Consistency

Keep a character or style stable across assets.

This is the same playbook the agent skill carries. Names like fast and default are aliases; see Models for what they run on the hosted service; model names in the research notes below may include ones only the local tool offers. Prices are in credits.

For a character, mascot, product or room that must look the same across many assets. All rules [measured] on DeepInfra FLUX-2 / nano-banana unless marked.

Views generated independently produce a different subject each time. Use the two-phase flow: generate N anchor candidates → a human picks one → every other view is generated conditioned on the picked anchor. Skipping the pick gives "three views of three different people".
One sample per view drifts too. Non-anchor views need N candidates (seed+k), a per-view pick and an acceptance check.
The appearance string must be complete: full outfit including lower body and footwear, exact colours. Anything unspecified is re-rolled per render. Inject it verbatim into every prompt; never re-describe per image.
Lock bodies, not just faces. Hands and legs give AI art away; a dedicated hands plate is the targeted defence [inference].
Identity first, set last in the reference order, and cap references yourself: the provider drops extras from the end without error.
Round-robin across subjects when several need references, reserving one slot for the background plate; otherwise one subject eats the budget.
Single-reference models take a collage: 2, 4, 6 subjects side by side in one image were all recognisable; 8 gave 7 of 8. Poses copy the lineup, so collages suit wide shots, not close-ups.
Background plates are copied near-literally. One plate per camera position; a plate from another angle puts the wrong room behind the shot.
Reference images must be caption-free: text in a reference teaches the model to draw text.
Test honouring before trusting: generate fixtures, verify each against its own description with a judge, then render a scene whose prompt never describes the fixtures, with a no-reference control. A returned image proves nothing; only an output showing the reference's distinctive features does.
Promote approved outputs to references sparingly: human-approved only, never more than one generation removed from the locked set, or drift compounds [inference].
Editing a locked identity invalidates everything after it. Warn at the point of edit and regenerate the reference set.
Deterministic seeds (base + index) make reruns repeat; identical payloads are cache hits, so "variants" must differ by seed.