Pipeline steps

Reference images: generate, upload, regenerate

Every character, environment, and object in a project gets one reference image, rendered against your style's template images. Those references are what every later shot renders against — they are the mechanism behind character consistency — and every asset must have one before the voiceover step unlocks.

Generate references

Each asset card has a Generate Reference action, and there is a bulk Generate all that fills the gaps. Bulk generation is fill-gaps-only on the server: assets that already have an image are skipped, never overwritten, so it is always safe to press. Retries are equally safe — assets with a generation already in flight are skipped, so a retry never double-bills.

Or upload your own

You can upload your own image for any asset instead of generating one. Uploaded references work exactly like generated ones downstream.

Orientations are fixed per type

• ·Characters render square.
• ·Environments render landscape, in two stages: an establishing canon shot first, then a 2×2 multi-view sheet built from it — two image generations per environment.
• ·Objects render square.

Pinned, not frozen

A reference is a lock for downstream shots, not a one-time commitment. At any point you can regenerate it with an edited prompt, swap the model, or upload a replacement image — and later shots will render against the new reference.

Models

Reference generation currently offers GPT Image 2 (with a low/medium/high quality knob) and Gemini 3.1 Flash, chosen per run in the expert drawer's model picker.

For why references exist and how they hold a look across fifty shots, see keep characters consistent across scenes and the character consistency page.