Lessons from Queso and READ ME, as recorded in the canon.
- Midjourney v8.2 with --raw --hd and her two reference images through --edit beats v7 --oref: v7 holds the face but is stiff and posed; v8.2 text-only is lively but the face is someone else; v8.2 --edit gives both (Obie: 'these are much much better').
- Leave her v7 profiles off with v8.2.
- Shots without her (food, hands, the room, the boyfriend alone) take no --edit references.
- Fix plates with an image-edit model (GPT Image 2.5), never by hand.
- Obie 2026-10-12: 'you should seriously consider using grok and gpt image 2.5 a lot more than midjourney'. Read Me live plates were re-shot with GPT Image 2.5 from the reference faces; Midjourney was kept only for a few frames and the parrot world.
- Wording that asked for stray hair produced stray hair: say what you want absent ('her face clear of hair') in every Seedance prompt.
- One costume per film. Read Me re-dressed 14 frames into the V2a bodysuit and added boots to 16.
- Seedance re-voices a vocal in its own timing and the mouth follows that soundtrack, so retime the take to the song, or for a spoken clip use the take's own soundtrack.
- An expression the video model will not perform must be edited into the still first.
- Keep each Midjourney prompt under about 800 characters before parameters.
- Match skin to the grade targets before making clips. Grid skin CIE Lab for Saindra: L 55, a 22, b 25 (tolerance about 4 in a and b, 6 in L).