IC-LoRA for LTX-2.5: Neural Rendering from 3D Playblasts

Lightricks has released the IC-LoRA model for the generative video model LTX-2.5, featuring 22 billion parameters. The tool transforms 3D playblasts into high-quality videos, opening new possibilities for post-production and VFX.

Technical Specifications

The model weighs approximately 1.31 GB and uses rank 128. It is specifically trained to interpret data from a 3D viewport or playblast as a control signal for neural rendering. This ensures precise alignment between scene geometry and the generated result.

The rendering process is optimized via a two-pass ComfyUI workflow. The first pass runs at 960x544 resolution in 8 steps. The second stage upscales the latent space by 2x and increases resolution to 1920x1088 in an additional 3 steps.

Basic distilled weights of LTX-2.5, a Gemma 4 12B encoder, video/audio VAE, and a spatial upscaler are required. Official ComfyUI nodes and workflows are also necessary.

Input Data and Integration

• A playblast from Blender serves as the primary input signal.
• The first frame of the playblast is processed in the desired style.

It is recommended to take the first frame of the playblast and run it through image-to-image for better control over visual style. This approach helps integrate 3D geometry with artistic requirements.

Expanding Lightricks' Tool Ecosystem

In addition to the new model, the company has developed Alpha Gen, an SDR-to-HDR converter, Native Resolution, Refine/Restore, and other tools. Lightricks aims to capture the niche of open-source post-production and VFX toolkits.

All materials are available for download: LoRA model, full file set, and ready-made JSON workflow.