Developer video products
Teams that need a documented API for text, image, or audio-conditioned video with selectable speed and quality.
Independent tool overview
LTX-2 Model Family at a glance
LTX-2 is Lightricks' open-weight audio-video model family for generating synchronized video and sound from text, images, video, or audio. The original LTX-2 API models have been removed, but the product line remains active through LTX-2.3 and the current LTX-2.5, available as downloadable weights, a self-hosted stack, and per-second APIs. LTX-2.5 adds native multi-shot continuity, stronger prompt adherence, automatic duration, improved motion, fine-tuning, and higher-end HDR and RAW workflows. It is a serious developer and production option, but self-hosting is hardware-intensive and the custom community license—not an unrestricted permissive software license—requires paid terms for entities at or above $10 million in annual revenue outside limited noncommercial use.
Visit the official LTX-2 Model Family site ↗
- Current flagship
- LTX-2.5
- Also supported
- LTX-2.3, including editing endpoints not yet available on 2.5
- Original API status
- ltx-2-fast and ltx-2-pro were removed August 16, 2026
- Generation modes
- Text-to-video, image-to-video, and audio-to-video with synchronized sound
- LTX-2.5 API variants
- Fast for speed and up to 4K; Pro for higher fidelity up to 1080p
- Orientations
- Native landscape 16:9 and portrait 9:16
- Frame rates
- 24 or 25 fps, with 48 or 50 fps on supported model and resolution combinations
- Durations
- 6–20 seconds depending on model, resolution, frame rate, and endpoint
- Deployment
- LTX API, downloadable weights, self-hosted or on-premises inference, and fine-tuning
- License
- Custom LTX-2.x Community License, not an unrestricted Apache or MIT model license
- Revenue threshold
- Entities at or above $10M annual revenue generally need paid commercial terms for commercial use
- Local hardware
- Vendor page advertises a 16GB minimum path; training guidance recommends 80GB standard or 32GB low-memory VRAM
- API billing
- Per generated second for text/image video and per input-audio second for audio-to-video
- Reviewed
- August 31, 2026 from current LTX model, API, repository, model card, pricing, changelog, and license sources
Overview
What LTX-2 Model Family is
LTX-2 began as Lightricks' joint audio-video diffusion-transformer family, generating visuals and synchronized dialogue, ambience, sound effects, or music in one process. The current flagship is LTX-2.5, while LTX-2.3 remains available for API workflows including Retake, Extend, and Reframe that 2.5 does not yet expose.
The original API identifiers, ltx-2-fast and ltx-2-pro, are no longer usable. LTX announced deprecation in July 2026 and removed them on August 16, 2026. Existing integrations must select ltx-2-3-fast, ltx-2-3-pro, ltx-2-5-fast, or ltx-2-5-pro according to endpoint support.
LTX-2.5 supports text-to-video, image-to-video, and audio-to-video through both synchronous and asynchronous APIs. Fast reaches landscape or portrait 4K and up to 50 fps; Pro currently reaches 1080p. Supported durations vary by model, resolution, and frame rate, with automatic duration available for text and image generation.
LTX-2.5's headline addition is native multi-shot generation: one run can create connected shots while trying to preserve character, setting, lighting, visual style, and voice across cuts. Other additions include a new diffusion video decoder, a Gemma 4-based text encoder and prompt enhancer, cleaner motion, fine-tuning, and professional HDR and RAW claims.
Developers can download weights and code, run the stack locally or on-premises, build ComfyUI and Python workflows, and train LoRA or IC-LoRA adaptations. Lightricks advertises inference paths starting around 16GB VRAM, but full training documentation recommends Linux, CUDA, and far larger GPUs—80GB VRAM for the standard configuration or 32GB with low-memory options.
The weights use Lightricks' LTX-2.x Community License. It permits broad use subject to restrictions for qualifying users, but entities with at least $10 million in annual revenue need a paid commercial agreement for commercial use. Redistribution, derivatives, acceptable-use obligations, downstream notices, and model-specific component licenses require review.
