Simple Video Watermark Remover — v1.0 — v1.1

Remove a watermark from every frame of a video. SAM3 finds it from a text prompt, Mask Refine grows and softens the mask, LaMa fills it, and Save Video rebuilds the clip at the source frame rate with the original audio.

Requirements (read this first)

ComfyUI , updated (I tested on 0.37.0).

AusBoss nodes 2.2.0 or newer: ComfyUI-Manager → search "AusBoss" (registry id ausboss-nodes ). Manager installs it.

ComfyUI-RMBG , for the SAM3 Segment node: ComfyUI-Manager → search "RMBG" (registry id comfyui-rmbg ). On its first run it downloads sam3.pt (3.45 GB) into ComfyUI/models/sam3/ by itself.

big-lama.pt (206 MB): download it here and put it in ComfyUI/models/lama/ yourself; the LaMa node never downloads models.

Nothing else. rgthree and Easy-Use are no longer needed.

Restart ComfyUI after installing the two packs.

What changed since v1.1

Two packs instead of four. v1.1 also needed rgthree (group switch, labels) and Easy-Use (the frame counter). Those nodes are gone; both branches are plain groups and a Workflow Note card explains the run.

"Install Missing Custom Nodes" works now. v1.1 tagged SAM3 Segment and the frame counter with the wrong node-pack id, so ComfyUI-Manager looked for them in the wrong pack. SAM3 Segment now names ComfyUI-RMBG. The AusBoss nodes are also on the Comfy Registry now; they weren't when v1.1 went up.

Runs as downloaded. v1.1 shipped with the full-video branch bypassed; now both branches run, and SAM3 unloads its model after each branch to leave room for LaMa.

Saves straight into the output folder instead of an AusBoss/ subfolder.

Files that use the v1.0 SimpleWatermarkRemover node still open: AusBoss nodes keep that node id as a compatibility node.

What's inside

Load Video : your clip; trim with the IN/OUT handles.

Detect : SAM3 Segment with the prompt "watermark" (confidence 0.29) builds a mask on every frame.

Refine : Mask Refine grows the mask by 8 px and softens it, so the fill covers the edges too.

Remove : LaMa Inpaint fills the mask frame by frame and leaves every unmasked pixel alone.

Save : Save Video writes the cleaned MP4 at the source fps, with the source audio.

Single-frame check : a second branch runs the same steps on one frame and shows the mask and a before/after slider, so you can check detection before a long run.

Quick start

Install the requirements above and restart ComfyUI.

Drag the workflow in and load your clip in Load Video .

Queue. The top branch writes the cleaned video; the bottom one shows the single-frame check.

For your own clip, check a single frame first: select the full-video group and mute it (Ctrl+M) to work on the still only. If the watermark has a more specific description, put that in the detection prompt. Then unmute and run the whole clip.

Models

big-lama.pt → models/lama · 206 MB

sam3.pt ( downloaded by ComfyUI-RMBG on first run ) → models/sam3 · 3.45 GB

Settings that worked, and what didn't

Detection: "watermark" at confidence 0.29 is the starting point. If SAM3 misses part of the mark or grabs other text, change the prompt or the confidence threshold.

Edges: if an outline stays visible, raise Expand on Mask Refine a little and add a bit more blur.

Limits: LaMa fills each frame on its own, so busy or moving backgrounds under the mark can flicker, and it can leave some distortion behind. Review the output before you run a long clip.

Speed

Measured on an RTX 5090 (32 GB): the 124-frame 768 × 1376 sample at 24 fps was cleaned in 42 s. The first run also downloads SAM3 (3.45 GB).

About the gallery

The first clip is a straight output of this workflow, audio included; the second is the untouched sample with its watermark, for comparison. The sample (a snowy Moscow street at night) carries an AusBoss watermark in the lower-right corner. The cover is a silent before/after of the sample and the result, side by side, plus a still of it.

License & credits

SAM 3 is by Meta, under the SAM License; ComfyUI-RMBG fetches the weights from 1038lab's Hugging Face mirror, which follows Meta's terms. ComfyUI-RMBG is by 1038lab (GPL-3.0). LaMa is by the authors of advimman/lama (Apache 2.0); the big-lama.pt export is hosted in Sanster's model releases. The workflow and the AusBoss nodes are free.

Changelog

v2.0 (2026-09-25): two packs instead of four (AusBoss nodes + ComfyUI-RMBG), fixed node-pack ids for Manager's missing-node install, both branches on by default, sample clip and Workflow Note card.

v1.1 (2026-03-18): SAM3 + LaMa with a single-frame test mode.

Questions or bugs: GitHub issues or the comments here. Post what you make with it; I read everything. Videos on YouTube @ausbossdev . I post new workflows on X @Zanzibased and GitHub .