LTX HDR LoRA brings 16-bit color to open-source video

The LTX HDR LoRA is a featherweight 340MB add-on designed to bring professional color depth to open-source video generation. Built on Lightricks' LTX-2.3-22b foundation model, it lets creators convert existing 8-bit SDR clips into 16-bit HDR footage or generate wide dynamic range video directly from text and image prompts.

What it actually does

Open-source AI video generators usually suffer from a quiet bottleneck: they export standard 8-bit SDR files. That means restricted dynamic range, clipped highlights, crushed shadows, and ugly banding whenever you try to push color adjustments in post-production.

This LoRA fixes the color grading ceiling. By lifting generations to 16-bit HDR, it significantly expands dynamic range and preserves nuanced tonal transitions across balanced highlights and deep shadows. Creators can drop generated clips straight into editing suites like Premiere or DaVinci Resolve and actually grade them without the image falling apart.

It works both ways: you can upscale standard 8-bit SDR generations to 16-bit HDR, or prompt fresh HDR videos while maintaining LTX's native synchronized audio output.

How it works under the hood

Based on the team's LumiVid paper, the tool uses an In-Context LoRA (IC-LoRA) architecture. IC-LoRA conditions generation on reference video frames at runtime, offering frame-accurate video-to-video guidance on top of the base diffusion model.

To squeeze a massive 16-bit dynamic range into numbers the model can easily handle, Lightricks implemented cinema-grade LogC3 transforms. The pipeline applies a logarithmic curve before diffusion to align latents within a manageable numerical range, then reverses the transform after generation to reconstruct full 16-bit HDR depth.

Because logarithmic mapping carries the heavy mathematical load, the checkpoint itself stays around 340MB. It runs with a Reference Downscale Factor of 1, meaning reference inputs match your final output resolution.

How to run it in ComfyUI

The safetensors weights are available on Hugging Face under the LTX-2 community license, alongside research code and the project page.

To use it locally, you need the LTX-2.3-22b base model and an up-to-date ComfyUI setup:

Drop the safetensors file into your models/loras folder.

Load the official LTX-2 IC-LoRA workflow from the community repository. Make sure your graph includes the dedicated nodes: LTXICLoRALoaderModelOnly loads the LoRA and sets the downscale factor, while LTXAddVideoICLoRAGuide pipes the reference latent into the pipeline as guidance.

Why it matters

AI video models have spent the past year obsessing over resolution and motion coherence, but standard 8-bit color has quietly kept them out of serious post-production pipelines. You simply cannot mix brittle SDR AI footage into professional color-graded workflows.

Lightricks solved this without retraining a monstrous base checkpoint. Slipping a 340MB IC-LoRA with LogC3 math into an existing pipeline is a remarkably pragmatic move. If you already generate video with ComfyUI and care about color fidelity, this is an instant download, provided your rig has the VRAM to handle the 22B base model.