# How to Install LTX-2.5 in ComfyUI: Setup, VRAM Needs, Real Results > Source: > Published: 2026-08-17 00:00:00+00:00 # How to Install LTX-2.5 in ComfyUI: Setup, VRAM Needs, Real Results A grounded guide to installing LTX-2.5 in ComfyUI, covering custom nodes, model downloads, VRAM requirements, and honest first-generation quality. ## What is LTX-2.5 and why does it matter? LTX-2.5 is a video generation model from Lightricks that produces video and synchronized audio together in a single generation pass, from a text prompt, a reference image, or another video. Instead of generating silent clips and bolting on sound later, the model renders visuals and soundtrack at the same time, which is why dialogue, ambient noise, and on-screen action can line up without a separate audio pipeline. It runs on consumer and prosumer GPU hardware through ComfyUI, which is what makes it interesting to people who build local AI video pipelines rather than rely on hosted APIs. ## TL;DR **LTX-2.5** generates video and audio in one pass, so sound effects and actions can be synced without a separate audio generation step.- The release adds **multi-shot generation** that aims to hold character identity, lighting, and voice consistent across multiple cuts in the same generation. - A new **diffusion video decoder** and a**22 billion parameter distilled checkpoint** are both part of this release, with the distilled version trading some quality for faster generation. - Running the model in **full precision demands roughly 66GB of VRAM**, which in practice means at least an 80GB card like an A100 or H100 for comfortable headroom. **Quantized versions** exist that reportedly run in the 30 to 35GB VRAM range, making the model workable on 48GB cards.- Installation requires a **custom ComfyUI node**, its Python requirements, and manually downloading gated model files from Hugging Face into the correct ComfyUI model folders. - Early, uncurated test generations showed **real limitations**: reversed motion, inconsistent traffic light states, and anime character details getting mangled, even though audio sync generally held up. ## How do you install LTX-2.5 in ComfyUI? The installation assumes you already have a working ComfyUI setup. If you don’t, you’ll need to get that running first before touching LTX-2.5, since this is an add-on model and node set, not a standalone application. Once ComfyUI is installed, the process breaks down into a few concrete steps: **Install the custom node.** Navigate to your ComfyUI custom_nodes directory in a terminal and install the LTX-2.5 node package, then install its Python requirements (typically via a requirements.txt file bundled with the node).**Download the model files.** Go to the LTX-2.5 model card on Hugging Face. This is a gated model, so you need a Hugging Face account and to accept the license terms before you can download anything.**Place files in the correct folders.** Hugging Face provides separate files for the diffusion model, the text encoder, and the variational autoencoder (VAE). Each goes into its matching subfolder inside ComfyUI’s models directory: diffusion models go in the diffusion model folder, the text encoder in the text encoder folder, and the VAE in the VAE folder. There’s also an optional upscaling model you can add if you want higher-resolution output.**Free up disk space.** Between the diffusion model, text encoder, VAE, and optional upscaler, expect to need somewhere around 70 to 80GB of free disk space before you start.**Load the example workflow.** Lightricks provides a reference ComfyUI workflow (a two-stage setup that takes an image and a text prompt) that you can load directly instead of building nodes from scratch. The VAE’s role here is worth understanding if you’re new to these pipelines: it converts between the compressed latent space the diffusion model works in and the actual pixel space of video frames and audio waveforms. If you’ve worked with other diffusion-based image or video models, this part of the pipeline will look familiar. ## How much VRAM does LTX-2.5 actually need? In full precision, loading and running LTX-2.5 consumes around 66GB of VRAM. That number alone rules out most consumer GPUs. Realistically, running this model comfortably in full precision means having access to an 80GB card, such as an A100 or H100, whether that’s owned hardware or a rented cloud GPU instance. For anyone without access to that tier of hardware, Lightricks also offers quantized versions of the model. These bring VRAM consumption down to roughly 30 to 35GB, which fits on a 48GB card. The tradeoff with quantization is generally some loss of output quality or fidelity in exchange for the drastically reduced memory footprint, though the exact quality delta will depend on the specific quantization method and checkpoint used. This release also includes a 22 billion