Boost Your Design Workflow: A Hands‑On Guide to LoRA Training for Stable Diffusion

In everyday design work, 3D projects often require time‑consuming modeling and rendering. By leveraging AI‑generated content (AIGC) and training LoRA (Low‑Rank Adaptation) adapters for Stable Diffusion, designers can significantly reduce total design time.

LoRA, originally used to fine‑tune large language models such as GPT‑3, adapts large models by training only low‑rank matrices. When injected into a Stable Diffusion model, it changes the model’s style or adds new characters/IPs with far less computational cost.

Plugin preparation : Install the recommended Stable Diffusion LoRA trainer (a bundled package with Python) and unzip it.

Training data preparation : Collect 5‑10 images of the target IP that share the same composition, angle, and style. Avoid mixing perspectives or unrelated images, as they can pollute the training output.

Labeling steps : Follow strict naming and grouping rules—create a

train

folder inside the SD

models

directory, then place the source images in an

ip‑in

sub‑folder.

Size settings : Resize every image to 512×512 pixels; larger sizes slow down labeling, while smaller sizes degrade quality.

Generation settings : In the Web UI, select

deepbooru

for tag generation after preprocessing.

Training process : Copy the training set into the LoRA trainer, set input and output paths, and start training.

Result showcase : After training, place the LoRA file into the SD

models

folder, refresh the UI, and use the new keyword to generate images. The generated outputs retain the original style and show recognizable features.

Changing colors : Export the generated image to Photoshop for color adjustments, then feed the edited image back into the generation pipeline to control color palettes.

The workflow can also be applied to batch‑produce 3D‑style icons, enabling designers to create consistent, perspective‑accurate icons quickly. By training on a set of 3D icons and using the LoRA model, new icons can be generated that match the original style.

Conclusion : As AIGC advances, model training becomes essential for differentiated design. Although the process is still exploratory and complex, mastering LoRA training for Stable Diffusion opens new possibilities for efficient, high‑quality visual creation.