The drift problem
Vanilla SDXL given the same prompt twice produces two visibly different people. For a companion product where users form an attachment to a specific face, that is a product killer.
Prompt engineering helps a little. Reference images via IP-Adapter help more. Neither closes the gap on its own.
Why LoRAs win for this
A LoRA trained on 20 to 60 curated images of one character locks the face, body, and styling into the diffusion model itself.
Inference-time cost is roughly the same as base SDXL. Training cost is a few dollars on a small GPU pod.
How LovlyChat trains them
Once a character is launched, a one-time job runs on a RunPod A40 with kohya_ss for around 2000 to 4000 steps.
Output LoRA file is a few hundred MB and is registered against the character id in our character store.
How the API uses them at request time
When POST /v1/generation/jobs is called with a characterId, the worker loads the matching LoRA into the SDXL workflow and applies it during sampling.
Per-character LoRAs are why every image you get back of the same character looks like the same person.
Veelgestelde vragen
How LovlyChat Keeps Character Images On-Model: Per-Character LoRAs Explained explains Why generic SDXL drifts between generations of the same character and how the LovlyChat stack uses per-character LoRAs trained once on cheap GPU pods to keep faces consistent. It is written for users comparing AI companion products, evaluating relationship-style chat experiences, or looking for a clearer way to understand this topic before trying a platform.
This article is for people researching sdxl character consistency, comparing companion platforms, or trying to understand which product features make relationship-style AI chat feel more personal and worth returning to.
sdxl character consistency matters because users searching this topic often have stronger intent than broad chatbot traffic. They usually care about memory, personality, and continuity rather than only novelty or one-off responses.
Users can explore these ideas directly on lovlychat by browsing companions, creating a custom character, and opening persistent chat threads in the browser.