Realistic AI Avatars
Photorealistic AI characters with a consistent identity.
Characters that stay recognisably the same across hundreds of stills — different scene, different light, same face. It runs in ComfyUI on a Colab GPU: the cyberrealisticPony SDXL checkpoint, a pair of LoRAs for realism and texture, a FaceDetailer with YOLOv8 and SAM that regenerates the face at high resolution, and an UniversalUpscaler pass. One run produces four takes of the same shot, each with a different LoRA pairing — the keeper is the one that holds the identity, not the prettiest one.
A closer look
One identity, three scenes
A beach, a castle wall, Paris — different light, different focal length, different clothes. The face has to stay recognisable through all of it, and that is what the whole pipeline is for: without the FaceDetailer pass the features drift on every generation and ten images later it is a different person.

A second character
Same pipeline, a different type. Harsh outdoor sun and soft indoor light are two entirely different problems for the model — skin has to behave like skin in both, not like plastic. That is what the realism and amateur-photography LoRAs are for; resolution alone will not do it.

How the pipeline runs
The graph is built around four parallel paths: same prompt and seed, different LoRA pairing. Each path goes through the FaceDetailer and a 4× upscale scaled back down by half — which lands on a clean 2× without over-sharpened artefacts. Models and custom nodes live on Drive and ComfyUI launches from a Colab notebook, so none of it needs a local GPU.