Q: How are each of these CFG hacks different from one another?

[TomLucidor]

Are there explainers on how each of these methods vary in quality?

• Smoothed Energy Guidance
• Semantic CFG
• Perturbed Attention Guidance
• CFG Interval / CFG Scheduler ("Clamp")

Bonus: what about these other new methods?

• Dynamic Thresholding ("X mimic Y clamp Z%") https://github.com/mcmonkeyprojects/sd-dynamic-thresholding
• CFG++ or "manifold-constrained" https://github.com/CFGpp-diffusion/CFGpp
• Skimmed CFG https://github.com/Extraltodeus/Skimmed_CFG
• CFG Rescale https://github.com/Seshelle/CFG_Rescale_webui https://github.com/MythicalChu/CFGRescale_For_Forge
• APG Scaling https://github.com/logtd/ComfyUI-APGScaling

[6DammK9]

Currently moving from "FreeU + DynCFG + PAG" to "CFG++ + PAG + SEG". You can see my findings in my repo.
Notice that I have agressive parameters so it may not match to "general usage".

Currently CFG++ is a new topic to explore, there is an "optimimal STEPs" to search, even it should be a common value coincidentally.

Trust me, I have a model very capable in this topic, unconditional image generation is a deep topic, and I believe in merging.

[TomLucidor]

Pretty lit @6DammK9 (wishing for an ELI5-ish article but oh well ChatGPT time)... Wait I think I missed these ones

• Characteristic Guidance https://github.com/scraed/CharacteristicGuidanceWebUI
• Particle Guidance https://github.com/gcorso/particle-guidance
• "Attend and Excite" (not even sure if this is relevant) https://github.com/yuval-alaluf/Attend-and-Excite

[6DammK9]

Characteristic Guidance = CHG. Tested, too long for inference. Roll 10 seeds instead.
Next 2 has no "UI" integration so no comment.