LTX Video (Lightricks): The Open-Weight Video Model That Added Audio First

Lightricks is best known for Facetune — the selfie-editing app that defined a category of mobile image tools. In late 2024, the company raised a Series E to fund its AI-first pivot, and in January 2026, released the model that makes that pivot legible: LTX-2, the first open-weight video generation model to natively produce synchronized audio alongside video.

The current iteration is LTX-2.5, released August 11, 2026, extending the same 22B-parameter dual-stream audio-video architecture with native multi-shot generation and clips up to 20 seconds with synchronized audio (up from ~10 seconds on LTX-2.3). As of late August 2026, Lightricks/LTX-2.3 still shows roughly 1.5 million HuggingFace downloads per month and Lightricks/LTX-2.5 roughly 790,000/month and growing, with a ComfyUI integration at 4,000+ stars. For developers building video pipelines, it sits in a distinctive position: open weights, native audio, real-time-class speed, and IC-LoRA controls — a combination that had no open-weight peer at launch and, as of review time, has only one known open-weight rival on the audio dimension (see Competitive Position, below).

This review covers the full LTX Video model family: what was built, how it evolved, what it does well, and where the tradeoffs are.

Company Background

Lightricks was founded in January 2013 by five PhD students from Hebrew University of Jerusalem: Zeev Farbman (CEO), Nir Pochter (CMO), Yaron Inger (CTO), Amit Goldstein (COO), and Itai Tsiddon. The founding team came from academic computer graphics and image processing — which shaped the company’s technical identity.

The product portfolio that followed reflects that lineage:

• Facetune (2013) — portrait editing, 200M+ downloads
• Photoleap — generative AI photo editing
• Videoleap — video editor, Apple’s 2017 App of the Year
• Popular Pays — influencer marketing platform (acquired 2022)
• LTX Studio — AI filmmaking platform (launched February 2024)
• LTX Video — open-weight video generation model family (launched late 2024)

Lightricks has grown into one of Israel’s most substantial AI companies through multiple funding rounds.

Confirmed funding history:

The Series D round was co-led by Insight Partners and Hanaco Venture Capital, with participation from Goldman Sachs Asset Management, Clal Tech, Harel Insurance and Finance, and Greycroft. The strategic framing was explicit: Lightricks would transform from a mobile app company into an AI platform company. LTX Video is the technical centerpiece of that transformation.

The Model Family

LTX Video has gone through two distinct architectural generations with four main versions worth understanding: the original LTXV 0.9.x line, then LTX-2, LTX-2.3, and LTX-2.5.

Generation 1: LTX-Video (LTXV 0.9.x)

Released in late 2024 alongside a technical paper (“LTX-Video: Realtime Video Latent Diffusion,” arXiv:2501.00103), the original LTXV introduced the architectural innovation that defines the entire family: an unusually aggressive Video-VAE with a 1:192 compression ratio.

Most video diffusion models work by compressing frames spatially, then applying the transformer. LTXV relocated the patchifying step — normally done at transformer input — to the VAE itself, achieving 32×32×8 pixel spatial-temporal downscaling per token. The result is that the transformer operates on a much smaller representation than competitors, enabling full spatiotemporal self-attention at manageable compute cost.

The practical claim: generate 5 seconds of 24 FPS video at 768×512 in approximately 2 seconds on an H100. Faster than real-time was not a figure of speech.

The LTXV 0.9.x branch continued iterating through early 2025:

• v0.9.6 (April 2025): Quality improvements, default 1216×704 at 30 FPS
• v0.9.7 (May 2025): 13B distilled variant, continued real-time generation claim
• v0.9.8 (July 2025): 60-second video support, distilled 2B and 13B variants, IC-LoRA control adapters

The 0.9.8 model lineup by size/speed tradeoff:

• ltxv-13b-0.9.8-dev — highest quality, full 13B
• ltxv-13b-0.9.8-distilled — 15× faster than dev, minimal quality drop
• ltxv-2b-0.9.8-distilled — fastest, lowest VRAM; full HD in ~10 seconds, low-res preview in ~3 seconds

All 0.9.x models are licensed under OpenRail-M (commercial use permitted). This is a permissive standard for open-weight AI models.

Generation 2: LTX-2 (October 2025)

LTX-2 is architecturally distinct from LTXV 0.9.x. The paper (“LTX-2: Efficient Joint Audio-Visual Foundation Model," arXiv:2601.03233, 29 authors) describes an asymmetric dual-stream transformer with bidirectional cross-attention: a 14B-parameter video stream and a 5B-parameter audio stream that attend to each other during generation.

