ComfyUI LC MaskMaker Nodes

ComfyUI LC MaskMaker Nodes

Mask, matting, segmentation and image-adjustment nodes for ComfyUI by lonecatone23.

• Repo: https://github.com/lonecatone23/ComfyUI_LC_MaskMaker_nodes
• Civitai: lonecatone23
• Support: Buy me a ☕
• Version: 0.19.1 · 19 Python nodes

Companion to ComfyUI_LC123_nodes. LC123 stays import-free. Anything that needs a model lives here and loads it on demand.

Release history lives in git tags. This page describes the pack as it is right now, not a changelog.

---

Nodes

| Node | What it does | |---|---| | LC Outpaint 🖌️➕ | Drag the canvas edges outward on the node. Drag inside the source to reposition it. free or locked to a common aspect ratio, grey fill. Outputs control_image, control_mask (white = new area), mask_image, width, height. hold_mask: held keeps the canvas you set up, run after run. snap_to_image passes the image through as is while nothing is pulled out; pull the edges out and run again and it continues with them, then the box snaps back to the image only after the whole run has finished (so a new image always starts clean, and a failed or cancelled run keeps your edges). Switching to snap_to_image snaps the box right away. block set to if_empty_mask stops everything downstream while nothing is pulled out: drag the edges out, run again, and it continues. Run once to get the preview. note: snap_to rounds the final size up to a multiple (8 by default), set it to 1 to turn that off. | | LC Image Rotate + Pad 🔄 | Turn, crop and pad a picture by dragging on the node. The round handle turns it (hold Shift for 15° steps, or type the angle), the frame's square handles pad (drag out) or crop (drag in), and dragging inside the frame moves the picture. The control_mask is white where the model has to paint: the padding and the corners a turn opens up, so it goes straight into LC Krea2 AnyPaint. fit: expand starts the frame around the whole turned picture, crop at the largest box with no empty corners. fill: gray (what Krea 2 and AnyPaint expect), edge (stretches the outer pixels), white, black. aspect locks the output shape, snap_to rounds the size up. Straight and quarter turns copy pixels, never resample. block: until turned or framed stops the run after this node while the picture is untouched, so the first run only puts the picture on the node; never always passes it on. keep_frame: keep (same image) holds the turn and frame run after run (post processing the same picture), snap back after run resets them once a run finishes so the next image starts clean. Run once to see the picture on the node. | | LC Face Crop 📐 | Crops to a format with the faces placed well, not a blind centre crop: one render, every social format, nobody's head cut off. aspect_ratio uses the same presets as LC Aspect Ratio Simplifier (or custom). framing: widest keeps as much picture as the format allows; upper body, head & shoulders and face frame tighter from the face size (character datasets). faces: all keeps every face in the crop, largest frames the main one (use it when posters or background people get picked up). output_size: preset size (e.g. 1080x1350) or keep pixels. SAM 3 finds the faces; wire a mask to frame anything else (a car, a product, a pet). Nothing found: a centre crop, and the console says so | | LC Create Mask 🖼️✏️ | Draw a mask right on the node. Click to add points (no limit), drag to move, shift+click removes one. Flip the toggle to pencil for freehand. smooth, invert, feather, and opacity are built in. Undo and Clear buttons on the node. block (top widget) set to if_empty_mask stops everything downstream until you have drawn something, same idea as Impact Pack's Preview Bridge. Run once to get the preview. | | LC Mask Refine ✨ | Tighten a mask with a trimap. edge_erode pulls the sure-foreground in from the mask edge, edge_dilate sets how far the unknown band reaches out. Only that band is decided by the method (vitmatte for hair and fine detail, guided_filter for speed, or none), so a narrower band means a tighter mask. Also grow (negative shrinks), fix_gap, black_point / white_point, feather, and edge_color_spread to cut background halos on the cutout. Plug a mask into detail_region (like hair) to give just that area a wider trimap (detail_erode / detail_dilate). Outputs cutout (RGBA), mask, trimap. Before/after wipe on the node, same as the LC123 previews: hover and the after image sits left of the pointer. preview_view switches between cutout, mask and trimap. Run once to get the preview. 💡 Start with 10 / 10, then shrink both toward 3 for a tight edge. Micro adjust: every pixel value (grow, fix_gap, edge_erode, edge_dilate, the detail_ ones, feather) takes decimals in 0.1 px steps, the points step by 0.001, and the edges are true circles, not blocks. Wire sliders in as FLOAT (INT sockets will not plug into these). ⚠️ Numbers are real pixels. | | LC Remove Background ✂️ | Cut the subject out with a BiRefNet model. Outputs cutout (RGBA), mask, and on_background (subject on a solid color). Optional refine (guided_filter or vitmatte) with the same trimap controls (edge_erode / edge_dilate), plus grow, black_point / white_point, and feather. No preview on the node (add LC Mask Refine after it if you want one). Uses ComfyUI's built-in background-removal loader, so no extra dependency. 