oumoumad/LTX-2.3-22b-IC-LoRA-Outpaint

LTX-2.3 22B IC-LoRA Outpaint

This is an Outpaint IC-LoRA trained on top of LTX-2.3-22b, designed to extend the canvas of an input video by generating new content in regions marked as pure black. You can use it to outpaint on the sides, top, bottom, or any combination — the model fills the black regions with content that is visually and temporally consistent with the original footage.

It is based on the LTX-2.3 foundation model.

Model Files

ltx-2.3-22b-ic-lora-outpaint.safetensors

Model Details

• Base Model: LTX-2.3-22b
• Training Type: IC LoRA
• Purpose: Outpaint / extend video canvas

How It Works

The model was trained with pure black pixels (RGB 0,0,0) as the sentinel for the region to generate. At inference, letterbox your source video to the target canvas size with black bars on the sides / top / bottom you want to extend, and the model will fill those regions with newly generated content consistent with the scene.

🔌 Using in ComfyUI

1. Copy the LoRA weights into models/loras.
2. Use the IC-LoRA workflow from the LTX-2 ComfyUI repository you can also use my workflow.
3. Load the LoRA using the LTXICLoRALoaderModelOnly node.

⚠️ Dark scenes: use a gamma correction trick

Because the model uses pure black as the "generate here" sentinel, very dark source footage (deep shadows, night scenes, underwater, etc.) can be ambiguous: the real fill bars may be interpreted as legitimate dark scene content and left un-generated.

The fix is a simple gamma round-trip:

1. Before feeding to the model, apply gamma 2.0 (brightening) to your letterboxed input. The real dark content lifts into clearly-colored territory while the pure-black bars stay black — giving the model an unambiguous signal of where to paint.
2. After the model produces its output, apply the inverse gamma 0.5 to return everything (original center + newly generated regions) to the original exposure.

Because gamma 2.0 and 0.5 are exact inverses, the round-trip is mathematically lossless on continuous values — the only loss is the same VAE round-trip every other clip experiences.

In ComfyUI, the Color Correct (mtb) node works well for both the forward and inverse steps.

License

See the LTX-2-community-license for full terms.