LC MOSS AnyPaint 🩹

If you're on MOSS Image, you've already hit the wall: no inpaint checkpoint, no inpaint LoRA, no ControlNet with a mask mode. Every guide you find either tells you to use Flux Fill (different model, different look) or to go install LanPaint and wait. LC MOSS AnyPaint is the third option - it fakes masked inpainting on MOSS properly, by keeping your pixels outside the mask at every single step instead of trusting the model to leave them alone.

Why you'd want this instead of an edit model

The one thing masks still own is bit-identical unmasked pixels. An edit model takes the whole frame into context and hands back a whole new frame, so everything nobody asked about drifts - and the drift compounds if you chain two edits. The 2026 answer is crop → edit → stitch, with the stitch protecting the rest of the picture.

This node is that pattern rebuilt around MOSS. It sits in LanPaint's slot - masked conditioning for a base that shipped without an inpaint variant - except it trades extra sampling for a hard keep: your untouched region is written back after every step, token-exact.

How it works

Say yes to crop_to_mask (the default) and your mask gets its own crop, plus padding px of context around it, scaled up to inpaint_resolution. Then:

• The hole is pre-filled by smearing the surrounding picture into it (OpenCV Telea fill), not with noise. MOSS starts from something plausible instead of white mush.
• A per-token keep mask is built. A latent token survives only if all of its pixels sit outside the painted area grown by boundary_px. MOSS latents are one token per latent pixel, so this is about as tight as you can get.
• Sampling runs at full denoise with a real negative and real CFG - the thing SDXL users miss most on Flux-era models - and skip_layer runs ComfyUI's skip-layer guidance on layer 22 for the first half of the steps - MOSS's own recommendation, and what keeps fine detail from quietly falling apart.
• On the way out, only the painted area plus half the border is pasted back with a soft edge, and colour drift measured in the outer border is corrected - the same seam fix the pack's Smart Detailer uses. Keep composite on keep original; raw hands you the whole redraw of the crop, which is for debugging.

The inputs that actually matter

model, clip and vae have to be MOSS Image's own trio - the class check will refuse anything else, and the encoder is Qwen3.5-9B with CLIP type moss_image. Then image and mask (white paints). The mask is genuinely optional: leave it off and wire LC Outpaint's control_image / control_mask in and the node outpaints the opened-up canvas instead.

prompt is where most people lose an afternoon. Describe only what goes into the mask. Write "a black sign reading MOSS on a rainy window" and you get a sign; write a description of the whole photo and MOSS paints a second person into the crop, because that's what you asked for. Want text in the image? Put it in quotes.

If you're on the 8-step turbo LoRA: steps 8, cfg 1, skip_layer off. Otherwise 25 steps, cfg 2.5, er_sde / simple. boundary_px stays at 8 - wider borders come back soft and leave a halo. And small text needs the full step count; the turbo LoRA garbles it.

Outputs are image and mask, where the mask marks what was actually pasted back, so you can chain another pass or hand it to the pack's refine nodes.

Installing it

ComfyUI Manager → search ComfyUI LC MaskMaker Nodes, or:

cd ComfyUI/custom_nodes
git clone https://github.com/lonecatone23/ComfyUI_LC_MaskMaker_nodes

Restart. There's no requirements.txt; the pack's install.py only pip-installs transformers, huggingface_hub and timm if they're missing, and it never touches torch, numpy or pillow. Nothing is bundled or auto-downloaded for this node - MOSS's weights, encoder, VAE and the optional turbo LoRA come from wherever you already got them.

Where it goes wrong

The prompt trap is the big one: a flat blob where a car used to be looks like a broken node rather than a prompt that described the wrong thing. Tell it what should be there instead.

crop_to_mask on for a small inpaint is right, but for outpaint the mask runs along the frame edge and the crop ends up being most of the picture anyway, so the siblings in this pack document turning it off. And inpaint_resolution above roughly 1.6 MP is where this family starts creasing limbs - it's an area, not an edge length, so 1024 means about a megapixel.

One caveat: MOSS Image has almost no community footprint yet, so there's no folklore for its settings to check your sanity against - the tooltips here are effectively the documentation. Follow them, and don't get creative with CFG until you have a result you like.