💡 Raise black_point to clear faint haze, use vitmatte for hair. | | LC Image Blend Advance 🎚️ | Place a layer on a background and blend it. x_percent / y_percent (0.01 steps), scale, aspect_ratio, rotate, mirror, and 29 blend modes, with opacity in 0.1 steps and an optional layer_mask (multiplied with the layer's own alpha). background_image is optional: leave it off and the layer goes on a transparent canvas the size of the layer (RGBA out, like LayerStyle's V3). transform_method: lanczos, bicubic, bilinear, nearest. Outputs image and mask (where the layer landed). Pure torch, no extra dependency.| | LC Segment Anything 🎯 | Select things by describing them: type hair, eyes, bikini and each word is found and combined. engine: grounding_dino + sam (GroundingDINO finds boxes, SAM makes masks) or sam3 (SAM 3 does both, using ComfyUI's built-in SAM 3 support). Then the same refine as Mask Refine (refine, edge_erode / edge_dilate, black_point / white_point) and a before/after wipe on the node. Outputs cutout (RGBA), mask, raw_mask. 💡 For sam3, sam3_threshold 0.5 is a good start; near 0.3 it invents matches for words that are not in the image. Every match is found (both hands, every face), and your own sam3.pt / sam3.safetensors work as well as the SAM 3.1 download. | | LC Smart Detailer 🩹 | Replaces the SEGS / Impact FaceDetailer chain with one node. Finds what you type (hands, face, eyes) with SAM 3 and redraws it at full detail. Each hand or face gets its own crop, scaled to about inpaint_resolution squared pixels, redrawn, then blended back in. The redraw reaches a little past the pasted area and its colour is matched to the original along the seam, so the edge does not show. tone_match (on): the redraw gets the original's brightness, contrast and colour back where it is pasted, keeping the new detail (a light redraw otherwise comes back flat and dull); turn it off when the redraw is meant to change colours. Changing denoise, steps or feather redraws without searching again. grow / feather shape the mask, padding is the context around each crop, blend mixes the original back in (1 = all redraw). Works with 4-D and 5-D latents (SDXL, Flux, Krea 2, Anima, Qwen, etc.). Leave negative empty for cfg 1 models, wire it for SDXL. Nothing found? The image passes through untouched and the node says so. Before/after wipe on the node with the found areas outlined. LC Smart Detailer (pipe) 🩹 takes the image, model, VAE and sampler settings from the LC pipe; wire an image, model or vae into it to use that one instead (a node without a pipe changed the picture, a model with LoRAs on it, or a different model). It puts its result back in the pipe's image and leaves two things for the nodes after it: every area the detailers redrew (merged, so face + hands + eyes add up), which LC VOSR2 Upscale (pipe) carries to the new size (its protected output), and what SAM 3 found, so the same word is never searched twice. Both are tied to the picture they were made for and ignored if it changed in between. 💡 Faces: denoise 0.3, blend 0.8. cfg defaults to 1 on both versions (the pipe version no longer takes the main sampler's cfg_1, which made redraws drastic). 1 works for Krea 2, Flux, Z-Image and, at detailer denoise up to about 0.4, SDXL too (tested cfg 1 to 7: the same result). It never goes below 1: below 1 the redraw comes out speckled. 💡 If your positive describes the whole scene, wire a short prompt for the area into inpaint_positive (e.g. "close-up of a woman's face, natural skin"). Otherwise, at denoise 0.3 and up, the scene can get drawn into every crop. | | LC VOSR2 Upscale 🧩 | Upscale that keeps the detailers' work. VOSR 2.0 (one step, built in, no other pack needed) sharpens and adds fine detail to the picture it is given instead of redrawing it with your model, so faces and hands come through as the detailers left them, with no new skin blotches. Run it after the detailers: render → LC Smart Detailer → LC VOSR2 Upscale. upscale_by 1.5x, 2x, 3x or 4x in one pass, 6x (3x + 2x) and 8x (4x + 2x) in two (one big pass breaks the skin up into patches), color_alignment wavelet keeps the colours, tile_size 512 (what it was trained at), vae_tile_size lower = less VRAM. LC VOSR2 Upscale (pipe) 🧩 takes the picture from the LC pipe (what the detailers left in it) and puts the result back, one wire after LC Smart Detailer (pipe). Its protected output is the detailers' areas at the new size: wire it into LC Skin Texture / LC Skin Upscale protect_mask so they do not add skin detail on top of the detailers' work (the plain version takes LC Smart Detailer's mask as protect_mask and gives the same output). 