Automatic CFG - Negative
Give your negative prompt its own strength dial
- model
- MODEL
The plain Automatic CFG node rescales guidance so nothing burns. This node takes that same engine and adds the one knob people keep wanting: a negative_strength slider that treats your negative prompt as its own adjustable quantity instead of a fixed "everything else" target.
Normally your negative prompt is all-or-nothing: CFG pushes away from whatever you typed, at the same strength as it pulls toward your prompt. This node instead lerps the positive and negative predictions toward each other by a factor - below 1 the negative gets weaker (blurrier, more forgiving), above 1 it gets pushed harder than the positive. The author borrowed the idea from the negative prompt weight project, and it's a "your CFG won't be your CFG anymore" situation: because the automatic rescaling is underneath it, cranking the negative past 1 doesn't instantly deep-fry the image like it would on a raw sampler.
The inputs that matter
- negative_strength - the dial. Range 0 to 5, default 1. At 1 it behaves like the standard node with a normal negative. Above 1 (try 1.5–2.5) the negative's influence is amplified; below 1 it's softened toward a "prompt only" look. One note from the update history: the scale was halved in a 2024 rework, so if you're coming from an older version, double what you remember.
- boost - on by default. The same low-sigma negative-skip as the main node: once sigma drops below ~1 the negative pass is skipped for speed. Off if you want the negative running the full render.
- model - the MODEL to patch, output wired into your KSampler.
Output is a single patched MODEL. Set the sampler's CFG around 8 like the rest of the pack and treat the node's own slider as the thing you actually tune.
When you'd reach for it
Two situations. First: your negative prompt is doing more harm than good and you want a softer-than-default negative without deleting the text - nudge below 1. Second, the more interesting one: your prompt keeps drifting toward an artifact the negative names but can't fully suppress (extra fingers, duplicate characters), and raising CFG burns the image. Crank negative_strength up instead; the anti-burn rescaling keeps you out of the deep-fry zone while the negative bites harder. Same trust rules as the whole pack - it patches ComfyUI's sampling function globally, so update ComfyUI before debugging anything that misbehaves.
cd ComfyUI/custom_nodes
git clone https://github.com/Extraltodeus/ComfyUI-AutomaticCFG
Restart and it appears under model_patches/Automatic_CFG/presets. It's the friendliest way to explore "what if my negative were a separate control" without opening the Advanced node - one slider, one toggle, done.
Inputs (3)
| Name | Type | Default | Description |
|---|---|---|---|
| model | MODEL | — | |
| boost | BOOLEAN | true | — |
| negative_strength | FLOAT | 1.00–5 | — |
Outputs (1)
| Name | Type | Description |
|---|---|---|
| MODEL | MODEL | — |