To clarify: this is an epsilon-prediction model, although some of the model used in merging is v-pred

V1.0_fix removes the v-pred indicator (an empty tensor), other than that it is identical with V1.0. If you can run V1.0 without problem, you don't have to redownload it, though you may need it if you upgrade your software to later version. Sorry for any inconvenience caused.

Alternatively you can enforce an epsilon-pred sampling to run V1.0, for example, you can use the following setting in reForge:

If you can understand and run python code, here is another simple way how you can fix it:

from safetensors import safe_open
from safetensors.torch import save_file
data = dict()
file_name = '/path/to/downloaded/CatWai.safetensors'
with safe_open(file_name, framework='pt') as model_file:
    for key in model_file.keys():
        if key != 'v_pred':
            data[key] = model_file.get_tensor(key).clone()
save_file(data, '/path/to/save/CatWai_fix.safetensors')

This is my first checkpoint merge, an experimental attempt to turn a V-prediction model into eps-prediction model.

It's mainly built on catTowerNoobaiXL_v11Vpred (v-prediction) and the final merge is in eps-prediction format so that you can use it in vanilla webui.

It was achieved by gradually shift its output layers to other eps models. The chosen model to replace those parts was (90% waiNSFWIllustrious_v70 + 10% ebara_pony_2.1). The converted model was further merged with waiNSFWIllustrious_v70 at 9:1 ratio so roughly speaking, it's about 70% CatTower + 28% wai + 2% ebara, that's why I call it CatWai.

The merge was done in ComfyUI. You can drag the model file or the json workflow into ComfyUI to see the detailed merge work flow. The json can be found in the zip uploaded as "training data". With the ComfyUI workflow you can save 6GB downloading if you have the mentioned 3 models on disk : ) The text version recipe is also included below.

Many thanks to source models' authors, especially the catTower model which is so nicely crafted that makes me excited on AI painting stuff again! The license for this merge is the same as catTower one as required: commercial use is prohibited.

Originally I wanted to post this one as some patch LoRA on catTower but it failed. I did get something close but the quality drop is significant even at dim=conv_dim=512.

Settings

I'm new to IL / Noob so this setup may not be optimal.

Positive Prompt:

masterpiece, best quality, absurdres, safe

Negative Prompt:

lowres, worst quality, low quality, bad anatomy, bad hands, 4koma, comic, greyscale, censored, jpeg artifacts, overly saturated, overly vivid

Sampling method: Euler

Sampling steps: 30

Resolution: personally I recommend 1024 x 1536 but people use 832 x 1216 more often

CFG Scale: 5

Clip skip: 2

Recipe

A = 0.9 * waiNSFWIllustrious_v70 + 0.1 * ebara_pony_2.1
B = x * catTowerNoobaiXL_v11Vpred + (1 - x) * A
where x is [1.0, 1.0, 1.0, 0.7, 0.5, 0.25, 0.1, 0.0, 0.0, 0.0] for output_block 0-8 and remaining out blocks. x is 1.0 for other blocks (i.e. input_blocks + mid_blocks + others)
result = 0.9 * B + 0.1 * waiNSFWIllustrious_v70

License

License is the same as catTower: commercial use is prohibited.