Generated video is not factual evidence and can contain identity drift, temporal artifacts, impossible physics, lip-sync errors, visual bias, offensive material, or misleading depictions. Production deployment needs provenance, consent, disclosure, rights review, content moderation, human approval, and conventional post-production.
Use cases
Who LTX-2 Model Family is best for
The strongest fit depends on the job you need the product to complete, not the size of its feature list.
Private deployment
Qualified organizations that want open weights and infrastructure control instead of a mandatory hosted API.
Custom video models
Technical teams with licensed datasets and suitable GPUs for LoRA, IC-LoRA, or full fine-tuning.
Multi-shot concepts
Creators exploring connected scenes with more consistent characters, settings, lighting, style, and voice.
Audio-led visuals
Music, dialogue, or sound-driven workflows where generated motion and imagery should follow an input track.
Production pipeline experiments
Studios evaluating 4K, HDR, RAW, portrait, high-frame-rate, and self-hosted generation inside conventional finishing workflows.
Capabilities
Core LTX-2 Model Family features
Joint audio-video generation
Generates video and synchronized sound within the same model family rather than stitching independent systems together.
Three input modes
Supports text-to-video, image-to-video, and audio-to-video in current LTX-2.5 API variants.
Native multi-shot
LTX-2.5 can create multiple connected shots in one run while attempting to hold identity, scene, lighting, style, and voice.
Automatic duration
Lets LTX-2.5 choose an allowed clip length from the requested action for text and image generation.
Resolution and orientation choices
Fast supports portrait or landscape output through 4K; Pro supports portrait or landscape through 1080p.
Camera controls
API parameters can request supported camera movements and first or last-frame conditioning.
Open weights and code
Provides model checkpoints, Python inference packages, training code, ComfyUI support, and reference workflows.
Fine-tuning
Supports trainable development checkpoints, LoRA, IC-LoRA, and full fine-tuning on appropriately licensed data.
Professional formats
LTX-2.5 is marketed with native 4K HDR and RAW-oriented workflows for downstream color and finishing.
Editing ecosystem
LTX-2.3 currently supplies Retake, Extend, and Reframe API endpoints while 2.5 focuses on generation.
Sync and async APIs
Returns a video directly for synchronous requests or a job ID for polling in longer asynchronous workflows.
Local desktop option
Lightricks maintains an open-source LTX Desktop application for local LTX-model video workflows.
Process
How the LTX-2 Model Family workflow works
Step 1
Choose the current model
Use LTX-2.5 Fast or Pro for current generation; retain LTX-2.3 when you need Retake, Extend, Reframe, or a lower published rate.
Step 2
Audit the license
Calculate revenue across the covered entity and affiliates, classify commercial versus noncommercial use, and review redistribution, derivative, attribution, acceptable-use, and downstream obligations.
Step 3
Choose API or self-hosting
Compare per-second cloud costs with GPU acquisition, storage, bandwidth, operations, security, electricity, model updates, and staff time.
Step 4
Clear every input
Obtain rights and consent for scripts, music, voices, faces, reference images, footage, trademarks, datasets, and fine-tuning material.
Step 5
Protect sensitive assets
Use approved storage and signed uploads, strip unnecessary metadata, set retention rules, and prefer a controlled self-hosted boundary when cloud processing is not permitted.
Step 6
Storyboard before prompting
Define shot purpose, subject, action, environment, camera, lens, light, timing, sound, continuity anchors, duration, aspect ratio, and exclusions.
Step 7
Prototype cheaply
Test prompt structure, motion, and composition with Fast or a lower resolution before paying for high-resolution or Pro renders.
Step 8
Lock continuity references
Use consistent, authorized reference assets, first and last frames, seeds, adapters, and descriptive anchors across shots.
Step 9
Moderate inputs and outputs
Block deceptive impersonation, non-consensual likeness use, sexual abuse material, dangerous content, rights violations, and other prohibited uses before generation or delivery.
Step 10
Review frame by frame
Inspect identity, hands, mouths, text, logos, continuity, physics, flicker, cuts, lip sync, audio artifacts, and unintended background content.
