Sign In

One-image LoRAs vol.1

45
1.2k
0
4
Type
LoRA
Stats
290
Reviews
Published
Apr 30, 2023
Base Model
Other
Training Images
Download
Hash
AutoV2
6B45108476

This is an experimental collection made of LoRAs trained on 1 image each. They are obviously overfitted, but this is the intended result.

Be aware that results are frequently unstable, so use or don't use trigger words accordingly. Don't hesitate to change the rate of LoRA or any trigger word if you are using one.

The base model is Suzumehachi.

UPDATE:

The workflow.

The way I create it is as follows. I find an interesting picture and either make it square or manually crop it into one or two interesting pieces. Then I use BLIP (sometimes with DeepDanbooru) to caption them and check if everything is ok with that. After that, I usually add a special trigger word just in case (to give some punch to LoRA if it will struggle in an unusual prompt environment)

I use kohya with the parameters (only the most important ones are shown):

number of steps for dataset image - 100-200

--network_alpha="128"

--text_encoder_lr=5e-5

--unet_lr=0.0001

--network_dim="128"

--lr_scheduler_num_cycles="1"

--learning_rate="0.0001"

--lr_scheduler="constant"

--train_batch_size="2"

--mixed_precision="bf16"

--clip_skip=2

--noise_offset=0.1

--min_snr_gamma=5

Because of my other workflow, I make 4 variants of a LoRA with different seeds and combine them pair by pair with around 0.7 mult. factor:

python.exe "networks\merge_lora.py" --save_precision fp16 --precision fp16 --save_to combined_lora.safetensors --models lora_1.safetensors lora_2.safetensors --ratios 0.7 0.7

If something goes wrong I recombine them with different mult. factor or exclude some LoRA-outlier.