parameter distilled checkpoint alongside the full model. Distillation here trades some generation quality for speed, which matters a lot in practice: video generation is slow even on high-end hardware, and a faster checkpoint can be the difference between an iterative workflow and a “start it and walk away” workflow. ## What do real LTX-2.5 generations actually look like? ### Built like a system. Not vibe-coded. Remy manages the project — every layer architected, not stitched together at the last second. This is where it’s worth being direct rather than repeating marketing claims. When you install the model locally, following Lightricks’ own workflow exactly, in full precision (not a quantized version), the output quality does not match the polish typically shown in official demo reels. In one uncurated test, a text prompt and reference image describing a car accelerating through an intersection as a traffic light turns green produced a video with several visible errors: the car appeared to reverse instead of moving forward, and the traffic light showed both green and yellow lit simultaneously. Some details were handled correctly, like wet road reflections, though a discrepancy also appeared where the prompt implied a driver but the generated video showed only hands on the steering wheel with no visible person. In a separate test animating an anime-style character (a prompt describing a smiling girl with long silver hair), the model altered the character’s appearance and got a ribbon accessory noticeably wrong, even though the audio track that came with the generation was reasonably in sync with the intended action. A third test involving a person on a trampoline and background audio of someone mowing a lawn showed similar results: imperfect visual physics on the main subject, but audio that stayed convincingly synced to the described action. The consistent pattern is that **audio synchronization tends to hold up better than fine visual detail or physical plausibility**. Motion errors, inconsistent object states (like the traffic light), and character detail drift showed up across different prompts and subject types. ## Is LTX-2.5 worth running locally right now? For anyone with access to enterprise-grade GPU hardware (or willing to rent it), LTX-2.5 offers something distinct: audio and video generated together, locally, without piecing together a separate text-to-video model and text-to-audio model and manually syncing them. That combined generation is genuinely useful for workflows where matching sound to action matters, like short-form content or prototyping. That said, the honest gap between polished demo footage and raw, uncurated first-pass generations is significant. Visual coherence, especially around physical motion and consistent object states, still needs work. Lightricks also offers LoRAs for the model, which are commonly used to fine-tune output toward specific styles or improve consistency for particular use cases, though results without any LoRA applied are, as shown, inconsistent. If your hardware situation limits you to the quantized checkpoints, expect an additional quality tradeoff on top of the baseline limitations already visible in full precision. For teams building production pipelines, this suggests LTX-2.5 is currently better suited to experimentation, rapid prototyping, and workflows with human review built in, rather than fully automated, no-oversight generation. ## Frequently Asked Questions ### What is LTX-2.5? LTX-2.5 is a video generation model from Lightricks that creates video with synchronized audio in a single generation pass, from text prompts, images, or reference videos, and runs through ComfyUI on local or rented GPU hardware. ### How much VRAM do I need to run LTX-2.5? Full precision generation consumes around 66GB of VRAM, effectively requiring an 80GB card like an A100 or H100. Quantized versions of the model reduce this to roughly 30 to 35GB, making 48GB cards viable. ### Does LTX-2.5 require a paid license or account? ## Other agents ship a demo. Remy ships an app. Real backend. Real database. Real auth. Real plumbing. Remy has it all. The model files are hosted on Hugging Face as a gated model, meaning you need a free Hugging Face account and must accept the model’s license terms before downloading the files. ### Is the audio generated by LTX-2.5 actually synced to the video? In testing, audio sync held up reasonably well across different prompts, generally tracking the described action more reliably than the visual details or physical motion of the generated video matched the prompt. ### Can I improve LTX-2.5’s output quality? Lightricks provides LoRAs for the model that are intended to improve or specialize output quality. Baseline generations without any LoRA applied showed clear errors in motion, object consistency, and character detail in testing.