Turnaround: clean method for generating a consistent character turnaround that actually holds at all four angles
Turnarounds are the foundation of everything (they're your LoRA dataset seed and your canon reference) and most people's turnarounds secretly don't hold — the back view is a different character. Here's a method that actually works.
The sequential-anchor method:
Lock your FRONT view as absolute canon. Everything derives from this.
Generate the 3/4 using the front as InstantID reference. It'll be ~85% right. HAND-FIX the 15% (eye spacing, nose, jaw) until it's undeniably the same person.
Generate the PROFILE using the CORRECTED 3/4 as the reference, not the front. Each angle anchors off the nearest already-corrected angle, so the identity information travels around the head gradually instead of trying to jump straight from front to back.
Generate the BACK from the corrected profile.
Now lay all four on one sheet at the same scale and eyeline, and eyeball-check: is this ONE person from four angles? If any view feels like a sibling, fix it before it poisons your LoRA.
Key insight: don't generate every angle from the front. Daisy-chain each angle off the nearest corrected one. Front-to-back is too big a leap for the model; front-to-3/4-to-profile-to-back is four small believable steps. Example sheet in comments.