LightX2V Studio pricing, free plan and limits

Last updated: 2026-10-28

LightX2V Studio is a hosted video, avatar and image platform that runs 19 models, among them Wan, LTX, MiniMax-H3, SekoTalk, GPT Image and Nano Banana, from one credit balance. It is built on the open-source LightX2V inference framework, which showed 2,964 GitHub stars in October 2026.

About LightX2V Studio

LightX2V Studio is the hosted creation platform built on LightX2V, an Apache 2.0 inference framework for image and video models that the ModelTC organization started on GitHub in March 2025. The framework is the engineering project; the Studio is the point-and-click product on top of it: choose a task, choose a model, spend credits, download the result. According to its terms of service, the Studio is operated by Beijing Liuguang Technology Store, an individual-business registration in Beijing.

Where Kling offers its own video models, the Studio works as a switchboard. The vendor's live model list, read on 2026-10-28, spans four SekoTalk avatar versions, Wan 2.2 for image-to-video, text-to-video and character replacement, LTX-2.3 from Lightricks and MiniMax-H3 for audio-synced clips, SwiftVR for restoration, and image models that include GPT Image 2, Nano Banana Pro (listed on HokAI as Gemini 3 Pro Image) and Qwen Image. The list marks every model as available on every plan, Free included, although heavier models burn credits faster.

Avatars lead the homepage. SekoTalk V3 handles single-speaker talking video, V3 Multi handles scenes with several speakers, and a separate flow re-dubs existing footage with new speech. Agent mode, introduced in the 2026-07-15 update, plans multi-step jobs from uploaded media, and Canvas mode saves a chain of generation steps as a reusable workflow. For scripts and third-party agents there is an OpenAPI schema and an official Skills repository. If you need a finished presenter library with templates and corporate controls, look at HeyGen first, then Synthesia; the Studio's draw is model choice and a low entry cost. Specialists cover single jobs better: D-ID for talking-head avatars, ElevenLabs when voice is the whole task, and Midjourney when still images are. For a dedicated video lab's own models, Runway fits better, and Pippit fits when posting the clip is part of the job.

The open-source side is the unusual part. The project publishes step-distilled acceleration LoRAs on Hugging Face that cut typical 40 to 50 step inference to four steps, and its About page reports hundreds of thousands of downloads for the Qwen-Image Lightning LoRA as of August 2026. A technical buyer can therefore check the speed claims or self-host the framework instead of trusting the hosted service. Developers who only want to call hosted open models over an API can compare Replicate, and anyone who mainly wants LTX output can go straight to Lightricks' own LTX Studio. The trade-off is operator risk: this is a young product from a small operator, so procurement teams should weigh that before picking it over Higgsfield.

Pricing

Free: claim 300 credits a day (they expire at the end of the day), and new accounts receive 1,000 credits valid for 7 days. Memberships are one-time purchases that do not renew automatically, sold on monthly, quarterly and annual terms. 99 for 300,000.

99, with credits granted in monthly batches. 99. 99 for the monthly term, so those pay prices are promotional and can change.

The Standard card also lists 30K, 50K and 175K credit sizes beside the 110K size priced here. 99 buys 160,000. 33 on monthly memberships, so packs only make sense as a top-up.

Completed payments are generally non-refundable apart from duplicate charges, non-delivery and payment errors. Prices were read in USD on the English site on 2026-10-28; the Chinese site bills in CNY.

Key Features

• Avatar video with SekoTalk: Four SekoTalk versions turn a portrait and speech into talking video, from V3 for one speaker to V3 Multi for several, with clips up to five minutes on Standard and Pro and two minutes on Lite.
• Image-to-video and text-to-video: Wan 2.2 distilled models cover image-to-video, text-to-video and first-and-last-frame clips, while MiniMax-H3 and LTX-2.3 add audio-synced output.
• Character replacement and motion transfer: Wan 2.2 Animate swaps a subject into existing footage or copies a reference performance onto a new character, with input capped at 20 seconds on Lite and 60 seconds on Pro.
• Image generation and editing: Seven text-to-image models include GPT Image 2.5 Flare, Nano Banana Pro and Qwen Image 2512, and image-to-image editing runs on GPT Image, Qwen Image Edit 2511 and Nano Banana.
• Agent and Canvas modes: Agent mode breaks a creative brief into steps using your uploaded media and Skills, while Canvas mode saves node-based chains so a repeated job runs the same way each time.
• API and Skills: An OpenAPI schema, a live model-spec endpoint and an official Skills repository on GitHub let scripts and third-party agents call the same models, and API access is listed as included on Standard and Pro.
• Voice tools: Voice cloning learns from up to 15 seconds of audio, and a dual-host podcast generator writes two to ten minute episodes that can feed avatar videos; both are free with any paid plan.

Pros

• Memberships are one-time purchases, so there is no surprise renewal charge to cancel.
• The open-source roots can be checked: the framework, acceleration LoRAs and docs are public, so a technical buyer can verify speed claims or self-host instead of relying on the hosted service.
• The privacy policy states that uploads and outputs are not used for model training by default, and a future training program would need a separate opt-in.
• You or the original rights holder keep rights in inputs and outputs, and the platform does not claim ownership just by running the job.

Cons

• The operator is an individual-business registration in Beijing, and no trust-center page with SOC 2, ISO 27001 or HIPAA documentation was found, which rules it out for most regulated teams.
• Face and voice uploads may be processed by outside compute, storage and model providers, and the policy says data may be handled outside mainland China.
• Costs are hard to predict because credit use depends on model, resolution, duration and parameters; the quote shown before you submit is the only authoritative figure.
• No independent review of output quality was found when this record was checked, so quality claims rest on the vendor's own samples.

Visit LightX2V Studio Official Website