ComfyUI-LTXVideo: LTX-2 Video Generation Nodes
ComfyUI-LTXVideo packages the custom nodes and workflows that give ComfyUI users access to LTX-2's advanced modes — HDR export, speech dubbing, motion-tracked control, and text-to-audio — beyond ComfyUI core. Reach for it if you run ComfyUI on a 32GB+ VRAM CUDA GPU and want IC-LoRA workflows for HDR or dubbing without hand-wiring nodes. Skip it if you lack that VRAM or don't use ComfyUI already, since this repo does nothing on its own.
Understanding ComfyUI-LTXVideo and LTX-2
ComfyUI-LTXVideo is a collection of custom nodes and example workflows Lightricks built to extend ComfyUI's support for LTX-2, its video model. LTX-2 already ships in ComfyUI core, so this repo adds extras: Union, HDR, Dub-It, and Pixel Spatial Upscaler IC-LoRAs, a text-to-audio node, and ready-made LTX-2.3/LTX-2.0 workflow JSON files you load into ComfyUI. It only works as an add-on inside an existing ComfyUI install, not standalone.
Key Capabilities and LoRA Models
- ✓Union IC-LoRA combines depth and edge (canny) control into a single LoRA that runs on downsampled latents, per the README, cutting memory use and speeding inference versus separate control LoRAs.
- ✓HDR IC-LoRA outputs linear HDR video encoded in ARRI LogC3; the LTXVHDRDecodePostprocess node returns both a Reinhard-tonemapped SDR preview and the raw HDR tensor, with optional 16/32-bit EXR sequence export.
- ✓Dub-It IC-LoRA dubs or rephrases speech in an existing video: it regenerates lip movement and audio to match new dialogue text in another language or the same one, while reference audio tokens keep the original speaker's voice.
- ✓Pixel Spatial Upscaler IC-LoRA re-renders a low-resolution clip at 2x or 4x with synthesized detail rather than interpolating pixels — the README frames it as a creative upsampler for a 'draft-then-upscale' workflow, not a pixel-accurate refiner.
- ✓A dedicated text-to-audio path: the LTXVAudioOnlyModel node flips off LTX-2's video cross-attention flags so the joint audio/video transformer denoises audio alone, paired with a placeholder 64x64 empty video latent so the sampler's input shape still lines up.
- ✓Motion-track and pose IC-LoRAs, plus camera-control LoRAs (dolly in/out/left/right, jib up/down, static) and effect LoRAs (colorization, deblur, decompression, day-to-night, in/out-painting, water simulation, instant shave), all listed as downloadable Hugging Face checkpoints.
- ✓Ready-made example workflow JSON files for both the current LTX-2.3 line and the older LTX-2.0 line, covering text/image-to-video, two-stage upsampling, and IC-LoRA control setups, installed straight into example_workflows/.
Installing LTXVideo Custom Nodes
The README's recommended path is Comfy Manager, not a manual git clone: open ComfyUI, open the Manager (or press Ctrl+M), choose Install Custom Nodes, search for "LTXVideo," click Install, then restart ComfyUI. After restart the nodes show up in the node menu under the LTXVideo category, and required models download automatically on first use of a workflow. Before you start, the README lists three prerequisites: ComfyUI itself already installed, a CUDA-compatible GPU with 32GB or more VRAM, and 100GB or more of free disk space for models and cache.
Exploring Example Workflows and Required Models
Once installed, the example workflows live at ComfyUI/custom_nodes/ComfyUI-LTXVideo/example_workflows/, split into a 2.3/ folder (current LTX-2.3 line) and a 2.0/ folder (older LTX-2.0 line). The 2.3 set covers single-stage and two-stage distilled text/image-to-video, the Union IC-LoRA control workflow, motion-track, HDR, Dub-It, pixel spatial upscaling, and a single-stage text-to-audio workflow — load any of these JSON files directly into ComfyUI to start from a working graph instead of wiring nodes from scratch. Running them means downloading the matching models yourself: an LTX-2.3 checkpoint (the dev or distilled-1.1 variant) into models/checkpoints, spatial and temporal upscaler models into models/latent_upscale_models, the distilled LoRA into models/loras for most two-stage pipelines, the full Gemma 3 text encoder into models/text_encoders, and whichever IC-LoRA or camera-control LoRA the workflow calls for, also into models/loras.
Strengths
- ✓Ready-made IC-LoRA workflows for genuinely different jobs — HDR export, speech dubbing, motion-tracked control, creative upscaling, text-to-audio — instead of one generic video-generation node set.
- ✓Example workflow JSON files for both the current LTX-2.3 line and the older LTX-2.0 line, so you load a working graph rather than assembling nodes from documentation alone.
