UPDATE V3: New dataset. Includes a variety of different subjects and landscapes (V1 & V2 only had women in the dataset). Better prompt adherence, better body anatomy. Includes different lighting.
This uses my trained Zi lora/checkpoints from Illustrious mixed in with Anima Base!
Works very well with natural language.
YOU NEED:
qwen base encoder & qwen image vae
All gallery images should have metadata so you can drag and drop to see the exact things you need.
SETUP:
Steps: 30 - 50
CGF: 4 - 5
Sampler: Eular/Eular A/res 2s (haven't tested others)
Description
New dataset
- 187 hand-picked images, up from 77. They were chosen from 900 candidates, and every caption was checked by hand.
- V2's dataset was almost all young women. V3 adds men, couples and groups, fantasy races (elves, orcs, androids), animals and creatures, scenery, food and objects.
- Ages range from the 20s to over 70, with a wide range of skin tones and body types.
- Rim light isn't only neon anymore: sunsets, moonlight, campfires, windows and stage lights all give the Zi glow. A few soft-light images are included so that soft-light prompts stay soft.
Captions
- Booru tags plus one or two plain-English sentences, so both tag prompts and natural-language prompts work.
Training
- V2: Prodigy, 1536 px only, the whole LoRA merged into the base.
- V3: AdamW at a constant 2e-5, trained at 512, 1024 and 1536 px. The release is epoch 16 (about 2,250 steps). Only blocks 7–20, the ones that carry the style, were baked into Anima base v1.0. The rest of the model is plain Anima.
What changed in testing (V2 vs V3, same 20 held-out prompts and seed)
- The Zi look is the same or stronger, with deeper blues and stronger rim light in night scenes.
- Compositions are fuller: a large subject stays large and detailed. A dragon over a castle went from tiny to filling the frame.
- Daylight and soft-light prompts stay soft instead of being pushed toward neon. The tone runs a little cooler.
- V2 printed garbled text on a shirt in one test. V3 didn't.