GOD Sampler (Advanced) Ext.

The GOD Sampler (Advanced) Ext. (class GOD Sampler (Advanced) Ext.) is the sibling of the plain GOD Sampler, and the difference is exactly one input: god_strength. Where the base version has a fixed behavior baked in, this one lets you dial how aggressively the sampler's latent-reading "synergy" adjustment steers the denoise schedule. If you're going to experiment with the GOD idea at all, this is the version to do it with - it gives you an off switch and a control group.

How it works

The machinery is shared with the base GOD sampler: at each step it calls the model, derives the noise direction, computes a synergy scalar S from the current latent's structure, and adapts the next sigma. The Ext. version adds two things on top. First, a gain (that's your god_strength) scales how much S can shrink the next sigma: sigma / (1 + gain * S). Second, instead of clamping to the scheduled value, it blends the adapted and scheduled sigmas 50/50, so the adjustment is always partial and can't fully override the schedule.

The range on god_strength goes 0 to 100 with default 1. The 0 setting is the one worth remembering: it makes the synergy term vanish, so the sampler degenerates toward following the plain schedule - effectively your control run. Past that, higher values make the adaptation more aggressive. The step is 1.0, which is coarse; the interesting territory is almost certainly the low single digits.

Inputs and output

• god_strength - FLOAT, default 1, min 0, max 100. 0 disables the adaptation; crank up from there to taste.

Output is a single sampler (SAMPLER) for the sampler input of SamplerCustom. Same as the base node, nothing else.

Installing it

It's in the Switchblade Pack: ComfyUI Manager → search MilitantHitchhiker-SwitchbladePack → install → restart, or:

cd ComfyUI/custom_nodes
git clone https://github.com/MilitantHitchhiker/MilitantHitchhiker-SwitchbladePack

Restart ComfyUI. No downloads, no Python deps.

The honest take

This is the node the pack should have shipped first: a strength knob is exactly what a speculative sampler needs, because it turns "does this idea help?" into a measurable question. The right experiment is boring and worth doing: same seed, same schedule, run god_strength at 0, 1, 3, and maybe 10, and diff the outputs. If 0 and 1 are indistinguishable, the synergy math is doing nothing at your resolution and step count. If they diverge wildly, treat the high settings with suspicion rather than excitement.

The provenance caveat applies here with extra force. Zero community signal, no writeups, no comparisons - this is one author's unpublished hypothesis about what "detail-aware" sampling should look like, wrapped in a very presentable node. The flow-matching caveat from the broader sampler debate applies too: adaptive sigma shaping is the kind of schedule surgery that tends to hurt Flux-class models, which prefer conservative schedules. On SD 1.5/SDXL it's a fun, cheap experiment with a built-in control group. Treat it as a science kit, not a solution.