Step 11
Finish conventionally
Edit, color-grade, mix, caption, normalize sound, clear music and likeness rights, add disclosures or provenance, and export through a controlled mastering process.
Step 12
Measure and cap spend
Set duration and resolution limits, disable unsafe auto top-up defaults, monitor failures and retries, and account for the maximum credit hold when using automatic duration.
Step 13
Deploy with human approval
Require an accountable reviewer before publishing, advertising, news use, client delivery, or any high-impact decision.
Cost
LTX-2 Model Family pricing and free plan
LTX-2.5 weights can be downloaded and self-hosted under the custom community license, with no per-generation fee from Lightricks for qualifying use; infrastructure is still a real cost. Entities with annual revenue of at least $10 million generally need a paid commercial agreement for commercial use. The hosted API uses prepaid or approved postpaid billing per second. Automatic-duration jobs temporarily hold enough prepaid credit for the maximum possible clip, then release the unused remainder.
Open-weight self-hosting
No LTX usage fee for qualifying users
For local, on-premises, fine-tuned, or embedded deployment under the current LTX-2.x Community License.
- Downloadable model weights and code
- No mandatory API or per-generation charge from Lightricks
- Free commercial use is marketed for entities under $10M annual revenue, subject to the full license
- Entities at or above $10M generally need paid commercial terms for commercial use
- GPU, storage, electricity, bandwidth, engineering, security, and maintenance costs are separate
- Component and dependency licenses may add obligations
LTX-2.5 Fast API
$0.09–$0.30 per second
The current speed-focused API with portrait and landscape output through 4K.
- 720p: $0.09 per generated second
- 1080p: $0.13 per generated second
- 1440p: $0.19 per generated second
- 4K: $0.30 per generated second
- Text-to-video and image-to-video bill output duration
- Audio-to-video bills input-audio duration
LTX-2.5 Pro API
$0.12–$0.17 per second
The current higher-fidelity API option at 720p or 1080p.
- 720p: $0.12 per generated second
- 1080p: $0.17 per generated second
- Maximum listed output resolution is 1080p
- Text-to-video and image-to-video bill output duration
- Audio-to-video bills input-audio duration
LTX-2.3 API
$0.03–$0.34 per second
Still-supported Fast and Pro models, including current editing endpoints.
- Text/image Fast: $0.03 at 720p, $0.06 at 1080p, $0.12 at 1440p, $0.24 at 4K
- Text/image Pro: $0.04 at 720p, $0.08 at 1080p, $0.16 at 1440p, $0.32 at 4K
- Audio-to-video Pro: $0.06 at 720p, $0.10 at 1080p, $0.18 at 1440p, $0.34 at 4K
- Retake currently uses LTX-2.3 Pro at $0.10 per input-video second
- Endpoint compatibility matters more than version number alone
Enterprise license and API
Custom
For larger commercial entities, negotiated licensing, volume, support, or deployment terms.
