Flux Erase Image

There's a specific annoyance in the "remove the thing" category that inpainting is overkill for: you don't want to replace the masked area with something new, you want it to have never been there - the watermark off the corner, the tourist out of the shot, the trash can out of the frame. Flux Erase is BFL's API tool for exactly that. It's the newest member of the Flux image family in core (landed June 2026, alongside Flux Virtual Try-On), and unlike Fill it has no prompt input at all. There's nothing to describe; you just point at what dies.

Like every node in the partner/image/BFL group, this is a paid remote call: your image and mask go to api.comfy.org as base64, BFL's servers run the erase model, and the credits come off your Comfy account balance. No model files, no VRAM, no login-free operation. The price is shown as an approximate range on the badge because BFL bills it by input size.

The mask is where the semantics matter, and the tooltip says it plainly: white areas are removed; black areas are preserved. That's the opposite of a fill mask's "white = what to regenerate" convention in spirit - you're marking deletion, not replacement. The node resizes the mask to match your image, so rough painting is fine.

The one input you'll actually fiddle with is dilate_pixels, default 10. It expands the mask outward (0–25px) before the job runs, so it eats the fuzzy edges of whatever you painted - hair, motion blur, soft shadows - that you'd otherwise leave as a halo. If you see a tell-tale ring of the erased object left behind, nudge it up. Too high and it starts chewing into things you wanted to keep; 10 is a good starting point, adjust by feel. seed just controls rerun behavior; results are nondeterministic anyway.

A couple of grounded gotchas. Your image must be at least 256×256 in both dimensions or the node refuses to run. And the mask is binarized before sending, so don't bother with soft feathered alpha in your paint - areas painted at less than half opacity get ignored. Also note the erase model reconstructs the background plausibly, not truthfully: on a busy background it will invent content that fills the hole. For a clean wall or sky that's invisible; for a crowd scene it can look a little wobbly.

Output is a single IMAGE tensor, already decoded - save it, or feed it onward. The honest framing: locally, the same job is a mask + FLUX.1 Fill at low-to-zero prompt, or even SDXL inpainting, all free. Flux Erase wins on convenience - one node, no denoise tuning, no seam blending - and you're paying BFL to do that fiddling for you. For occasional cleanups it's a delight. For a thousand-product batch, do the math on credits before you commit.