Outverse Flux Model Pred

OutFluxModelSamplingPred is the sibling of InFluxModelSamplingPred and the node that sits on the resample side of an RF-Edit or FlowEdit graph - the half that turns your edited noise back into a clean image. Where the inverse wrapper treats the model output as-is, this one restores the normal "predict the clean sample" interpretation, which is what a forward generation pass expects.

What it changes

It builds the same ModelSamplingFlux base with the same shift math as its inverse counterpart - max_shift (1.15), base_shift (0.5), width and height (1024) - but swaps the sampling type. By default it uses the OutCONST regime, where calculate_denoised recovers the clean sample as model_input - model_output * sigma: the standard Flux flow-prediction interpretation. There's one extra toggle:

• reverse_ode (BOOLEAN, default false) - flips it into the ReverseCONST regime instead, which keeps the model output raw. You turn this on when the reverse pass is being run as part of an ODE reversal, which is the Fireflow/RF-Edit reversed path.

So the node is really "the resample-side model wrapper, with a switch for which direction the ODE is going."

Outputs

One MODEL. In the RF-Edit example workflow it's placed on the decode/generation branch opposite the inverse wrapper, and in the FlowEdit example it's the wrapper in front of the FlowEditGuider. Either way, it needs ConfigureModifiedFlux upstream so the patched attention blocks are in place.

Installation

Standard for the pack: ComfyUI Manager → search "ComfyUI-Fluxtapoz" → install → restart, or git clone https://github.com/logtd/ComfyUI-Fluxtapoz into custom_nodes. No downloads; einshape is the only dependency.

Common issues

• Using the same wrapper on both ends. The inverse and outverse nodes are a matched pair for a reason - each side's prediction regime is the mirror of the other. Swap them, or use one for both, and the structure of the original image won't survive the round trip.
• Wrong resolution inputs. The shift is computed from width × height. If you loaded a 512×768 image but left the defaults at 1024×1024, the schedule is tuned for a resolution you're not using. Keep width/height in sync with your actual latent.
• reverse_ode confusion. It defaults to false because the ordinary resample path doesn't need it. Only flip it if you're following a Fireflow-style reversed ODE graph; on a normal FlowEdit graph it's left alone.

It's a plumbing node, but it's the plumbing that decides whether your edited image comes back sharp and stable or as structured noise.