- Contact LTX sales
- Review the binding license rather than relying on the marketing summary
- Clarify affiliates, revenue calculation, derivatives, fine-tunes, redistribution, indemnity, support, and model updates
- Enterprise API terms may differ from public prepaid pricing
Pricing checked . Check current pricing at the source ↗
Assessment
LTX-2 Model Family strengths and limitations
Where it stands out
- Current family combines synchronized audio and video generation in one model stack
- LTX-2.5 supports text, image, and audio conditioning through API or self-hosting
- Native multi-shot generation addresses a real continuity need across cuts
- Fast reaches native portrait or landscape 4K and high frame rates on supported settings
- Open weights and code support inspection, private deployment, customization, and fine-tuning
- Documented Python, ComfyUI, training, sync API, and async API paths
- Per-second API pricing is public and granular by model and resolution
- LTX-2.3 remains available for editing endpoints and lower-cost options
- Qualifying smaller entities can self-host without LTX usage fees under the community license
- Model card explicitly states factual, bias, prompt-following, and inappropriate-output limitations
What to consider
- The original ltx-2-fast and ltx-2-pro API identifiers have been removed and now return errors
- The latest model is LTX-2.5, so older tutorials and checkpoints can be stale or incompatible
- LTX-2.5 does not currently support the Retake, Extend, or Reframe API endpoints available with LTX-2.3
- The custom community license is not equivalent to an unrestricted Apache, MIT, or OSI-approved model license
- Entities at or above the $10M revenue threshold face paid-license requirements for commercial use
- The license contains use restrictions, derivative and redistribution duties, downstream-notice requirements, and other legal obligations
- Self-hosting avoids API fees but requires substantial GPU, memory, storage, energy, deployment, security, and maintenance capacity
- Full training is far more demanding than the advertised minimum inference path
- Local weights and dependencies create software-supply-chain, checkpoint-integrity, and update-management risks
- Generated output is not factual and can depict events, people, products, or locations inaccurately
- Identity, wardrobe, objects, text, lighting, motion, voice, and environment can drift within or across shots
- Hands, mouths, collisions, camera motion, physics, cuts, lip sync, and audio can contain visible or audible artifacts
- Multi-shot consistency is an attempted capability, not a guarantee
- High resolution and frame rate do not make a clip production-ready or rights-cleared
- Automatic duration can require a larger prepaid credit hold and may produce a longer clip than budgeted
- Retries, failed generations, prompt testing, storage, transfer, and downstream rendering add cost
- Fine-tuning on faces, voices, branded footage, music, or copyrighted work requires rights, consent, governance, and abuse controls
- Open deployment shifts content moderation, provenance, cybersecurity, privacy, and legal responsibility to the operator
- Deepfakes, deceptive media, non-consensual likeness use, political content, and youth safety require strict policy and human review
- API or app privacy, retention, subprocessors, residency, deletion, and training-use terms must be checked before sensitive uploads
Compare
LTX-2 Model Family alternatives
The right alternative depends on the specific output, workflow, controls and budget your project requires.
Content Creator
Runway Gen-4.5
Choose Runway Gen-4.5 for a managed creative platform and commercial video workflow without self-hosting model weights.
Explore Runway Gen-4.5 →Content Creator
Veo 3.1
Choose Google Veo 3.1 for Google's hosted high-end video generation and its surrounding Flow and Vertex AI ecosystem.
Explore Veo 3.1 →Content Creator
Sora 2
Choose Sora 2 for OpenAI's integrated consumer and creative video experience.
Explore Sora 2 →Content Creator
Wan 3.0
Choose Wan 3.0 when evaluating Alibaba's current multimodal video family and alternative deployment options.
Explore Wan 3.0 →Content Creator
Kling 3.0
Choose Kling 3.0 for another hosted native-audio video model with long-clip and consistency features.
Explore Kling 3.0 →Content Creator
Hailuo AI
Choose Hailuo AI for a simpler hosted video-generation interface aimed at creators rather than model operators.
Explore Hailuo AI →Questions
LTX-2 Model Family FAQs
What is LTX-2?
LTX-2 is Lightricks' open-weight audio-video foundation-model family. It generates synchronized video and sound from text, images, video, or audio and can be accessed through weights, self-hosted tools, and LTX APIs.
Is LTX-2 still active?
Yes, as a model family. The original ltx-2-fast and ltx-2-pro API IDs were removed in August 2026, but LTX-2.3 and the current LTX-2.5 are active successors.
What is the latest LTX-2 model?
LTX-2.5 is the current flagship as of August 31, 2026. It adds native multi-shot generation, stronger prompt adherence, cleaner motion, automatic duration, fine-tuning support, and professional HDR and RAW-oriented workflows.
What should developers replace ltx-2-fast and ltx-2-pro with?
Choose among ltx-2-3-fast, ltx-2-3-pro, ltx-2-5-fast, and ltx-2-5-pro. Confirm the endpoint matrix first because Retake, Extend, and Reframe currently remain on LTX-2.3.
Is LTX-2 open source?
Lightricks publishes weights and code and markets the family as open source. The model weights are governed by a custom LTX-2.x Community License with revenue thresholds, use restrictions, and redistribution duties, so teams should describe it precisely as open-weight and review the binding terms.
Can I use LTX-2 commercially for free?
