Synth Species '24

👉 Check out the Pony XL version! 🐎

👉 Check out Cynfall's version of this lora! They're pretty similar, try both!

👉 Learn more about Vader-San's Synths on their homepage!

You want an even better LoRA?

Me too! Contact me on the Furry Diffusion server, I'll help any way I can. (Ask for "Mo.")
I don't care whose LoRA is best, I just want great AI-generated Synths!
👉 https://discord.gg/furrydiffusion

How was the LoRA improved?

I spent a week improving the dataset - adding images, cropping some, cleaning up tags - and another week re-training the LoRA on 420 Synth images + 420 regularization images... so 40 epochs was 33600 image-steps at 960 resolution. I was comparing different versions using the methodology above. I arrived at these settings:

$num_epochs = 40
$train_batch_size = 3
$optimizer_type = "Prodigy"
$learning_rate = 1 * $train_batch_size
$text_encoder_learning_rate = 1 * $train_batch_size
$network_dim = 96
$network_alpha = 96
$scheduler = "cosine"
$clip_skip = 1

--shuffle_caption
--enable_bucket
--mixed_precision="bf16"
--min_snr_gamma=5
--scale_weight_norms=1.0
--v_parameterization
--zero_terminal_snr

I trained on the fluffyrock-576-704-832-960-1088-lion-low-lr-e159-terminal-snr-vpred-e132 model, but the LoRA works well with EPS models too - like IndigoMix I used for the showcase.

Disclaimer!

I originally included the following comparison between this lora and the other 2 available on CivitAI. However, in May 2024 it turned out Cynfall's lora ("FRL27nO") was broken all this time, so I was comparing my lora to one that never worked properly. On May 23rd he uploaded a fixed version, so I edited this page as now I recommend trying both mine and his.

Of course these 6 are cherry-picked to show good images from my model. But the 100 batches I rated in total to compare these 3 LoRAs were not cherry-picked.

Testing methodology

I generated 4 images from randomized prompts: 3 using the 3 LoRAs and one without a LoRA. I used 4 models: Easyfluff 11.2 (40 images), IndigoMix v110 Realistic (30 images), IndigoMix SE01 vpred (20 images), IndigoMix v100 Anime (10 images).

I shuffled the images in each batch, so I wouldn't know which was from which LoRA. (That's the blind in blind testing.) I rated images in each batch on a scale from 1 (awful) to 9 (perfect). On average, the three models scored, mine: 6.04, NBP: 4.76, and FRL27nO: 3.99. Images with no LoRA scored 3.37.

After un-shuffling the results, I used the Wilcoxon signed-rank test (scipy.stats.wilcoxon) to see if the score differences are statistically significant.

Synth '24 v17 (6.04) >>     NBP (4.76):  T = 331.0,  p =  0.00%
Synth '24 v17 (6.04) >> FRL27nO (3.99):  T =  96.0,  p =  0.00%
Synth '24 v17 (6.04) >> no lora (3.37):  T =  80.5,  p =  0.00%

With differences this large the Wilcoxon test was pointless, but I also used it to compare the 16 versions of my new LoRA which came before v17. Sometimes the scores were much closer.