2x of a 1 MP picture: about 5 s and under 7 GB on an RTX 5090. Weights download on first run (about 6.5 GB). | | LC Krea2 AnyPaint 🩹 | Krea 2 inpaint and outpaint with yijunwang2's AnyPaint LoRA (Here, put it in your loras folder), run the way it was trained. Everything outside the mask comes back exactly as it was. The model sees a small copy of the whole picture while it paints, so the new part matches the light and the scene, and a thin border around the mask is redrawn so the seam disappears. prompt: describe the whole finished picture, not just the change. crop_to_mask on for inpaint (the area is redrawn in its own crop at inpaint_resolution), off for outpaint. composite: keep original pastes back only the painted area, raw is the whole redraw. Use Krea 2 Turbo, 8 steps, euler / simple, cfg 1. 💡 "Remove the car" leaves a flat blob: describe what should be there instead. | | LC Z-Image AnyPaint 🩹 | The same inpaint and outpaint for Z-Image (Turbo / Base): the same crop, keep, border and seam fix, painted by Alibaba's Z-Image Fun ControlNet Union 2.1 in inpaint mode (Here, put it in models/model_patches; the 8steps one is for Turbo). The only difference from the Krea 2 node is the top two widgets: control (the union patch) and control_strength. Describe the whole finished picture in the prompt. | | LC SDXL AnyPaint 🩹 | The same inpaint and outpaint for any SDXL checkpoint (SDXL, Pony, Illustrious, your finetunes), no LoRA needed. Painted by xinsir's ControlNet Union SDXL ProMax in repaint mode (Here, put the promax file in models/controlnet). prompt: describe what goes in the mask, short ("a red bucket full of fish on the pier"). To remove something, describe what should be there instead, or leave it empty and it fills from the picture around the mask (outpaint still wants the scene described). SDXL uses a real negative and cfg (25 steps, cfg 5, dpmpp_2m / karras). fooocus_inpaint (optional, off by default): lllyasviel's Fooocus inpaint patch (Here, models/inpaint, needs Acly's ComfyUI Inpaint Nodes) blends best on outpaints and fills but rebuilds what was there more than it adds. ⚠️ With both on, drop control_strength to about 0.5: Pony and Illustrious turn to noise at 1. ⚠️ Outpaint: control_strength about 0.2. At 1 the straight edge of the old picture reads as a real edge, and SDXL paints a window frame, fence or line along it. Inpaint stays at 1 (the mask is surrounded by your picture, so it holds the light and edges through the hole). 💡 A boolean and two floats into control_strength switches it with your inpaint / outpaint toggle. | | LC Person Mask 🧍 | Pick the parts of a person: face, hair, body, clothes, accessories, background, combined into one mask. engine: segformer (higher resolution, body is arms and legs only) or mediapipe (body includes torso and neck skin). remove_features cuts the eyes, brows, lips and teeth out with SAM 3. 💡 Skin only: mediapipe, face + body, remove_features on, then feed the mask to LC Skin Beauty, LC Skin Upscale or LC Sharpen Pro. Then the same refine as Mask Refine. No preview on the node (add LC Mask Refine after it if you want one). Outputs cutout (RGBA), mask, raw_mask. 💡 For skin work, enable face and body. | | LC Auto Adjust 🔆 | One-click auto levels. mode: RGB (each channel on its own, also removes color casts), lum + sat, luminance, saturation, or mono. Then red / green / blue balance, brightness, contrast, saturation trims, and strength to blend back toward the original. All in 0.1 steps. An optional mask limits both where the levels are measured and where the result is applied (an alpha channel works as the mask when none is connected). Pure torch, no extra dependency. | | LC Depth Anything V2 🌊 | Depth map from one image, normalized 0-1 per image, bright = near unless invert is on. Reads the original Depth Anything V2 weights, so a copy you already have in models/depthanything works (it also finds the ones in comfyui_controlnet_aux). Outputs depth (3-channel image) and depth_mask. resolution in steps of 14. Matches comfyui_controlnet_aux's Depth Anything V2 (correlation 1.000). Model code is vendored, no extra package. The model list reads the same on every machine (Small / Base / Large), so shared workflows never show a missing model; older saved workflows still load. | | LC Normal Map (BAE) 🗺️ | Surface normal map (RGB = XYZ) with the BAE network, the same style ControlNet's normal_bae makes. resolution, flip_y for OpenGL vs DirectX, and optional blur_radius / sigma (ComfyUI's own Blur Image, so it matches chaining that node after this one; radius 0 = off). Matches comfyui_controlnet_aux's BAE (correlation 0.9999). The model entry reads the same on every machine, and older saved workflows still run. Needs the timm package (the installer adds it if missing). 💡 Both maps are what LC Lighting Control in LC123 wants as its depth_map and normal_map. |