This made LTX-2 the first open-weight video model to natively generate synchronized audio. The audio is not post-processed or added from a separate pipeline — it is generated jointly with the video, conditioned on the same text prompt. Speech, background sounds, and foley are all within scope.

The claimed efficiency: “up to 50% lower compute cost than competing models” with “state-of-the-art audiovisual quality and prompt adherence among open-source systems” and results “comparable to proprietary models at a fraction of computational cost.” These are Lightricks’ claims; independent benchmark verification was not accessible at review time.

Key specs:

• 19B parameters total (14B video + 5B audio)
• Up to 4K/50 FPS (video-only mode)
• Up to 10 seconds with synchronized audio
• Multilingual text encoder
• License: LTX-2 Community License Agreement (custom, not standard OSI)

The license change from OpenRail-M is worth noting. The LTX-2 Community License permits free use for personal, research, and commercial applications for organizations under $10 million in annual revenue; entities at or above that threshold must obtain a paid commercial license (Section 2.1 of the license text). It is not an OSI open source license. Review the full terms before deploying commercially.

Generation 2.3: LTX-2.3 (March 2026)

LTX-2.3 extended LTX-2 to approximately 22B parameters (checkpoint naming: ltx-2.3-22b), improving on LTX-2 across audio quality, visual quality, and prompt adherence, and shipped with a desktop video editor for local, consumer-hardware execution. As of late August 2026, HuggingFace reports roughly 1.5 million downloads/month for the LTX-2.3 checkpoint family — down from a peak near 1.73M/month in spring 2026 as usage has begun shifting to the newer LTX-2.5 release below.

Available variants: ltx-2.3-22b-dev, ltx-2.3-22b-distilled, ltx-2.3-22b-distilled-1.1, plus spatial and temporal upscaler checkpoints and IC-LoRA adapters.

Generation 2.5: LTX-2.5 (Current, August 2026)

Lightricks released LTX-2.5 on August 11, 2026, a 22B-parameter update to the same asymmetric dual-stream architecture (arXiv:2601.03233 covers the underlying LTX-2 design; the same paper ID is listed on the LTX-2.5 model card). It adds native multi-shot generation (a single generation can contain multiple connected cuts that preserve character, environment, lighting, and voice), an optional “auto duration” head that predicts appropriate clip length from the prompt, a new diffusion video decoder aimed at high-motion footage, native 4K HDR output, and a fine-tuned Gemma 4 12B text encoder. Clip length extends to roughly 6–20 seconds with synchronized audio — a substantial jump over LTX-2.3’s ~10-second audio-sync cap. Lightricks also shipped a dedicated checkpoint tuned for physical-AI/robotics simulation use cases. HuggingFace shows roughly 790,000 downloads/month for the new checkpoint family as of late August 2026, and Lightricks has said the LTX family overall has passed 33 million cumulative downloads.

Technical Architecture

Understanding what LTX Video does differently requires understanding the architecture at a level of specificity that most coverage skips.

The VAE as the key innovation. Video diffusion transformers have a fundamental tension: spatiotemporal attention over video frames is expensive, but coarser attention loses quality. Most models address this by limiting video length or reducing resolution. LTXV’s approach is to move patchifying (the process of chunking inputs into tokens) from the transformer input stage into the VAE encoding stage. The VAE compresses 32×32×8 pixels per token — meaning 32 pixels spatially in both dimensions and 8 frames temporally collapse into a single token. This 1:192 compression ratio is more aggressive than comparable architectures, reducing the sequence length the transformer operates on while preserving enough visual information for high-quality decoding.

The dual-stream transformer (LTX-2+). Audio and video are processed in separate transformer streams, with bidirectional cross-attention allowing each stream to condition on the other. This is architecturally different from concatenating audio features into the video token sequence — bidirectional cross-attention means the video stream influences audio timing and the audio stream influences visual motion. The practical result is lip-sync and foley coherence without a separate alignment step.

IC-LoRA controls. The IC-LoRA (Image Conditioning LoRA) adapter system lets users attach lightweight control signals to the generation pipeline without fine-tuning the base model. Available adapters: depth maps, human skeleton/pose, Canny edge detection, Union control (combined), HDR lighting, and motion tracking. Each adapter is a small checkpoint (~200MB range) compatible with ComfyUI workflows.