- ✓The HDR path is unusually complete for a ComfyUI add-on: linear HDR output in ARRI LogC3, an SDR preview alongside the raw tensor, and optional 16/32-bit EXR sequence export for grading.
- ✓Installs through Comfy Manager like any other custom node pack, with required models fetched automatically on first use rather than needing a separate manual download step first.
- ✓A documented low-VRAM path (low_vram_loaders.py plus --reserve-vram) for people whose card sits at or near the 32GB minimum instead of comfortably above it.
Hardware Requirements and VRAM Management
- △Baseline hardware is steep: the README's own prerequisites call for a CUDA GPU with 32GB or more VRAM and 100GB or more of free disk space just to get started.
- △Fitting generation into 32GB VRAM at all requires extra steps — the README points to low-VRAM loader nodes in low_vram_loaders.py and a --reserve-vram ComfyUI launch flag, not something that works unmodified on lower-spec cards out of the box.
- △The README doesn't state a license, so check the repository directly before using this in anything beyond personal experimentation.
- △Two-stage pipelines that need the distilled LoRA, plus separate spatial and temporal upscaler models, mean tracking and downloading several multi-gigabyte files correctly rather than one checkpoint.
- △The README doesn't mention a version number or changelog, so there's no documented way to tell which LTX-2.3 model revision a given workflow file expects beyond the filename itself.
Exploring Other ComfyUI Video Generation Tools
Frequently Asked Questions
ComfyUI-LTXVideo's README lists a CUDA-compatible GPU with 32GB or more VRAM and 100GB or more of free disk space for models and cache as prerequisites, on top of an existing ComfyUI install.
Open ComfyUI, launch the Manager (Ctrl+M), choose Install Custom Nodes, search for "LTXVideo," click Install, and restart ComfyUI — the nodes then appear under the LTXVideo category and required models download on first use.
ComfyUI-LTXVideo's LTX-2.3 example workflows sit in the example_workflows/2.3/ folder inside the installed custom node, covering text/image-to-video, IC-LoRA control, upscaling, and text-to-audio JSON files you load directly into ComfyUI.
The Union IC-LoRA in ComfyUI-LTXVideo combines depth and edge (canny) control conditions into one LoRA that runs on downsampled latents, so a single model handles multiple control signals instead of needing separate LoRAs for each.
Yes — ComfyUI-LTXVideo's HDR IC-LoRA produces linear HDR video encoded in ARRI LogC3, with a decode node that outputs both an SDR preview and the raw HDR tensor, plus optional EXR sequence export.
Yes — the LTXVAudioOnlyModel node in ComfyUI-LTXVideo puts LTX-2's joint audio/video transformer into an audio-only mode, skipping video output entirely to generate audio from a text prompt.
The problem it solves
LTX-2 already ships inside ComfyUI core, but core support only gets you the base model — depth/edge/pose-conditioned control, HDR output, speech dubbing, and creative upscaling aren't built in, and wiring the transformer's raw video and audio latents into a custom node graph by hand is fiddly and specific to LTX-2's joint audio/video architecture. ComfyUI-LTXVideo packages that wiring as ready-made nodes, IC-LoRA checkpoints, and example workflow JSON files, so someone who wants HDR export or dialogue redubbing doesn't have to reverse-engineer LTX-2's latent layout themselves.
Best use cases
- •Adding HDR grading output to a ComfyUI video pipeline: run the HDR IC-LoRA workflow to get an EXR sequence for downstream color work in DJV or a compositing tool, instead of tone-mapped SDR only.
- •Redubbing an existing clip into another language without re-shooting: the Dub-It two-stage workflow regenerates lips and audio from a new text prompt while keeping the original speaker's voice.
- •Upscaling a fast, low-resolution LTX-2 draft into a final-resolution render: generate at roughly 280p to lock composition and motion, then run the Pixel Spatial Upscaler IC-LoRA at 2x or 4x.
- •Driving video generation from an existing video's depth, pose, or edge maps in one pass via the Union IC-LoRA, instead of chaining separate single-condition LoRAs.
- •Generating a standalone audio track — narration or sound design — from a text prompt using the T2A workflow, with no video render needed.
Who should try it — and who should skip
Try ComfyUI-LTXVideo if you already run ComfyUI on a 32GB+ VRAM CUDA card and want a specific LTX-2 capability — HDR export, speech dubbing, motion-tracked control, or a fast draft-then-upscale pipeline — without building the node graph from scratch. Skip it if you don't have that VRAM headroom, since even the low-VRAM path via low_vram_loaders.py and --reserve-vram is still built around fitting inside 32GB, or if you're not already a ComfyUI user, since this repo has no standalone functionality of its own.
Related repositories
Is ComfyUI-LTXVideo worth your time?
ChatGPT, Claude and Perplexity can all read this page. Ask one of them what it makes of ComfyUI-LTXVideo.
