Epsilon (recipe) - Rotate the previous model with comradeship v14T13A because both use 0.5 as a base.

V-prediction (recipe) - Added this epsilon merge on a v-prediction itercomp merge using perpendicular component.

The tool used is sd_mecha.

You can use these merges with LoRAs trained on their base, for improved results.

• Epsilon 0.5

• V-prediction test

LoRA settings I use can be found here: epsilon, v-prediction

(Settings are compatible with this UI)

(You have to train the LoRA on the base model, not these merges)

Recommended method for high-res:

Scaling: 1.5x (optional latent upscaler)

Model: 4x-NomosUni-span-multijpg

0.5 strength with this yolov8 model

I recommend using a CFG around 4 or 5.