ComfyUI nodes that treat 32GB VRAM as the bare minimum

What it does

This repository is essentially glue code: a set of custom ComfyUI nodes and pre-built workflows for Lightricks’ LTX-2 video generation model. Since LTX-2 already ships inside ComfyUI core, these nodes expose the model’s advanced features—such as HDR output, motion tracking, and lipdubbing—through the familiar node-graph interface. It also provides model loader tricks to help the 22B parameter model squeeze into 32 GB VRAM.

The interesting bit

The standout additions are highly specialized IC-LoRA nodes that turn a text-to-video model into a post-production tool. There is a Union LoRA that fuses depth and edge control into a single adapter, an HDR LoRA that emits linear ARRI LogC3 data with an optional EXR export, and a Lipdub LoRA that rephrases or translates dialogue while preserving the original speaker’s voice via reference audio tokens.

Key highlights

• Union IC-LoRA combines depth and canny edge control into one adapter, processing downsampled latents to save memory.
• HDR IC-LoRA outputs both a Reinhard-tonemapped SDR preview and raw linear HDR tensors, with optional 16/32-bit EXR sequence export.
• Lipdub IC-LoRA supports multilingual dubbing and same-language rephrasing through a two-stage pipeline that locks audio during upscaling.
• Includes spatial and temporal upscaler workflows for two-stage generation.
• Low VRAM loader nodes attempt to fit the full pipeline into 32 GB VRAM.

Caveats

• Requires a CUDA GPU with 32 GB+ VRAM and over 100 GB of disk space for models and cache.
• Demands a long list of manually downloaded checkpoints, LoRAs, upscalers, and a Gemma text encoder.

Verdict

Worth a look if you are already running ComfyUI and want professional video controls like HDR grading or automated lip-sync inside your node graph. Everyone else—especially anyone without workstation-grade hardware—should probably just watch the demo.