Configure Modified Flux

ConfigureModifiedFlux is the unglamorous foundation of the whole ComfyUI-Fluxtapoz pack. It takes a model in, gives you the same model out, and in between swaps Flux's transformer blocks for patched versions that the rest of the pack's nodes hook into. Every RF-Inversion, RF-Edit, Fireflow, FlowEdit and regional-prompting workflow in the repo's example_workflows runs this node first. If you're following one of those and something isn't doing anything, nine times out of ten it's because this node got missed.

What it actually does

Flux is a DiT (diffusion transformer), and every editing trick in this pack - attention-mask patching for regional prompting, query perturbation for PAG, gaussian-blurred queries for SEG, attention-bank injection for RF-Edit - is implemented as a patch on the attention blocks. Stock ComfyUI doesn't expose those hooks. ConfigureModifiedFlux calls inject_flux() and inject_blocks(), which replace the DoubleStreamBlock and SingleStreamBlock classes with subclasses that honor mask_fn, post_q and the rfedit transformer options those nodes rely on.

It's a bit like installing a tap on a pipe: the water is the same, but suddenly the other tools have somewhere to attach.

Inputs and outputs

The node is deliberately boring. One required input:

• model (MODEL) - your Flux checkpoint as it comes off the UNETLoader or after a LoRA stack.

And one output:

• MODEL - the same model, with the patched blocks.

That's it. No sliders, no strength, no gotchas. Put it immediately after your model is loaded and before anything else from this pack. It doesn't hurt to leave it in a plain txt2img workflow either, but it's doing nothing for you there.

Installation

Same story for every node in this pack. ComfyUI Manager: search "ComfyUI-Fluxtapoz" (by logtd), install, restart. Or the manual route:

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

Then restart ComfyUI. The only Python dependency is einshape, which installs itself. There are no model files to download - you bring your own Flux checkpoint, T5 text encoder and VAE.

Common issues

• Everything else is wired up and nothing changes. Check the model path: the node that actually applies the effect needs the patched model output from here, not the raw one from the loader. Trace the MODEL line and confirm ConfigureModifiedFlux is upstream of it.
• "This only works on Flux." Correct. It's patching Flux's specific transformer layout; the same graph on an SDXL or Z-Image checkpoint either errors or silently ignores the patches. This is a Flux pack, full stop.
• The node reports as a no-op in output (same MODEL type out), which makes people suspect it's broken. It's not - the visible work happens one or two nodes downstream.

It's the most boring node in the pack and the one you can't skip. Wire it in once and stop thinking about it.