A1111 Style Prompt (Negative)

This is the negative twin of the A1111 Style Prompt node from the same pack. It does exactly what you'd guess: you feed it your negative text in A1111 syntax - weights, BREAK, wildcards - and it produces a conditioning output you wire into the negative slot of your sampler, next to the positive node's output. Same look, same inputs, same encoding logic.

But there's one rule the README is upfront about, and it will save you a wasted afternoon: scheduling and alternation don't work here. The main node's party trick - [forest:city:0.5], [white|black] - requires a per-step hook that only registers on the positive conditioning path. The negative node detects that syntax, quietly uses the first step's prompt, and logs an informational message. So 1girl, [a|b], bad hands doesn't error out; it just encodes 1girl, a, bad hands and moves on. Keep your negative prompt static and you'll never hit it.

Inputs and outputs

| Input | Type | What it does | |---|---|---| | clip | CLIP | The same CLIP you feed the positive node. Required. | | text | STRING (multiline) | Your negative prompt. A1111 weights and wildcards work; scheduling doesn't. | | normalization | BOOL (default off) | The Norm/No Norm toggle, same meaning as the positive node. Leave it off unless you're on SD1.5 and chasing saturation shifts. | | debug | BOOL (default off) | Prints the resolved prompt to the console. |

Output is a single conditioning, wired to the sampler's negative input - nothing else changes in your workflow.

Install and gotchas

Installation is the pack's, so this part is shared: ComfyUI Manager, search "A1111 Prompt Node" (pack Enferlain/ComfyUI-A1111-cond), or

cd ComfyUI/custom_nodes
git clone https://github.com/Enferlain/ComfyUI-A1111-cond
pip install lark

then restart. lark is the only non-trivial dependency; no model downloads.

Two grounding facts worth remembering, straight from how the ecosystem actually works:

A negative prompt only does anything when CFG is above 1. It's not a filter - it replaces the empty string in the unconditioned pass of classifier-free guidance, and at CFG 1 that pass is skipped entirely. On distilled models (Turbo/Lightning checkpoints, most 2026 releases) the negative box does nothing at their default settings. This node is for CLIP-era models - SD1.5, SDXL, Illustrious, NoobAI - running real CFG, where negatives genuinely pull quality.

Don't inherit the forty-keyword boilerplate. worst quality, low quality, bad anatomy, ... mostly fails an A/B test on a fixed seed. Name the specific defect you can see in your output, or a coherent source region ("deviant art"). Fewer, targeted negatives beat the wall of text, and they keep your token count in the green.

It's a small node with a single gotcha. If you're already using the positive A1111 Style Prompt, add this one for symmetry and move on - just keep the negative static.