Workflow Share: my full character-consistency pipeline from design to finished shot, start to end
Enough people asked that I'm writing the whole thing out. This is how I take a character from a sketch to a shot that holds. It's a lot of steps because consistency is a lot of steps.
Lock the design. I generate ~20 candidate character sheets, pick ONE, and treat it as canon. No more design changes after this or everything downstream drifts.
Build the turnaround. Front, 3/4, profile, back. I use the chosen face as InstantID reference to generate the angles, then hand-fix until all four are unmistakably the same person. This is the dataset seed.
Train the LoRA on the turnaround + ~25 generated-and-cleaned variations (expressions, lighting). Diverse dataset, face constant.
Per shot: LoRA for the character + InstantID face-lock from the canon front view + pose/composition control (ControlNet or a reference). The LoRA does the body/outfit, InstantID enforces the face.
Motion: animate the still with image-to-video (Kling/Hailuo for body), then drive the face separately with Act-One for dialogue shots.
Face-lock cleanup pass: even with all this, 1 in 4 shots drifts a little. I fix the worst frames with a targeted face inpaint using the canon reference.
It's not push-button. Consistency is a PIPELINE, not a setting. Full node graph + the character bible in comments.