Case-Hardened
CaseH does not mean hentai case; instead, CaseH stands for Case Hardened.
It derives from the SDXL distillation, creating a 1.5 model.
Recommended configuration:
- Sampler: DPM++ (series)
- Steps: 15-55 (recommended 35)
- CFG: 2-5 (recommended 4)
> The higher the CFG value, the fewer steps should be used.
How
This model represents an idea I have been experimenting with. It seems I've managed to produce a release-worthy version.
Actually, my concept is quite straightforward, based on the following two assumptions:
- sd15 model has ample capacity to accommodate more knowledge.
- Consistency in architecture implies the possibility of distillation (and we could focus on learning just the realistic part).
Therefore, I attempted to use the SDXL model (one of the more popular versions on Civitai) as the teacher model to distill knowledge into the SD1.5 model (the beta version is based on the LOFIv4 tune-up).
> To my surprise, this round of fine-tuning has significantly surpassed previous distillation (raw-xs) versions. After the tune-up, LOFIv4 has almost completely abandoned the asian-style and shifted entirely towards SDXL.
Issues
ti model lora model may not be used as expected
Due to the strong leaning towards learning from SDXL, it could cause the action on the 1.5 TI model to fail, such as with the ng_deepnegative_xx series models.
backup hugginface:
This is a test model, fine-tuned based on lofi_v4
Files
caseh_beta.safetensors