Verified: 6 days ago
SafeTensor
Type | |
Stats | 260 0 |
Reviews | (25) |
Published | May 20, 2025 |
Base Model | |
Hash | AutoV2 24CC8DE78B |
What I see with 'realistic' versions of Illustrious, is that they loose capabilities. I've been tinkering with it trying to push it toward more realistic feel bit by bit. Basically training a lora, checking how much of it can I mix in without loosing too much. Merge lora, train on the new checkpoint and repeat. After six passes I started to feel I'm not getting anywhere. So there it is. Somewhat trained, somewhat merged, somewhat realistic, somewhat illustrious.
Suggested settings: scheduler - Euler or DMP++ 2M SDE, cfg 2-8, steps 16-40, depending on desired effect. If you go for illustration feel, less steps, more cfg. More steps, less to mid cfg for realistic results. Default mode is realistic, it needs proper prompt to go elsewhere.