Skip to content

Chains

Chaining tools

Stack the steps in the order you want them to run, tune each one, and press Run once.

Drop tools into a chain frame in the order you want them to run. The first step takes your input, and every step after it takes what the one above produced.

  1. Start from the sequence you already ran by hand. The settings you tuned on those cards are the settings the chain wants
  2. Add the steps in order. Position is execution order, so the top row runs first
  3. Set each step’s model and settings. Every step keeps its own model pill and its own gear, exactly as the standalone card would
  4. Check the summed price on Run before pressing it
  5. Press Run once

The result lands on the canvas the way any single generation does, as one asset from one press.

The order that works is almost always cheap-and-structural first, expensive-and-final last.

Put early Put late
Generation, which decides what the thing is Upscaling, which multiplies the cost of everything before it
Background removal, which changes the composition Grading and finishing, which only makes sense on the final frame
Remeshing, which changes the topology everything else sits on Texture regeneration at 4K or 8K

Upscaling early is the classic expensive mistake. Every step after it works on a bigger image, which usually costs more and always takes longer, and you have paid to enlarge something you were about to change anyway.

Each row carries the same controls its card would: the model pill, and the settings behind the gear. Anything you change shows in the grey line under the step’s name, so you can see what a row is set to without opening it. Swapping a model inside a chain changes that step’s price, and therefore the total.

A chain hands you its last step and throws the middle away. Turn on Keep steps and every step’s result lands on the canvas instead, in a row beside the chain, in the order they ran.

It is off by default and only offered once a chain has more than one step, because the usual case is that the middle is scaffolding. Turn it on when you are still deciding whether the order is right, since seeing what step two handed step three is the fastest way to find the step that is letting the chain down.

The chain runs as one operation, so a failure part way through stops the sequence rather than delivering something half-processed. The frame names the step that stopped it.

The failure reports the same way any run does, with a code and a tier telling you whether it was your input, a transient provider problem, or something on our side. See when a run fails.

Closing the tab mid-run is not the same as a failure. Reopen the board and the chain says it was still running when the tab closed, and offers Pick up where it stopped. The steps that finished kept their results, so it continues from there rather than paying for the whole sequence again.

Only the browser that started the run sees that offer. A collaborator opening the same board is not asked to finish somebody else’s work, and neither are you on a different machine. The offer also stops appearing after about six hours, on the grounds that a run you abandoned yesterday is not one you are still waiting on. Dismiss clears it early.

A chain you have got right can be saved instead of rebuilt. Reusing a chain covers how.