleted from the gallery comes back empty so you can see what is missing.

The assistant and the steps

The assistant sets up and runs the same steps. Ask it to make a talking shot and the steps appear here with their inputs filled; ask it for another operation after a result and it adds a step rather than replacing the one you can see. You can open any of the steps it wrote, change them, and run them yourself.