Details

A LoRA trained on Midjourney V5 generated portraits, using SD 1.5 as base.

It generates a portrait of a happy/smiling subject (a close-up most of the times), on a sunny day, with a depth of field effect behind it.

Models I personally use this with:

• Deliberate V2

• RealisticVision V2.0

• Custom merge of (0.5*Deliberate V2 + 0.5 * realisticVisionV20_v20)

Training

Training method: Dreambooth LoRA (Koya's GUI)

Training data: 20 images

Repeats: 10

Epochs: 10

Batch size: 3

LR: 0.0001 (no warmup)

LR scheduler: cosine_with_restarts

TE_LR: 5e-5

Unet LR: 0.0001

Optimizer: Lion

Network Rank, Alpha: 128,128

Other: No xFormers + Memory efficient attention

Trained using AMD Radeon 6800XT GPU on Linux.

Docker images available here:

https://www.reddit.com/r/StableDiffusion/comments/136ncb8/amd_rocm_dockerfiles/

Example

Example prompt: (((best quality))), (masterpiece), a portrait of an italian woman, 30yo <lora:SunnyPortraits:0.7>

Disclaimer

This is model cannot be used on generation services, but generated images can be sold.

This licensing rule applies to merges as well.