GOD Sampler (Advanced)

The GOD Sampler (Advanced) (class GOD Sampler (Advanced)) is the ambitious one in this pack: instead of walking a fixed noise schedule like every sampler you're used to, it computes a per-step "synergy" score from the current latent - how peaky, how structured, how stable it looks - and uses that score to adapt how fast it denoises at that exact moment. Whether the idea works is an open question; that it's interesting is not.

How it works

It's a k-diffusion-style ODE sampler wrapped in comfy.samplers.KSAMPLER, which is what lets it stand in wherever ComfyUI expects a SAMPLER. At each step it:

1. Calls the model to get the denoised estimate.
2. Derives the noise direction (epsilon) the standard way.
3. Computes a scalar "synergy" S from the current latent - a mix of peak sharpness, gradient orientation stability versus the previous step, and an interior/exterior contrast ratio.
4. Uses S to shrink the next sigma (sigma / (1 + S)), clamped against the scheduled value so it can't run away.
5. Projects deterministically to the next state - no random noise added, so it's reproducible given a seed and schedule.

That "reads its own latents" behavior is the whole pitch: the sampler is reacting to the image as it forms, spending more effort where the structure is still unstable, rather than trusting the schedule blindly.

The inputs and output

Here's the surprising part: there are no inputs. The node is a factory - you drop it on the canvas and it emits a SAMPLER output (sampler) wired into the sampler input of SamplerCustom. If you want a dial, that's the sibling node GOD Sampler (Advanced) Ext., which adds a god_strength control. This one is the fixed-behavior version.

Installing it

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

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

Restart. No dependencies beyond ComfyUI's own sampler machinery (which the pack imports directly).

The honest take

This is where the pack's sampler corner requires the most skepticism. Handwritten samplers are a known genre of "sounds great, barely different from Euler in practice," and this one has zero community signal behind it - no threads, no workflow shares, nobody posting side-by-sides. The math is real and the implementation is careful (the source is well-commented, handles schedule direction, calls the model the way ComfyUI expects), but careful and good are different things.

Because it adds no noise, it behaves like a deterministic ODE sampler: it will converge, and it's reproducible. Test it the right way - fixed seed, same schedule from the pack's GOD Scheduler, then swap in your usual sampler (say, DPM++ 2M Karras on SD 1.5/SDXL, or Euler on Flux) at the same seed and steps, and compare. Expect it to be different; judge whether it's better. And beware the flow-matching trap the KB hammers on: adaptive sigma shaping is exactly the kind of aggressive schedule surgery that flow models tend to dislike. Try it, A/B it, and treat any "it's amazing" claim you hear as unverified until you've seen it with your own seeds.