MasterBedroom is a proof of concept exploring whether a LoRA can be trained on a single, specific room to create a more consistent background. I chose a relatively simple bedroom as the subject to test this idea.
The LoRA is somewhat flexible, allowing you to choose which elements of the room you want to include in your image. The main prompt contains the elements found in the training images, but they can be omitted or combined with your own prompts to create different variations of the room.
Since the source material is realistic, the LoRA naturally leans toward a realistic appearance. However, using a lower LoRA weight together with appropriate prompting can produce semi-realistic and, to some extent, anime-style results as well.
Main prompt:
masterbedroom, indoors, bed, plaid blanket, pillow, window, window blinds, cityscape, fog, floor lamp, hanging light, potted plant, bonsai, rug, wooden floor, grey floor, white wall
The training dataset includes a variety of aspect ratios: 1:1, 2:3, 3:4, 3:2, 4:3 and 16:9, allowing the LoRA to work across a range of image compositions.
Description
General tips and recommendations.
Lora Weight: 0.6-1
Resolution: 1024x1536, 1152x1536
Steps: 30-40
CFG: 4-6
Sampler: Euler, Euler a, DDPM
Upscalers:
4x_foolhardy_Remacri for realism
4xOpscaler_mixed for anime
Adetailer with segment models is highly recommended


