🅛🅣🅧 APG Guider

Plain classifier-free guidance has an annoying failure mode: crank the CFG to get better prompt adherence and your output oversaturates, contrast blows out, and things start looking fried. APG - Adaptive Projected Guidance - is a smarter replacement that lets you turn guidance up without that penalty. This node is LTX's implementation of it, straight from the paper (arXiv 2410.02416).

How it works

Standard CFG just adds the full "difference between conditioned and unconditioned prediction" back into the sample, and the part of that difference pointing along the current image direction is what over-cooks colors. APG splits that guidance vector and down-weights the component that causes oversaturation, keeping the part that actually improves prompt following. It also adds a momentum term across steps so the guidance is smoother rather than jerky. The result is a GUIDER you feed into a SamplerCustomAdvanced in place of the basic CFG guider - same slot, better behavior at higher guidance.

On LTX this matters because the distilled models run at CFG 1 and the dev models want real guidance to follow a prompt; APG is how you give the dev model more push without the video going garish.

The inputs that matter

• cfg_scale (default 1) - the guidance strength. This is where APG earns its keep: you can go higher than plain CFG tolerates.
• momentum_coefficient (default -0.9) - the smoothing term across steps. Negative by default per the paper; it's what keeps guidance from oscillating.
• eta (default 1) and norm_threshold (default 0) - the projection controls. norm_threshold caps how large a guidance update can get; leaving it at 0 disables the cap. These are the knobs you touch last, once cfg_scale is dialed.

It takes the model, positive, and negative conditioning, and outputs a single GUIDER.

Installing it

• ComfyUI Manager - search LTXVideo, install, restart.
• Manual - cd ComfyUI/custom_nodes && git clone https://github.com/Lightricks/ComfyUI-LTXVideo, then restart.

No model of its own - it wraps the model you already loaded.

Common issues

Colors still blow out at high cfg_scale. Set a norm_threshold (try a moderate value above 0) to actually cap the guidance magnitude - with it at 0 you've turned off APG's safety net.

Output looks flickery or unstable step to step. That's the momentum term. Keep momentum_coefficient near its default; wandering far from -0.9 is where things get twitchy.

No visible improvement over plain CFG. If you're on a distilled LTX model at CFG 1, you're barely guiding at all, so there's little for APG to fix - its benefit shows up when you're actually pushing guidance on a dev model. Match the tool to the situation: APG is for when you need adherence and CFG is fighting you.