Download
1 variant available
This checkpoint includes a config file, download and place it along side the checkpoint.
New notes on the canvas point at what to do: load your video, what the two boxes do, and the two LoRAs you need.
The notes inside the boxes get a big line and small print.
A YouTube link above the page note.
Every group and note is pinned, so a stray drag can't move them. Select one and press P to unpin it.
It runs the same as v1.1: same models, same settings.
Needs AusBoss nodes 2.15.0 or newer. Already have them? Update first.
Show more
Extend a video on any side, and LTX 2.3 paints the new area. Your video goes back on top of the result, and its sound is kept.
How to use it
Load your video.
Pick a ratio. It opens on 16:9, which turns a phone clip into a wide one. The small button next to the ratios flips portrait and landscape.
Press Run.
What you need
AusBoss nodes 2.15.0 or newer. In ComfyUI-Manager, search "AusBoss". Already have them? Update first.
The LTX 2.3 models and the two LoRAs:
ltx-2.3-22b-dev_transformer_only_fp8_scaled.safetensors →
models/diffusion_models· 23.5 GBgemma_3_12B_it_fp8_scaled.safetensors →
models/text_encoders· 13.2 GBltx-2.3_text_projection_bf16.safetensors →
models/text_encoders· 2.31 GBLTX23_video_vae_bf16.safetensors →
models/vae· 1.45 GBltx-2.3-22b-distilled-1.1_lora-dynamic_fro09_avg_rank_111_bf16.safetensors →
models/loras· 2.74 GBltx-2.3-22b-ic-lora-outpaint.safetensors →
models/loras· 1.31 GB
Tips
Keep the fill color black. Pure black is what the IC-LoRA paints over. Gray or any other color comes back as a flat bar.
Crop out edges that confuse it. Noise, a frame line or a busy strip along the edge of your clip can make the new area come out strange. Drag the blue handles on Video Crop + Rotate + Pad inward to cut those edges off first.
It makes 97 frames from IN, about 4 seconds at 24 fps. Drag IN on the timeline to pick the start. On a 60 fps clip, set Every nth to 2 or 3 for more seconds.
Very dark footage can come back flat. Night or underwater clips can read as a bar. Brighten the clip first (gamma 2.0) and darken the result after (gamma 0.5).
Out of memory? Lower Megapixels on the video node, or use the smaller int8_convrot transformer in the same Kijai folder.
Part of the new area still black? Run it again with another seed.
About 2 minutes a run for a short clip on an RTX 5090, model loading included.
Credits
LTX-2.3 is by Lightricks, under the LTX-2 Community License: companies making US$10M or more a year need a paid license from Lightricks. The ComfyUI files are Kijai's split, and Gemma 3 12B is Google's model, packaged by Comfy-Org. The outpaint IC-LoRA is by oumoumad, under the same license. The clip in the gallery was made with Qwen Image 2.1 and MiniMax H3, and their licenses cover it.
v1.2 (2026-10-20): new notes on the canvas, everything pinned, and a YouTube link. It runs the same as v1.1. v1.1 (2026-09-30): a smaller workflow, and the new area gets painted on any clip. v1.0 (2026-09-25): first release.
Questions or bugs: GitHub issues or the comments here. Videos on YouTube @ausbossdev. More from me on X @Zanzibased and GitHub.
