Local image models still can't match ChatGPT at character consistency from one reference

tammy_orbit · reddit · 2026-10-13

A Reddit user asks whether any local model can generate consistent character images from a single reference (a Hogwarts-style character shown in three views). Out-of-the-box attempts in ComfyUI, including Qwen Image 2.1, produce good-looking but clearly different characters. They wonder if fine-tuning could close the gap or if the capability remains out of reach for local models.

Original post →

More from Multimodal

Multimodal channel →