You need to agree to share your contact information to access this model

By clicking "Agree and Access" you acknowledge the Privacy Policy and consent to receive offers and updates. You can unsubscribe at any time.

LTX-2.3 22B IC-LoRA Inpainting & Outpainting

This is an Inpainting & Outpainting IC-LoRA trained on top of LTX-2.3-22B. It supports two related tasks:

Inpainting — filling or removing a masked region within the frame (e.g. removing a person, replacing an object).
Outpainting — extending the canvas beyond its original boundaries, generating plausible new content that blends seamlessly with the existing video.

It is based on the LTX-2.3 foundation model.

Model Files

ltx-2.3-22b-ic-lora-in-outpainting-0.9.safetensors — the single released checkpoint.

Model Details

Base Model: LTX-2.3-22B Video
Training Type: IC-LoRA (video-to-video, mask-conditioned)
Control Type: Reference video + binary mask. The model generates content for the masked region while leaving the unmasked area intact.

Intended Use

Inpainting: Remove or replace content within a masked region of the video. Common uses: object removal, background replacement, face swap, wardrobe change.

Outpainting: Extend the video canvas horizontally or vertically. The model generates new content that transitions naturally into the original footage.

Pipeline Details

Both tasks use a two-stage pipeline to produce clean results at the boundary:

Stage 1 — coarse generation over the masked region.
Stage 2 — a second pass with refined boundary handling, including careful masking and blending the original content back (via Laplacian blend) to ensure a smooth, artifact-free transition between the generated and original regions.

Prompting

The IC-LoRA can work without a prompt — an empty or minimal prompt is sufficient for most use cases.

If you want to guide what appears in the masked/outpainted region, you can describe only that region in the prompt (e.g. "a sandy beach with crashing waves" for the outpainted strip). Avoid describing the full scene — prompting elements that already exist in the unmasked area may cause the model to duplicate objects or people.

Mask Guidelines

Mask quality is critical, especially for inpainting:

Dilate the mask beyond the visible boundary of the object you are removing. The model understands scene causality — if a shadow or reflection of the removed object is still visible, it may re-inpaint the object from that cue alone.
Include causally connected regions: shadows, reflections, cast light, and contact areas should all be covered by the mask.
For outpainting, the mask is the new canvas region; no dilation is needed.

Usage

🔌 ComfyUI

Copy the LoRA weights into models/loras.
Load the LTX-2.3-22B base model and add ltx-2.3-22b-ic-lora-in-outpainting-0.9.safetensors as the LoRA.
Use an IC-LoRA (video-to-video) workflow from the LTX-2 ComfyUI repository that supports mask conditioning. Connect your source video as the reference and supply the binary mask for the region to fill or extend.