Adaptive Projected Guidance for Forge2 webUI

APG code from ELIMINATING OVERSATURATION AND ARTIFACTS OF HIGH GUIDANCE SCALES IN DIFFUSION MODELS

credit to Seyedmorteza Sadat, Otmar Hilliges, Romann M. Weber

https://arxiv.org/pdf/2410.02416

---

controls described in the paper

eta affects saturation/contrast

rescale threshold ... highly variable depending on model, experiment

momentum between -0.75 and -0.25 recommended

---

The paper mostly compares APG with too-high CFG, which I'm not sure is all that useful - we already know high CFG burns badly, so a better comparison would be between APG and reasonable CFG to see if we really do get better quality, better prompt adherence, etc. Anyway, it does something.

sdxl model haveall cfg3.3

sd1.5 model swizz8real cfg 15

---

Also,

TraSCE: Trajectory Steering for Concept Erasure https://arxiv.org/abs/2412.07658

Intent from the paper is to improve the ability of the negative prompt to remove/control concepts: using a negative is necessary - no effect if the negative prompt is empty.

Basically, cfg_result = empty_conditioning + guidance * (positive_conditioning - negative_conditioning).

Standard is cfg_result = negative_conditioning + guidance * (positive_conditioning - negative_conditioning).

---

Also, some other experimental stuff. And SkipLayerGuidance for SD3.