RunningHub OneReward Eraser

Object removal, minus the part where you describe what goes in the hole. The OneReward Eraser is the pack's Sampler with the prompt input deleted and the word remove hardcoded in its place. You mask the photobomber, the watermark, the stray cable in the frame, and it figures out what belongs behind them on its own. Sometimes that's exactly what you want - no prompt engineering, no "empty beach, no people" phrasing to get right. Just mask and erase.

It works because it is the Sampler. The Eraser class subclasses the Sampler and overrides only the sampling call, swapping your prompt for a fixed "remove" and keeping everything else identical: the same OneReward pipeline input, the same mask handling, the same true-CFG denoising loop, the same hardcoded nsfw negative prompt (still there, still invisible, still harmless). So all the mechanics from the Sampler article apply - true_cfg past 1.0 runs the doubled unconditional/conditional batch and takes roughly twice as long per step, and 50 default steps on a quantized consumer card is a wait. Crank true_cfg toward 1.5 and drop steps to 20–30 if you're impatient.

The inputs you'll actually touch:

• mask - the whole job. OneReward was trained on mask-conditioned object removal, so a clean mask around the target usually does the rest. Feather it a little to avoid a hard seam; paint a bit of margin so it removes the whole object, not a crisp slice of it.
• seed - lock it while you compare results. With the prompt fixed, seed and true_cfg are your only real dials.
• num_inference_steps / guidance_scale - leave the defaults until something looks off.

guidance_scale is FLUX's native guidance embedding, not CFG; the CFG dial is true_cfg (default 4.0, which is a lot - back it down if fills get noisy or overworked).

The one output, image, is the whole frame with the masked area regenerated. Same caveat as the Sampler: this is whole-image inpainting, so if the pixels outside the mask must survive exactly, composite the output back over the original yourself - nothing in this pack does the crop-and-stitch trick for you.

Where the Eraser earns its keep, and where it doesn't: pure removal with no strong idea of what's behind the object, it's the node you reach for. The moment you do have an opinion about the background - "this is a cafe, I want the table and nothing else" - use the Sampler instead, because the fixed "remove" prompt is exactly the kind of under-described fill that makes a model invent clutter. That's the one real limitation here, and it's the author's trade-off, not a bug.

Install and models are the pack's shared story: clone https://github.com/HM-RunningHub/ComfyUI_RH_OneReward into custom_nodes (or search "ComfyUI OneReward" in Manager), download the ~30 GB of FLUX.1-Fill-dev plus OneReward weights into the exact folders the Loader expects, and restart. The repo ships no requirements.txt despite the README saying it does, so make sure diffusers and optimum-quanto are in your environment. Mask, click, and the cable in your product shot quietly stops existing.