Training an img2img LoRA on just 35 pairs to turn photos into Unreal Engine style

NeoHisto · reddit · 2026-09-21

A ComfyUI beginner shares a project converting photorealistic images into an Unreal Engine look: they generated photorealistic counterparts with GPT Image 2.5 for 35 Unreal-style images, then trained an img2img LoRA in reverse, but results fell short—likely because AI-generated "photoreal" images lack diversity. They ask whether to build a better paired dataset, how to preserve quality (upscaling tips welcome), and whether the approach is worth the effort versus alternatives.

Original post →

More from Multimodal

Multimodal channel →