Flux Inverse Sampler

FluxInverseSampler is the plainest sampler in the Fluxtapoz pack, and sometimes that's exactly what you want. It has no inputs, no sliders, nothing to misconfigure. It just runs the ODE forward step by step - a plain Euler-style walk in the denoising direction - and hands you the result as a SAMPLER.

What it's for

The pack's headline inversions (RF-Inversion, RF-Edit, Fireflow) use controlled samplers that steer the trajectory toward a target. FluxInverseSampler is the uncontrolled baseline: no eta, no gamma, no attention bank. Think of it as the "brute force" inversion - run the image latent forward, then try to walk it back with the same prompt. It's genuinely useful as a control in a comparison: does your controlled sampler preserve more structure than this does? If not, your tuning is off. It also gives you a sanity check that the pack's plumbing (ConfigureModifiedFlux, the model-sampling wrappers, the whole inversion rig) is working at all, because the sampler itself can't be misconfigured.

How it works

The implementation is a straight Euler step: at each sigma it calls the model to get the denoised prediction and moves x by denoised * dt. No churn, no ancestral noise, no convergence tricks. On a flow-matching model like Flux the trajectory is near-straight, so this simple walk is actually a reasonable inversion baseline - the same reason Euler-family samplers became the flow-matching default in the first place.

Outputs

One SAMPLER output, meant for the sampler port of SamplerCustomAdvanced. Because the node takes no conditioning or latent inputs, everything about the inversion is determined by whatever else is in your graph: the model patching, the sigmas, the noise schedule. It slots into the same spot in the workflow where the fancier samplers sit, so swapping it in for an A/B test is a two-second drag.

Installation

Standard for the pack - ComfyUI Manager, search "ComfyUI-Fluxtapoz", install, restart. Or:

cd ComfyUI/custom_nodes
git clone https://github.com/logtd/ComfyUI-Fluxtapoz

Only dependency is einshape, no models to download. It runs on your existing Flux checkpoint.

Common issues

There isn't much to break, which is the point. The one thing that surprises people is the same as everywhere else in this pack: it outputs a SAMPLER object, so it wires into SamplerCustomAdvanced, and it needs ConfigureModifiedFlux upstream of the model if you're using it inside one of the pack's inversion workflows. Standalone, on a stock Flux model with stock sigmas, it's just an ordinary deterministic forward pass - which is a perfectly honest thing to want when you're debugging everything else.