---

LC VOSR2 Upscale🧩

What is it?

• An upscaler built for speed and low VRAM, with the highest quality possible.
• It uses a dedicated upscale model (VOSR 2.0) that sharpens and adds fine detail without redrawing the picture.
• 💡Hook it up after my detailers with the pipe version. It passes along what the detailers already fixed, so LC Skin Upscale and LC Skin Texture don't turn skin pores into freckles or craters.

Render → LC Smart Detailer (pipe) → LC VOSR2 Upscale (pipe). One wire.

Things to notice:

• Skin: clean and smooth, with the lashes and brows sharp. No painted-on grain.
• Fabric: real herringbone and cable knit instead of mush.
• Small faces: the old photo's faces come back as faces.

---

Speed and size chart*

| upscale_by | Output size | Time | GPU memory | |---|---|---|---| | 1.5x | 1596 x 2028 | 2.8 s | 4.7 GB | | 2x | 2128 x 2704 | 4.4 s | 4.8 GB | | 3x | 3192 x 4056 | 8.6 s | 5.1 GB | | 4x | 4256 x 5408 | 16 s | 5.7 GB | | 6x (3x + 2x) | 6384 x 8112 | 46 s | 8.4 GB | | 8x (4x + 2x) | 8512 x 10816 | 75 s | 12.0 GB |

*All tests based on a 1064x1352 render on an RTX 5090, warm.

⏱️ Time will vary by card, but the output size will not. About 5 GB covers 1.5x - 3x, so it can almost run on a 🥔

• 6x and 8x run two passes on purpose. One big pass breaks the skin up into patches.

---

Settings ⚙️

• upscale_by: 1.5x - 4x run in one pass, 6x and 8x in two. 2x is the sweet spot, 1.5x is a third faster.
• color_alignment: wavelet.
• tile_size: 512, what it was trained at. Leave it.
• vae_tile_size: lower = less VRAM, a little slower.
• seed: only changes the finest grain.
• protected: the detailers' areas at the new size. Wire it into protect_mask on LC Skin Upscale and LC Skin Texture so they do not add skin detail twice.

