RF-Edit Single Layers Override

Of the two layer-override knobs in the Fluxtapoz RF-Edit workflow, this is the one you'll actually touch. RFSingleBlocksOverride ("RF-Edit Single Layers Override") gives you 38 switches - one per single block in Flux's transformer - and the author's defaults already have 18 of them on. It's the main dial for the thing people complain about most with image editing: "the style changed, but so did my subject's face."

The mechanism is shared with its sibling RFDoubleBlocksOverride, so the setup is the same story. RF-Edit is logtd's implementation of RF-Solver-Edit. The RF-Edit Forward Sampler unsamples your image to noise and, during its final save_steps, copies attention values from whichever layers you've switched on into an attention bank. The RF-Edit Reverse Sampler then regenerates with your new prompt and re-injects those exact values during its first inject_steps. Switch on = that layer's attention is pinned to the original, so its structure survives the edit. Switch off = the layer regenerates freely and the prompt gets more room to act.

Why single blocks are where the likeness lives: after the 19 double blocks (image and text mixed together) come 38 single blocks that refine the image stream on its own - local shape, edges, the detail that decides whether the person in the output still looks like the person in the input. The defaults enable blocks 20–37 and leave 0–19 off. That's the author's tuned starting point: the deeper single blocks carry the composition worth preserving, the shallow ones stay free so the edit has room to move.

That maps directly onto the classic RF-Edit complaint. A common scene: you're restyling a photo with a style LoRA and the workflow nails the look but quietly swaps your subject's features - people describe it as "I want at least 90% likeness." The answer is exactly this node. If the features drift, flip on the shallow single blocks (0–19) that default to off; each one you enable pins more of the original's local structure in place. If the result feels stiff and uneditable, flip the deeper ones off instead. Start with a couple at a time, not all eighteen at once.

Wiring it up: the output, SINGLE_LAYERS, feeds the single_layers input on both the RF-Edit Forward Sampler and the RF-Edit Reverse Sampler. Run the same node's output to both. If the forward and reverse passes disagree on the layer set, the reverse silently skips whatever wasn't banked - you think you configured preservation and it quietly isn't happening.

Installing it is just installing the pack. In ComfyUI Manager, search "Fluxtapoz", or:

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

Then restart ComfyUI. The only pip dependency is einshape; the real requirement is a working FLUX.1 dev stack, since none of this does anything without the model files.

Two gotchas worth knowing. First, if you toggle switches and see zero difference, check save_steps and inject_steps before you blame the node - injection only happens inside those windows, and if inject_steps is set higher than save_steps the reverse sampler raises an error telling you to save at least as many steps as you inject. Second, don't expect the example workflows to use this node - they run on the sampler defaults, so this is purely an optional tuning layer for when the default preservation balance isn't right for your image.