LayerUtility: SD3 Negative Conditioning(Advance)

This node only matters if you're running Stable Diffusion 3 (or 3.5). SD3 has a well-known quirk: apply a negative prompt across the whole denoising process the normal way, the way you would on SDXL or SD 1.5, and you tend to get washed-out, over-corrected results instead of a clean negative effect. ComfyUI's own official SD3 example workflows work around this with a specific combination - zero out the negative conditioning, then only let the real negative kick in over a limited range of the later timesteps, rather than the full denoising pass. This node is the README's own words: "Encapsulate the four nodes of Negative Condition in SD3 into a separate node" - it's that same fix, collapsed from a four-node combo into one, so you don't have to rebuild it by hand every time you wire up an SD3 graph.

The inputs and outputs that matter

• conditioning - your negative conditioning, same as you'd plug into any KSampler's negative slot.
• zero_out_start - where the split happens between "zeroed out" and "actually applied," on a 0–1 scale of the denoising range, default 0.1. Below this point the negative is zeroed (has no effect); the README notes this value doubles as the end value you'd otherwise set on a separate ConditioningSetTimestepRange node for the negative - so raising or lowering it moves how much of the early denoising process the negative sits out.

Output is a single CONDITIONING, ready to wire directly into your sampler's negative input in place of your raw conditioning.

How to install it

Recommended: ComfyUI Manager, search "ComfyUI Layer Style Advance". Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/chflame163/ComfyUI_LayerStyle_Advance.git

Then install_requirements.bat / install_requirements_aki.bat, or pip install -r requirements.txt plus repair_dependency.bat. Restart. No model download involved - this is pure conditioning math, not something that touches VRAM beyond what your sampler already uses.

Common issues & troubleshooting

You wired this into a non-SD3 workflow (SDXL, Flux, SD 1.5) and it did nothing useful, or made things worse. This node exists to fix an SD3-specific behavior. Other architectures don't share the same negative-conditioning quirk, so there's no reason to use it outside SD3/SD3.5 graphs - a plain negative-conditioning input is the right choice everywhere else.

Negative prompt still isn't landing. If zero_out_start is set too high, you're zeroing out most of the denoising range and leaving the negative almost nothing to act on. Bring it down toward - or below - the 0.1 default and see if the effect returns.

Back to the original washed-out problem. The opposite failure: zero_out_start too low means the negative is active across almost the whole process again, which is the exact behavior this node exists to avoid. There's no single correct value stated anywhere - treat 0.1 as a starting point and adjust per prompt, the same way you'd tune CFG.