AI video inpainting & object removal
Feed LTX a clip and a binary mask. It fills the masked region and leaves every other pixel untouched, handling rig, crew, logo, and rights cleanup at scale. Open weights, self-hosted, no per-edit license.
.svg)
Key capabilities
Mask-conditioned removal & replacement
Supply a clip and binary mask; LTX fills only the masked region while the rest of the frame stays completely untouched.
Mask what's hidden, not visible
Mask shadows, reflections, and contact points or the model may regenerate the object from those cues. Dilate your mask to cover all of them.
Dailies and delivery cleanup
Strip rigs, crew, boom mics, and set signage without frame-by-frame rotoscoping. Feed batches through your pipeline and deliver clean dailies at scale.
Rights clearance
Remove trademarked logos or unlicensed elements before a shot ships. The model fills the region generatively rather than leaving a blank space, so final output looks finished.
Object and background replacement
Swap a masked subject or backdrop for newly generated content. Prompt only the masked region to describe what should appear; avoid describing the surrounding scene.
Pipeline-native batch cleanup
Run it self-hosted in your infrastructure at volume. No per-edit license, footage never leaves your systems, and you maintain complete control over the workflow.
How it works
Input
Technical characteristics:
- Source — Source video plus binary mask over the region to remove or fill.
- Mask coverage — Dilate the mask to include shadows, reflections, and contact areas.
- Prompt — Optional prompt describing only the masked region; empty or minimal works best for removal.
- Base — LTX-2.3-22B + ltx-2.3-22b-ic-lora-in-outpainting-0.9.safetensors in the ComfyUI IC-LoRA mask-conditioning workflow. (Compatibility with LTX-2.5)
Output
Technical characteristics:
- Edited clip — The masked region is removed or replaced while the unmasked area is preserved.
- Boundary refinement — A refinement pass reduces visible seams between generated and original pixels.
- Run it — Self-host on GPU via ComfyUI or the open-source pipeline, or call through the LTX API.
What users are building with LTX
FAQs
How does it remove an object?
Supply the clip plus a binary mask over the object; the model fills the masked region and leaves the rest untouched. An empty or minimal prompt works best for removal.
Why does my removed object keep reappearing?
The model reconstructs the masked area from the surrounding context. If the mask doesn't cover the object's shadow, reflection, or contact points, the model can regenerate the object from those leftover cues. Dilate the mask to include them.
Can I replace an object instead of just removing it?
Yes. Prompt only the masked region with what should appear there. Don't describe the rest of the scene, or it may duplicate existing subjects.
How do I run it, and what does it cost?
Open weights on HuggingFace via a ComfyUI IC-LoRA mask-conditioning workflow. Self-host on GPU under the LTX-2 Community License, free under $10M ARR with no per-edit fee. Or call it through the LTX API if you would rather not manage the infrastructure.
