This repository will contains the needed safetensors to upgrade SDXL models based on Illustrious XL when new versions are released.
Since a lot of SDXL models are now based on Illustrious XL, each time a new version is released, it is painful to "upgrade" our merged models to benefit from the new capabilities.
Here, i'll store my "helpers" for my own method of upgrade:
- Replace the CLIP of the model with the new CLIP using a 0.8~0.9 weighted sum (so replacing it with most of the new CLIP but still keeping a bit of the original model)
- Merge provided LoRAs at preferred strength to update the UNET (the LoRAs are UNET only and built using SVD with Kohya_ss script using Illustrious checkpoints)
Provided safetensors (for now):
- Illustrious XL V2.0 CLIP: illus2-clip.safetensors
- V0.1 to V1.0 UNET LoRA: ZeroOne_To_OneZero.safetensors
- V1.0 to V2.0 UNET LoRA: OneZero_To_TwoZero.safetensors
- fp16 "SDXL-ready" version of both of these models (CLIPL, CLIPG): UniversalCLIPbase.safetensors
Example procedure to upgrade a model based on V0.1 to V2.0:
- Replace the CLIP using python script (will be provided later)
- Merge both LoRAs at strength 0.8 (using Kohya_ss Pluslora.py)
Update: Added IXL2NAI.safetensors obtained by SVD "extraction" (using Kohya_ss extract_lora_from_models.py). This will allow using a merge at strength -1 to remove a large chunk of Unwanted NoobAI influence on models which might have been "tainted" during merges.
A quick and dirty python script (compare.py) for cosine similarity using IXL 0.1/2.0 and NAI is included for reference (providing those checkpoints to the script is needed)
python compare.py HoJ_IXL-V4.0.fp16.safetensors
100%|ββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββββ| 2266/2266 [02:04<00:00, 18.18it/s]
Mean cosine similarity:
IXL1<=>NAI: 0.9564101696014404
IXL2<=>NAI: 0.9381213784217834
IXL1<=>Input: 0.9821427464485168
IXL2<=>Input: 0.9513874650001526
NAI<=>Input: 0.9625059366226196
UniversalCLIPbase.safetensors is provided for test purposes and is made leveraging the great work of Felldude.
Please refer to the provided links above regarding usage agreement.