A manga creator says background consistency is harder than characters in NoobAI workflows

Boubbay · reddit · 2026-07-28

A manga creator says character LoRAs work well in NoobAI/ComfyUI, but background consistency remains the bottleneck. Training a background LoRA did not help much, so they are currently using DALL·E to generate multiple consistent views of the same locations.

The hard part is compositing characters into those scenes without the background changing. They have tried inpainting and ControlNet with simple sketches, but both often alter the background or even clothing colors. Their current fallback is to generate backgrounds first, sketch characters in Procreate on a separate layer, then use ControlNet or manual compositing plus a light img2img blend. They ask whether anyone has a workflow that reliably preserves the scene every time.

Original post →

More from Multimodal

Multimodal channel →