Where does character consistency break first: training, prompting, or inference control?

remixeconomy · reddit · 2026-09-08

A Reddit poster examines the attribution problem behind character consistency failures in image generation: it's hard to separate "bad character model" from "bad way of testing the character."
A LoRA can look rock solid on prompts near the reference set, but drifts badly with new camera angles, full-body shots, clothing changes, extra people, or face occlusions — which everyone then blames on training.
In reality multiple factors mix: how much identity the model learned, prompt-identity competition, reference/control strategy at inference, and distance from the training distribution.
Possible fixes listed: retrain/change dataset, change captions or trigger strategy, adjust LoRA strength, add reference/control at inference, restructure prompts, or accept some shots need edit/inpaint instead of fresh generation.
The author argues drawing the line between "model isn't good enough" and "inference/control problem" matters far more than another gallery of perfect generations.

More from Multimodal