LTX-2.3 Remove Person (inpaint) ⚡

Videos, unlike photos, come with motion - and removing a person from a moving clip is the kind of task that shows exactly why LTX-2.3's masked inpainting is built the way it is. This node is stage one of the pack's verified person-replacement pipeline: it regenerates the masked region (typically a person) using the official LTX-2.3 in/outpainting IC-LoRA, following the recipe Wan2GP proved out. The result is a clean plate: your person gone, the background reconstructed, and - because of the blend step that follows - every pixel outside the mask left pixel-identical.

The mechanism is worth understanding because it's nothing like img2img inpainting. The per-frame mask gets dilated 5px, painted chroma green (#66FF00), green-padded at the tail, and appended as clean guide tokens at strength exactly 1.0, alongside the ltx-2.3-22b-ic-lora-in-outpainting-0.9 LoRA (loaded in-node from your loras folder). The LoRA learned "green = regenerate"; the mask deliberately never touches the denoise mask. Sample, crop the guides, decode, and the green silhouette is what tells the model where to rebuild.

The inputs that matter

• video, vae / audio_vae (the loader's video_vae and audio_vae outputs), model, clip - the kit stack. video sets length and frame rate from itself.
• mask - per-frame mask, white = regenerate. Produce it with comfy-core's SAM3 nodes (SAM3_Detect with a prompt like "woman", or SAM3_VideoTrack → SAM3_TrackToMask); a single-frame mask tiles across the clip. You'll need sam3.1_multiplex_fp16.safetensors from Comfy-Org/sam3 in models/checkpoints.
• prompt - describe the scene without the removed subject: "An empty room with a wooden floor and beige walls, doorway in the background, static camera." This is where people trip up - describe what should fill the gap, not the person you're erasing.
• inpaint_lora - the outpainting IC-LoRA, loaded at its trained strength 1.0. none disables the whole mechanism (for A/B only - the green fill would just be reproduced verbatim).
• keep_original_audio (on) - keeps the source clip's own audio; off lets the model generate new audio.
• mask_dilation (default 5) - the pre-fill dilation radius in pixels; the trained default.
• start_image - optional identity anchor: becomes the control video's first frame verbatim (no green fill on frame 0). For person replacement, that's an image of the new person in the same scene and starting pose - the green silhouette carries the motion, this frame carries who fills it, and the prompt should then describe the scene with the new subject.

Outputs: model, positive, negative, latent, frame_rate for the sampler, plus source_frames and blend_mask, which you feed straight into LTX-2.3 Mask Blend afterward.

The chain

LTX-2.3 Remove Person → LTX-2.3 KSampler (distilled) → LTX-2.3 Crop Video Guide → LTX-2.3 AV Decode → LTX-2.3 Mask Blend (with source_frames/blend_mask). Decode's video → GetVideoComponents → Mask Blend's images. Then mux the original clip's audio into the final CreateVideo rather than the decoded audio latent - the round-trip through the audio VAE is lossy, and keep_original_audio only holds audio for conditioning, it doesn't guarantee the output track is bit-perfect.

Installing

cd ComfyUI/custom_nodes
git clone https://github.com/ChrisColeTech/ComfyUI-GGUF-Loader
pip install --upgrade gguf

Restart. You need the LTX-2.3 A/V kit (DiT, Gemma-3 encoder, video + audio VAEs), the outpainting IC-LoRA, and the SAM3 checkpoint. Node under 🤖 CCTech/LTX-2.3.

Gotchas

Frame math again: width/height divisible by 32, frames divisible by 8 plus 1. If the removed-person region looks like it's being filled with the same content, check the mask isn't all-black (white = regenerate) and that inpaint_lora isn't none. And this node alone is not the whole job - without the Mask Blend stage, you'll see the regenerated region but also the subtle changes everywhere else.