Not loaded
Download
1 variant available
bf16 SafeTensor
EyesEyelids_SapphireEmerald_V1_Krea2turbo_DiT_TE_R32_3400st.safetensors
BF16, good balance • 218 MB
Verified: 21 hours ago
470 1 2 3 4 5 6 7 8 90 1 2 3 4 5 6 7 8 9
00 1 2 3 4 5 6 7 8 9
(4)
Oct 17, 2026
Dataset: 100 high-detail images (50 per eye color)
Captioning: 100% manual tagging (no auto-AI captioning)
Training Steps: 3,400
Network Rank (Dim): 32
Learning Rate: 0.00015
Text Encoder Trained: True (DiT + TE)
Trainer: Ostris AI-Toolkit
Hardware: NVIDIA GeForce RTX 5090 / Intel Core Ultra 9 285K / 96GB DDR5
Show more

520 1 2 3 4 5 6 7 8 90 1 2 3 4 5 6 7 8 9
310 1 2 3 4 5 6 7 8 90 1 2 3 4 5 6 7 8 9
Ever tried giving your elves, dwarves, beastfolk, or fantasy creatures truly vibrant, mythical eye colors, only to have the base model default back to its boring, washed-out training data?
Say goodbye to muddy shades of gray, dull hazel, and generic brown. Now you can easily dial in rich, striking sapphire blue or deep emerald green for any character in your frame.
Dial the weight to taste — this LoRA scales cleanly anywhere from 0.1 all the way past 1.0. My sweet spot is usually between 0.6 and 0.9.
And yes — full heterochromia is totally doable (one blue eye, one green eye).
---
### Did you like the girls? :)
If you liked the showcase models from the preview cards, you can grab them here:
* Helena Blau (Blonde / Sapphire Eyes):
https://civitai.red/models/3007155/helena-blau-ai-generated-fake-character
* Keira O'Green (Redhead / Emerald Eyes):
https://civitai.red/models/3007217/keira-ogreen-ai-generated-fake-character
---
### 💡 Pro Tips & Prompting:
* Target the irises, not the whole eye: After your trigger eyesk2sapphire or eyesk2emerald, add simple tags like vivid blue irises or vivid green irises to the specific character. Avoid generic prompts like blue eyes — if a female character wears makeup, the model might tint her eyeliner too.
* Diverse Dataset: Trained on 50 high-res images per color across different ethnicities (Caucasian, Asian, African), various ages, genders, with and without makeup, and even kids. It knows how eyes actually sit on different face structures.
* Eyelids & Eyeliner: It refines eyelid folds and contours, but keep the strength in check — cranking it too high might trigger unwanted winged eyeliner. The dataset includes diverse eyelid types to keep it versatile.
* Architecture: Trained locally with Ostris AI-Toolkit targeting both DiT and Text Encoder (DiT+TE, Rank 32). Because of the trained TE, if you're stacking this with character LoRAs trained without a text encoder (like quick fal.ai runs), you can dial back the strength a bit to avoid frying the prompt.
Drop your gens, reviews, and test runs in the comments — always hyped to see your feedback and what you build with it!



