Flow Edit Guider

FlowEditGuider is the steering wheel for the pack's FlowEdit implementation - an inversion-free way to edit images. Everything else in this pack (RF-Inversion, RF-Edit, Fireflow) unsamples your image to noise and resamples it back. FlowEdit never does that. It runs the whole denoising process under two sets of conditioning and switches between them partway, which is a lot less work and a lot less failure surface.

How it works

FlowEdit (the paper is "FlowEdit: Inversion-Free Text-Based Editing Using Pre-Trained Flow Models") exploits the straightness of flow-matching trajectories. Since Flux's path from noise to image is close to a straight line, you can define a correspondence between the source image and the target edit, then generate by walking from noise while interpolating the conditioning. The guider holds two conditionings:

• source_cond (CONDITIONING) - the prompt (or conditioning) describing the original image.
• target_cond (CONDITIONING) - the prompt describing what you want it to become.

During sampling it applies the source conditioning for the early steps and the target conditioning for the rest - the exact switch point is determined by the FlowEditSampler's skip_steps, not by this node. You also feed it model (MODEL), your Flux model with ConfigureModifiedFlux and OutFluxModelSamplingPred already applied.

Output

One GUIDER, which plugs into a SamplerCustomAdvanced (or a BasicGuider slot) in place of the normal guider, paired with FlowEditSampler. The node sets its own CFG to 1.0 internally - this is a flow model doing what flow models do, and CFG guidance is not part of the equation.

Installation

Standard for the pack: ComfyUI Manager → search "ComfyUI-Fluxtapoz" → install → restart, or git clone https://github.com/logtd/ComfyUI-Fluxtapoz into custom_nodes. Only einshape, no models to download.

Common issues

• Edits that barely change anything usually mean the switch to target conditioning is happening too late - raise skip_steps on the FlowEditSampler so the edit gets more of the schedule.
• Edits that throw away the source entirely are the opposite: too much of the schedule is running on target conditioning. This is a two-node tuning problem (guider holds the conditionings, sampler holds the switch point), and it's worth treating them as one unit.
• Since there's no inversion, there's no latent to carry the source's structure - the correspondence is built from the flow itself. That's what makes it fast, and also what makes it less appropriate for edits where you need pixel-level faithfulness to the original. For looser semantic edits it's genuinely the easiest path in the pack.