Lightricks says entities under $10 million in annual revenue can use it commercially without an LTX usage fee, subject to the full community license. Entities at or above that threshold generally need paid commercial terms. Legal counsel should evaluate the actual entity, affiliates, purpose, deployment, derivatives, and distribution.
How much does the LTX-2.5 API cost?
LTX-2.5 Fast is $0.09 per second at 720p, $0.13 at 1080p, $0.19 at 1440p, and $0.30 at 4K. Pro is $0.12 at 720p and $0.17 at 1080p. Text and image generation bill output duration; audio-to-video bills input-audio duration.
Can LTX-2.5 generate 4K video?
Yes, the Fast API supports native portrait and landscape output through 4K on supported duration and frame-rate combinations. The current Pro API tops out at 1080p.
Can LTX-2.5 create multiple shots?
Yes. Native multi-shot generation is a headline 2.5 feature intended to preserve character, setting, lighting, style, and voice across connected cuts. Review output closely because continuity is not guaranteed.
How much VRAM does local LTX-2.5 need?
LTX's product page advertises a minimum path around 16GB VRAM, but the exact requirement depends on checkpoint, precision, resolution, duration, backend, and pipeline. Training documentation recommends 80GB VRAM for a standard setup or 32GB with low-memory options.
Does LTX-2.5 include audio?
Yes. It can generate synchronized audio with text- or image-generated video and can generate visuals driven by an input audio track. Sound, speech, music, timing, consent, and licensing still require review.
Is LTX-2 safe for production use?
It can be part of a production pipeline when the operator supplies security, rights management, moderation, provenance, quality control, human approval, and post-production. The model card says it is not factual and may produce biased, inappropriate, or prompt-misaligned output.
Should I use LTX-2.3 or LTX-2.5?
Use 2.5 for its newer generation quality, native multi-shot, automatic duration, and current foundation. Use 2.3 when you need current Retake, Extend, or Reframe endpoints, broader Pro resolution tiers, or its lower published API pricing.
Who owns LTX-2 output?
The model license says Lightricks generally claims no rights in generated output, but the user remains accountable for inputs and uses. That does not clear copyrights, trademarks, publicity rights, privacy, consent, music, dataset rights, or local-law requirements.
Bottom line
Our LTX-2 Model Family verdict
LTX-2 remains one of the most important open-weight video families, but the page must be read as a moving product line rather than the frozen January 2026 release. LTX-2.5 is now the strongest starting point for new generation work, while 2.3 still matters for editing and lower-cost API paths. The combination of synchronized audio, multi-shot continuity, downloadable weights, fine-tuning, and public per-second pricing is unusually attractive to technical teams. The tradeoff is operational and legal ownership: capable GPUs, a custom revenue-gated license, model and dependency security, moderation, rights clearance, and frame-level human review all sit with the deployer.
Visit LTX-2 Model Family website ↗Similar Content Creator Tools
View all Content Creator AI tools →
Hunyuan Motion 1.0
Hunyuan Motion 1.0 - Tencent's open-source model for 3D character animations from text prompts
Scribe v2
Scribe v2 - ElevenLabs' SOTA transcription model with top accuracy, multi-language support, keyterm prompting, and more
Ray 3.2 Modify
Ray3 Modify - Edit and reimagine videos with precise keyframe and character reference controls
GLM-Image
GLM-Image - Z AI's new open-source image generation model
AI training for the future of work.
Get access to all our AI courses, hundreds of real-world AI use cases, live expert-led workshops, an exclusive network of AI early adopters, and more.
AI Courses
Get unlimited access to all of our current & upcoming industry-specific AI courses for the duration of your subscription.
Daily Guides
To keep up with the rapid pace of AI, our team publishes AI implementation guides daily. Our library contains 300+ practical use cases to automate real-world work.
Workshops
Join weekly, live, interactive sessions with industry leaders who are at the forefront of AI for hands-on implementation guidance and exclusive insights.
Community
Network with an exclusive community of AI-first professionals who are working smarter with AI. Learn how early adopters are using AI in their work and businesses.