Distilled variants. The distilled checkpoints use knowledge distillation from the dev (full-quality) model to achieve dramatically faster inference with minimal quality loss. The 13B distilled is claimed at 15× faster than 13B dev — bringing 13B-quality output within reach of consumer hardware in practical workflows.

What It Generates

Video without audio: Text-to-video and image-to-video generation. Up to 60 seconds at 30 FPS at up to 4K resolution (practical limits depend on VRAM). Multi-keyframe conditioning lets users specify multiple reference images at temporal positions, enabling scripted narrative sequences. Forward and backward video extension from existing clips. All LTXV 0.9.8 features.

Video with audio: The LTX-2.x line generates synchronized speech, ambient sound, music elements, and foley from the same text prompt that drives the video. A scene with rain, dialogue, and footsteps can be prompted as a single generation pass. LTX-2.3 capped audio-synchronized output at approximately 10 seconds; LTX-2.5 extends this to roughly 6–20 seconds, plus native multi-shot generation that chains multiple cuts while preserving character, environment, and voice consistency across the sequence.

Control-conditioned video: IC-LoRA adapters allow generation conditioned on depth maps (spatial structure), pose skeletons (human motion), edge maps (structural outlines), or motion fields. This is useful for animation workflows where the spatial layout or motion arc needs to be precise.

Deployment Options

LTX Video is available through multiple access paths:

Self-hosted:

• HuggingFace: Lightricks organization — 38+ model checkpoints, including base models, FP8 quantized variants, IC-LoRA adapters, and upscalers
• GitHub: Lightricks/LTX-Video — 10,900+ stars, 1,100+ forks (verified via GitHub API, August 2026); a separate Lightricks/LTX-2 monorepo (inference + LoRA trainer for the LTX-2.x line) has passed 9,200 stars, 1,500 forks
• Requires Python 3.12+, CUDA 12.7+, PyTorch 2.7+ for LTX-2.x
• VRAM: 8GB estimated for 2B distilled; 24GB+ recommended for 13B distilled; 32GB+ for 13B dev workflows

ComfyUI:

• Lightricks/ComfyUI-LTXVideo — 4,000+ stars, 480+ forks (verified via GitHub API, August 2026), with official day-one workflow templates for LTX-2.5
• Available via ComfyUI Manager
• Supports all model variants, IC-LoRA adapters, HDR workflows, tiled VAE, two-stage upscaling pipelines
• Active maintenance with dedicated workflow examples

Diffusers: Native support via LTXConditionPipeline and LTXPipeline classes in Hugging Face Diffusers library.

Cloud APIs (pay-per-second, not per-video):

• Fal.ai — LTX-2.3: Fast variant $0.04/s at 1080p (up to $0.16/s at 4K); Pro variant $0.06/s at 1080p (up to $0.24/s at 4K)
• Fal.ai — LTX-2.5: Fast variant $0.09/s at 720p, $0.13/s at 1080p, up to $0.30/s at 4K; Pro variant $0.12/s at 720p, $0.17/s at 1080p
• Replicate also hosts multiple LTX-2.x endpoints (lightricks/ltx-2-fast, ltx-2-3-pro, ltx-2-retake, etc.); per-second pricing varies by endpoint and was not consistently published at review time, so no single figure is quoted here — check the Replicate model pages directly before budgeting
• Both platforms allow API-based generation without local hardware requirements

LTX Studio: Lightricks’ own consumer-facing filmmaking platform, now at ltx.io/studio (the old ltx.studio domain 301-redirects there; the in-app generation UI still runs at app.ltx.studio). Powered by LTX-2.5 but also integrating VEO 3.1, Kling, and other partner models. Credit-based subscription model ($0–$125/month, Standard plan at $35/month is the commercial-use entry point). The January 2026 ElevenLabs partnership adds audio-to-video generation capabilities. May 2026 “Canvas” update added collaborative workspace features.

MCP Server

There is no official MCP server from Lightricks or LTX Video. One community implementation exists on GitHub (sbdsam/ltx-video-mcp, created April 29, 2026) connecting Claude Code to the LTXV 0.9.8 pipeline via Docker and ComfyUI. It runs CPU-only with 5–60+ minute generation times — not practically useful for production workflows, and as of August 2026 it has not attracted meaningful community adoption (zero GitHub stars).

