Angelo — click to refine

If you've ever generated something 95% right and then burned ten minutes fixing it - save the image, open MaskEditor, paint a mask, wire it back through a second KSampler, requeue, repeat for the next thing - AngeloRefine exists to make you stop doing that. It's a sampler with an interactive canvas bolted onto the front: generate the base image in Sampler Mode, flip to Edit Mode, then click, paint, or drag directly on the node's own preview to fix what's wrong. One node instead of a KSampler plus ADetailer plus MaskEditor plus an outpaint-pad chain, and every fix redraws immediately, in place.

The underlying idea isn't new - it's the same detect-mask-inpaint loop that made A1111's ADetailer beloved and that Impact Pack's FaceDetailer ported to ComfyUI: crop a struggling region, refine it at real resolution, composite back so everything else stays untouched. What's different is where the pieces live. ADetailer and FaceDetailer are separate detect → crop → sample → composite nodes you wire together; Angelo folds the mask, the crop-and-upscale (called Xtra-Fine here), and the composite into one sampler, driven by your mouse instead of a YOLO model. Same payoff - a 64-latent-unit face jumping to roughly 1000 units of working room - with no detector graph to assemble.

Two modes gate everything. Sampler Mode is Angelo acting like a plain KSampler off your Empty Latent - sampler_denoise, sampler_seed, and the seed-control widget live here. Flip Mode to Edit Mode and those grey out; instead you get denoise (per the node's own tooltip: 0.3 for a subtle touch-up, 0.6 for a real redo, 0.9+ to essentially regenerate the region), click_radius for how big a single click's refine area is, and feather_radius to soften the mask edge. inpainting_mode decides how a painted region gets treated: Refine partially denoises whatever's already there and works on any sampler-compatible checkpoint - SDXL, SD 1.5, FLUX 1, anything. Smart Inpaint and Smart Guided Inpaint zero the masked latent and inject reference_latents, which only an edit-trained model can use - FLUX 2 Klein 9B or Qwen-Image-Edit. That's the same reference-latent-plus-InpaintModelConditioning trick people wire by hand elsewhere to get true masked inpainting out of an edit model; Angelo just does it natively when you drag a rectangle or pick a location from a dropdown. Outpaint grows the canvas instead, with a review-before-commit step so a bad extension costs nothing.

Required wiring is unremarkable: model, positive, negative, vae. clip is optional but you'll want it connected - without it, the in-node Area Prompt box and both Smart modes are dead. latent is optional too; skip it and use Load Image instead if you're editing a real photo rather than something you generated. There's also an overrides slot that takes the paired Angelo - Overrides node, for driving steps/cfg/sampler/scheduler from elsewhere in your graph instead of the toolbar. Outputs are image, latent, and source_image - that last one is the original, pre-edit base, useful for wiring straight into a before/after compare node. Most of the remaining schema fields (click_x, stroke_points, rect_points, seg_polygon, and a pile of *_seq counters) aren't things you type - they're state the toolbar JavaScript writes when you click, paint, or drag. Ignore them and use the actual toolbar.

Installing it is a plain custom-node clone: cd ComfyUI/custom_nodes && git clone https://github.com/shootthesound/ComfyUI-Angelo, then restart. No extra Python dependencies for the core node, which is genuinely nice given how much it does. The one opt-in extra is Detect, which lets you type "the face" instead of clicking it - that needs Meta's SAM 3, which isn't on PyPI, so it ships as a separate installer (install_sam3_support.sh / .bat) you run with ComfyUI closed; the ~1GB weights pull down automatically the first time you use it. Skip it if you don't need "type a word, get a mask" - everything else in the node works without it.

Where people actually get stuck: pointing Smart Inpaint or Smart Guided Inpaint at a plain checkpoint instead of an edit model. A base Qwen-Image (not the Edit variant) has the reference-latent code path but was never trained for it, so you get color-distorted output rather than a clean error - regular Refine works on anything, the Smart modes specifically need Klein 9B or Qwen-Image-Edit. Second: ComfyUI's global latent-preview setting can render into the node mid-sample and squash the editing canvas down to nothing - a real, reported annoyance; the fix (disable_live_preview) lives on the companion Overrides node, not here. And go in knowing this is in-process state: undo history and cached refinements live in the running ComfyUI server, not the saved workflow JSON, so a restart clears the canvas back to nothing and you're loading the base image again.