👉 Check out the <a target="_blank" rel="ugc" href="https://civarchive.com/models/459373">Pony XL version! </a><a target="_blank" rel="ugc" href="https://emojipedia.org/horse">🐎</a>👉 Check out <a target="_blank" rel="ugc" href="https://civarchive.com/models/208127">Cynfall's version of this lora!</a> They're pretty similar, try both!👉 Learn more about Vader-San's Synths on <a target="_blank" rel="ugc" href="https://synthspecies.com/">their homepage!</a><h3 id="you-want-an-even-better-lora-f98z598hh">You want an even better LoRA?</h3>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! <a target="_blank" rel="ugc" href="https://emojipedia.org/backhand-index-pointing-right">👉 </a><a target="_blank" rel="ugc" href="https://discord.gg/furrydiffusion">https://discord.gg/furrydiffusion</a><h3 id="how-was-the-lora-improved-4g0xns5g5">How was the LoRA improved?</h3>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:<pre><code>$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</code></pre>I trained on the <code>fluffyrock-576-704-832-960-1088-lion-low-lr-e159-terminal-snr-vpred-e132</code> model, but the LoRA works well with EPS models too - like IndigoMix I used for the showcase.<h2 id="disclaimer!-014bz983r">Disclaimer!</h2>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.<img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/9e4ee640-fc62-4fdd-a8c0-99700ec7553c/width=525/9e4ee640-fc62-4fdd-a8c0-99700ec7553c.jpeg" />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.<h3 id="testing-methodology-647niihgo">Testing methodology</h3>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 (<code>scipy.stats.wilcoxon</code>) to see if the score differences are statistically significant.<pre><code>Synth '24 v17 (6.04) &gt;&gt; NBP (4.76): T = 331.0, p = 0.00%
Synth '24 v17 (6.04) &gt;&gt; FRL27nO (3.99): T = 96.0, p = 0.00%
Synth '24 v17 (6.04) &gt;&gt; no lora (3.37): T = 80.5, p = 0.00%</code></pre>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.

Files with SHA256: c4477c8130676a39979a3bc5cc464eff410c5e1d88cc54b7aa228d92c76b94f2