For AI-agent integration, the Fal.ai and Replicate APIs are the practical path. Both expose REST endpoints that any MCP server with HTTP capability can reach.

Ecosystem

The developer community around LTX Video is substantial for an open-weight model this young:

• LTX-Video GitHub: 10,900+ stars, 1,100+ forks; LTX-2 GitHub: 9,200+ stars, 1,500+ forks
• ComfyUI node: 4,000+ stars, 480+ forks
• Trainer repo: 466 stars, 66 forks
• HuggingFace: 38+ checkpoints, 100+ community Spaces; LTX-2.3 ~1.5M and LTX-2.5 ~790K monthly downloads (August 2026)
• Community tools: ComfyUI-LTXTricks (optimization workflows), TeaCache for LTX-Video (inference caching), LTX-VideoQ8 (8-bit quantization)
• Platform integrations: Fal.ai, Replicate, HuggingFace Diffusers, ComfyUI Manager

Competitive Position

LTX Video occupies a specific, and no longer uncontested, position in the open-weight video model landscape:

vs Wan2.1 (Alibaba/WanX): Wan2.1 has stronger raw visual quality in independent benchmark comparisons. Its Apache 2.0 license is more permissive than the LTX-2 Community License. But it generates video only — no audio. For pure visual quality, Wan2.1 is competitive or ahead. For audio-video workflows, Wan2.1 requires a separate model.

vs HunyuanVideo (Tencent): HunyuanVideo has the strongest benchmark scores among open-weight models for cinematic visual quality. But it is video-only, and inference is slower than LTX’s distilled variants. LTX-2 claims parity or better on compute efficiency.

vs CogVideoX (Zhipu/THUDM): CogVideoX tops out at 5B–10B parameters. LTX-2.3 at 22B is substantially larger, and adds audio. CogVideoX has a cleaner OSI-compatible license.

vs Sora/VEO 3.1/Kling (proprietary): Open-weight means local deployment, no usage fees at inference time, fine-tuning capability, and privacy. The tradeoff is setup complexity and hardware cost.

vs MOVA (OpenMOSS): LTX-2 is no longer the only open-weight option for single-pass, synchronized audio-video generation. OpenMOSS released MOVA on January 29, 2026 — about three weeks after LTX-2’s January 6 open-source launch — also generating video and audio in one inference pass with released weights, training code, and LoRA fine-tuning scripts. LTX-2 was first to production-ready open-weight audio-video generation, but the field is no longer uncrowded on that specific dimension; a direct head-to-head quality comparison between LTX-2.x and MOVA was not available at review time.

The current position: LTX Video was first to open-weight, single-pass audio-video generation and remains the most mature option on ecosystem size (GitHub stars, ComfyUI integration, cloud API availability), but it is no longer the only one. If you need coherent sound generated jointly with video — without stitching together separate models — LTX-2.5 and MOVA are, as of review time, the two known open-weight options.

Pricing

LTX Studio (Cloud — Lightricks):

Commercial rights begin at Standard tier. Pro adds VEO 3.1 access and collaboration features. Enterprise adds SSO, SOC 2/GDPR/ISO compliance, and custom model training options.

Cloud APIs (2026 pay-per-second pricing, not per-video):

• Fal.ai LTX-2.3: from $0.04/s (Fast, 1080p) to $0.24/s (Pro, 4K)
• Fal.ai LTX-2.5: from $0.09/s (Fast, 720p) to $0.30/s (Fast, 4K)
• Replicate hosts multiple LTX-2.x endpoints with per-second pricing that was not consistently published at review time — check current rates on the Replicate model pages before budgeting

Self-hosted: Free. Hardware cost is your own capital expense.

Limitations

Audio duration cap. LTX-2.3’s synchronized audio generation topped out around 10 seconds; LTX-2.5 raises this to roughly 20 seconds with native multi-shot chaining, but that is still short of the 60-second video-only capability inherited from LTXV 0.9.8. Long-form audio-synchronized video in a single generation pass is still not available.

License nuance, including a revenue cutoff. The LTX-2 Community License is not equivalent to OpenRail-M or Apache 2.0. It permits free use and commercial applications, but only for organizations under $10 million in annual revenue (Section 2.1); entities at or above that threshold must negotiate a paid commercial license. Review the full agreement for deployment-specific conditions, particularly around commercial derivative works.

