MODEL GUIDE

LTX-2.3 Clean Plate: remove people from video, no masks required

In January 2026 Lightricks released LTX-2, an open-weight video foundation model — and alongside the LTX-2.3 update came a small adapter with an outsized use case: the Clean Plate IC-LoRA. Feed it a clip, and it returns the same shot with the people, pedestrians, and vehicles removed and the background rebuilt behind them.

What makes it different

Traditional video object removal needs a mask: you (or a tracker) outline the subject on every frame, then an inpainting model fills the hole. The Clean Plate LoRA skips all of that. It is an in-context LoRA trained video-to-video — the original clip is the conditioning signal, and the model regenerates the whole shot without its dynamic subjects. Full-frame, no region selection, no rotoscoping.

• Removes people, crowds, and vehicles globally in one pass
• Keeps architecture, ground markings, foliage, street furniture
• Works with both static and moving cameras
• Frame-aligned output: same length, same framing as the input

Model facts

• Base: LTX-2.3-22B, a 22-billion parameter DiT video model by Lightricks
• Adapter: rank-32 IC-LoRA (~330 MB), trained on 26 paired clips of scenes with and without people
• Training resolution: 1024×576 and 576×1024 at 49 frames / 25 fps, validated up to 1920×1088
• License: LTX-2 Community License — free commercial use for entities under $10M annual revenue

Run it online vs. running it yourself

The official route is a ComfyUI workflow: download the 22B base model and the LoRA, wire up the IC-LoRA video-to-video graph, and run it on your own GPU — realistically a 24 GB+ card for HD output. It works well, and if you already live in ComfyUI, we compare both routes here.

This site runs the exact same model on cloud GPUs. Upload a clip, get the clean plate back in 2–3 minutes, download an MP4. One free clip a day, no signup, no CUDA errors at midnight.

Prompting: the part everyone gets wrong

The LoRA was trained on captions describing the empty result, not removal commands. “Remove the man” underperforms; “an empty street with no people anywhere in the frame” is what works. Naming stubborn objects in both the positive and negative prompt is the difference between a bike disappearing and its handlebars haunting your shot. Our generator bakes the official prompt template in automatically — you can optionally name extra objects to remove and we splice them into both prompts the right way.

CleanPlate AI is an independent service built on the open-weight LTX-2.3 Clean Plate model. We are not affiliated with or endorsed by Lightricks. “LTX” is referenced solely to identify the underlying model.