• ✋ The first run downloads the weights (about 6.5 GB) into models/vosr2. This takes a few minutes.

---

Example workflows

| Workflow | What it does | |---|---| | workflows/LC Maskmaker nodes.json | Every node in the pack wired to one image, each with its own preview, so you can compare them side by side. Also uses LC Image Desaturate and LC Skin Beauty from LC123. Bilingual (English / Chinese) note on the canvas. | | workflows/Lonecats Krea2 AnyPaint V1.1.json | Inpaint and outpaint with Krea 2 Turbo and LC Krea2 AnyPaint. Paint the mask on LC Create Mask, drag the edges out on LC Outpaint, or turn and pad the picture with LC Image Rotate + Pad, and switch between the groups with the Groups Bypasser. Needs LC123 and the AnyPaint LoRA. |

---

Models and licensing

• ✋ Models are never bundled or redistributed. Model-backed nodes download from the original source the first time you use them, or reuse a copy you already have.
• Each model in a dropdown is labelled with its license, so you can see what you are picking.
• VITMatte (LC Mask Refine): vitmatte-small-composition-1k and vitmatte-base-composition-1k, both Apache-2.0. Saved under models/vitmatte. A copy you already have in models/vitmatte is picked up automatically.
• BiRefNet (LC Remove Background): general, matting, portrait, and HR 2048 variants, all MIT. Saved under models/background_removal, the same folder ComfyUI's own loader reads.
• RMBG-2.0 (BRIA): used if you already have it in models/RMBG/RMBG-2.0. Not downloaded for you. ⚠️ Non-commercial license.
• Older BiRefNet .pth files (like BiRefNet-ep480.pth) use an older layout that ComfyUI cannot load, so they are not listed.
• GroundingDINO + SAM (LC Segment Anything): grounding-dino-tiny / -base and sam-vit-base / -large / -huge, all Apache-2.0, Hugging Face format. Saved under models/grounding-dino and models/sam. Your original-format .pth files there are not used (different layout).
• SAM 3 (LC Segment Anything, sam3 engine): Meta's original facebook/sam3 is gated (manual approval), so it is never downloaded for you. Your own sam3.safetensors in models/sam3 (or ComfyUI's models/detection) is picked up. For a one-click option, the dropdown offers SAM 3.1 (sam3.1_multiplex_fp16.safetensors, 1.7 GB) from Comfy-Org/sam3.1, the same ungated copy ComfyUI's own SAM 3 template uses, saved to models/sam3. SAM License terms apply (Hugging Face lists it as other).
• Segformer clothes (LC Person Mask): mattmdjaga/segformer_b2_clothes, saved under models/segformer_b2_clothes. ⚠️ Its Hugging Face card says license other and it is built on NVIDIA's SegFormer, so treat it as not cleared for commercial use until you have checked.
• MediaPipe selfie multiclass (LC Person Mask): Google's selfie_multiclass_256x256.tflite, saved under models/mediapipe. Check its model card for terms. The mediapipe Python package is not installed by this pack.
• Depth Anything V2 (LC Depth Anything V2): Small is Apache-2.0, Base / Large / Giant are CC-BY-NC-4.0 (non-commercial). Saved under models/depthanything. Small is listed first so it is the default.
• BAE normals (LC Normal Map): scannet.pt from lllyasviel/Annotators, saved under models/normalbae. ⚠️ Trained on ScanNet and the Hugging Face card only says other. Treat as not cleared for commercial use until you have checked.
• VOSR 2.0 (LC VOSR2 Upscale): from CSWRY/VOSR, Apache-2.0, including its DINOv2-L encoder (Meta) and Qwen-Image VAE. Saved under models/vosr2/VOSR2 (if you already use the ComfyUI-VOSR2 pack, the same files are used).
• ⚠️ Some models are non-commercial. The label tells you. Checking the license before commercial use is on you.

---

License

MIT for the code in this pack.

Third-party model code in vendor/ is credited in THIRD_PARTY.md.