Hardware requirements at scale. The 22B LTX-2.3/LTX-2.5 dev checkpoints require 32GB+ VRAM for ComfyUI workflows. The distilled variants reduce this significantly, but the best-quality output still demands professional GPU hardware; Lightricks’ own LTX-2.5 speed benchmarks were run on paired NVIDIA GB200 superchips, not consumer hardware.

Leaderboard position is now independently measurable, and it’s mid-pack. Artificial Analysis’s Video Arena places LTX-2.3 Pro at 13th of 27 models on its crowdsourced quality Elo score (1105.46) as of review time — a real independent data point that was not available at the previous audit, and it does not support “state-of-the-art” framing on raw visual quality. Lightricks’ own claims of parity with proprietary models on compute efficiency and audiovisual quality remain self-reported and should be weighed against this ranking.

Speed claims are hardware-specific. “Faster than real-time” applied to the original 2B model on an H100. Lightricks’ LTX-2.5 speed claim (a 10-second, 720p clip in ~6.8 seconds) was measured on two NVIDIA GB200 superchips, and its managed API took 23.7 seconds for the same clip at 1080p — both very different from typical consumer or single-GPU hardware.

LTX Studio credit friction. The cloud platform’s credit system (8,000/month on Lite, 28,000/month on Standard) can be opaque for estimating generation costs. Free tier is 800 credits total (not per month) — effectively a trial, not a usable free tier.

Who It’s For

LTX Video (open-weight) is for:

• Developers building video generation pipelines who need audio-video synchronization without separate model stitching
• ComfyUI power users building production workflows with control adapters
• Researchers who want to fine-tune on custom datasets (LoRA trainer available)
• Teams who need local deployment for privacy or cost reasons at scale
• Anyone building on Fal.ai or Replicate who wants a cost-effective open-weight alternative to proprietary APIs

LTX Studio is for:

• Individual creators and small teams who want the Lightricks model without setting up local infrastructure
• Commercial projects needing a filmmaking pipeline with multi-model access (VEO 3.1, Kling) under one subscription
• Enterprise teams needing SOC 2/GDPR compliance and admin controls (Enterprise tier)

LTX Video is probably not for:

• Users who want the absolute highest visual quality from an open-weight model (Wan2.1 and HunyuanVideo are competitive or ahead on raw visuals)
• Users who need a fully OSI open source license (the LTX-2 Community License is custom)
• Non-technical users — the self-hosted path requires GPU hardware and technical setup

Assessment

LTX Video is the most technically ambitious open-weight video model currently available. The 1:192 VAE compression that enabled real-time generation in 2024, the dual-stream audio-video transformer in LTX-2, and the IC-LoRA control system represent genuine architectural contributions rather than scale-only improvements. Combined LTX-2.x monthly HuggingFace downloads (roughly 1.5M for LTX-2.3 plus ~790K for LTX-2.5, as of August 2026) and 10,900+ GitHub stars on the original repo indicate real, sustained developer adoption — Lightricks shipped a full point-release (LTX-2.5) within seven months of LTX-2’s January 2026 launch.

The primary limitation is that “best open-weight model” is a competitive position that shifts, and it is now independently measurable: Artificial Analysis’s Video Arena places LTX-2.3 Pro 13th of 27 models on quality Elo, meaning Wan2.1 and HunyuanVideo remain genuinely competitive or ahead on visual quality for video-only use cases. What LTX Video has that they don’t is audio — though that gap has already started closing: OpenMOSS’s MOVA shipped a competing open-weight single-pass audio-video model in January 2026, three weeks after LTX-2.

The business model is coherent: open weights build developer community and trust, which funnels into LTX Studio subscriptions, gated by a $10M-annual-revenue commercial license threshold for larger users of the open weights themselves. Lightricks has demonstrated consistent growth across its app portfolio and multiple funding rounds, positioning it to keep the model family competitive.

Rating: 4/5. The audio-video architecture is genuinely differentiated, and LTX-2.5’s multi-shot generation and extended audio duration (up to ~20 seconds) address real gaps from LTX-2.3. The ComfyUI ecosystem is mature. The distilled variants make real-time generation accessible on consumer hardware. Deductions for the custom license (not OSI, with a $10M revenue cutoff), visual quality that independently verified benchmarks now show trailing Wan2.1/HunyuanVideo, and lack of an official MCP server for AI agent integration.

ChatForest reviews AI tools based on publicly available research. We do not conduct hands-on testing. This review was last updated August 25, 2026.