🔥 Winnougan LTX NAG Guidance

You write "no text, no watermark, no hands turning to jelly" in your negative prompt, and the distilled LTX-2.3 model shrugs. That's not you doing it wrong - the distilled model was trained without classifier-free guidance, so standard negative prompting barely registers. The community fix is NAG, Normalized Attention Guidance, and the Winnougan LTX NAG Guidance node is the most beginner-friendly way to apply it.

NAG started as a research technique (ChenDarYen's Normalized-Attention-Guidance repo) for guidance-distilled image models, then Kijai ported the idea to LTX-2 as LTX2_NAG in ComfyUI-KJNodes. This node is that same math with a friendlier face: one conditioning input instead of two, plain-English parameter names, and preset strength profiles. It's the difference between patching the attention mechanism by hand and clicking a dropdown.

How it works

Where CFG operates at the noise-prediction level - push the denoised output away from the negative - NAG operates inside the cross-attention layers. It computes attention with both your positive and negative context, then blends them so the negative steer is normalized and clipped by a threshold (nag_tau). The result: the negative prompt actually guides what the model attends to, which is the lever a distilled model can respond to. The node patches the model clone in memory and hands you a model output to wire straight into your sampler. If your LTX-2.3 build has audio, it patches the audio cross-attention blocks too, and nag_info confirms whether that happened.

The inputs that matter

• model - your loaded LTX model. The output model replaces it in the sampler.
• preset - Subtle, Balanced (recommended) (the default), Strong, Maximum, or Custom. Balanced at scale 11 / alpha 0.25 / tau 2.5 really is the right starting point for most generations.
• negative_conditioning - wire the same negative CONDITIONING you send to your sampler. Critical gotcha: leave this unconnected and NAG silently disables itself.
• nag_scale / nag_alpha / nag_tau - only used when preset is Custom. Scale pushes harder, alpha blends guided vs original attention, tau is a clipping threshold where lower is more aggressive.
• inplace - in-place tensor ops to save a sliver of VRAM; it changes numerical results slightly, so leave it off unless you're truly pinned.

Wire it like this:

[LTX Model] → NAG Guidance (model)
[Negative Conditioning] → NAG Guidance (negative_conditioning)
NAG Guidance (model out) → Sampler

Install

It's one of eight nodes in the ComfyUI_WLTX_nodes pack, under Winnougan LTX:

cd ComfyUI/custom_nodes
git clone https://github.com/Winnougan/ComfyUI_WLTX_nodes

Restart ComfyUI, or search ComfyUI_WLTX_nodes in ComfyUI Manager. No extra pip dependencies.

Where people get burned

The number one failure is the empty negative_conditioning input - the node runs, everything looks fine, and you've got plain CFG-1 behavior because there was nothing to steer with. Second: cranking Maximum because you want "more negativity" and getting overcooked, oversaturated output. The presets are conservative for a reason; if Balanced feels weak, bump to Strong, don't jump straight to 25. And since this patches the model in-place in the graph, remember the patched model out must feed your sampler - wiring the original model there silently bypasses NAG entirely.