- The choice is one number against another: the LTX API bills per second of generated video, starting at $0.09 for
ltx-2-5-fastat 720p, while self-hosting has no meter and costs you GPUs instead. Everything else follows from that split. - The crossover is utilization, not preference. Metered billing wins at low or bursty volume; fixed infrastructure wins at high steady volume, and the gap widens with every clip.
- Self-hosting is the only route that gives you fine-tuning on your own material, data that never leaves your network, and the full pipeline surface. The gate is 80GB or more of VRAM.
LTX-2.5 runs four ways, and the choice comes down to one number against another. The LTX API bills per second of generated video, starting at $0.09 per second for ltx-2-5-fast at 720p. A self-hosted deployment has no API meter at all and costs you infrastructure instead. Everything else, including setup effort, fine-tuning access and data residency, follows from that split.
The four routes are the first-party LTX API, self-hosting on your own GPUs, Hugging Face as the distribution point for the weights, and third-party serverless platforms such as fal and Replicate. Most teams weigh the first two. The third is a prerequisite rather than an alternative, and the fourth is the one people forget exists.
What actually differs
Differences reduce to five axes, scored in the table further down: cost model, setup burden, control over fine-tuning and pipeline choice, data residency, and latency. Cost model and setup burden decide it for most teams.
Every route runs the same model family, and LTX-2.5 generates synchronized audio and video from a single model rather than adding sound to silent footage afterward.
Route 1: the LTX API
The LTX API is the fastest path from zero to a generated clip: get a key from the developer console, send a POST request, get a video back. It exposes endpoints for text-to-video, image-to-video, audio-to-video, extend and retake, each authenticated with a Bearer token.
Pricing is per second of generated video, set by model variant and output resolution, with no request fees or per-asset charges. For ltx-2-5-fast that is $0.09 per second at 720p, $0.13 at 1080p, $0.19 at 1440p and $0.30 at 4K. For ltx-2-5-pro it is $0.12 at 720p and $0.17 at 1080p. Audio-to-video is priced at $0.13 per second on Fast and $0.17 on Pro, at 1080p.
Both variants output 720p through 4K at 24, 25, 48 or 50 FPS. Duration is a fixed set rather than a range: 6, 8 or 10 seconds in most combinations, extending to 20 in even steps only on ltx-2-5-fast at 720p or 1080p on 24 or 25 FPS.
Billing behaviour matters before you write retry logic. Prepaid organizations have credits reserved before generation starts and are charged only on success, so a failed job releases its reservation rather than consuming balance. Insufficient balance returns 402 Payment Required.
Enterprise organizations bill postpaid by invoice with no pre-generation check. An unauthenticated GET /pricing endpoint returns the full schedule for every SKU, which beats hardcoding rates into a cost estimator. Our guide to integrating a video generation API covers implementation.
Route 2: self-hosting
Self-hosting removes the meter entirely. Under the LTX Model License, use is free for organizations under $10M in annual revenue, with a commercial license required above that threshold that includes the full model weights plus engineering support for production deployment on your own infrastructure. There is no per-generation charge on a license. You pay for the GPUs.
The gate is hardware. The codebase is pure PyTorch, is tested with CUDA 12.7 or newer, and targets NVIDIA GPUs with 80GB or more of VRAM, with distilled variants supported on 32GB when quantization is enabled. The quick-start component set is roughly 66 GiB on disk.
What you get is not only cheaper generation at volume. Self-hosting is the one route that permits fine-tuning on proprietary IP, and the one where footage never leaves your network. Our private deployment guide covers the enterprise case.
Route 3: Hugging Face
Hugging Face is where the LTX-2.5 weights are published, one file per component, so you download only the parts your pipeline needs. That makes it the starting point of every self-hosted deployment rather than a separate route.
Two notes save time. The repository is gated: a 401 or 403 means you have not accepted the model terms or are not logged in with a Read token, and fine-grained tokens need the "read gated repos" scope enabled. And checkpoints are not interchangeable across versions. Every pipeline also runs on LTX-2.3, but its files bundle components differently.
Route 4: fal and Replicate
Third-party inference platforms host the model on their own serverless GPUs and bill per use, removing the hardware gate. LTX-2.5 is listed on fal, which also carries a large library of LTX fine-tunes, and LTX models are published on Replicate under the Lightricks account.
The tradeoff is specificity. These platforms expose whichever variants and parameters they chose to surface, a subset of what the local pipelines accept. Open-source-only settings such as cfg_scale and the guidance parameters are not necessarily available, and neither is arbitrary LoRA loading. In exchange you get no setup and no weights to store.
Platform catalogues move faster than documentation, so check the current listing on the platform itself before building against a specific variant.
How do the routes compare?
Comparison across the five axes:
Where is the cost crossover?
The crossover is a utilization question, not a preference. Below a certain volume the API is cheaper, because a metered service with zero fixed cost beats idle hardware. Above it, self-hosting wins and the gap widens with every generation, because the fixed cost is already paid and one more clip costs electricity.
You can locate your own crossover with two numbers. Take your monthly seconds of generated video at your target resolution, multiply by the published per-second rate, and compare against the monthly cost of a GPU that clears the 80GB tier plus the ops time to keep it running. Pull rates from GET /pricing rather than from any blog post, including this one.
Two caveats. Self-hosting has a setup cost that recurs on every upgrade and never appears in a per-second comparison. And the crossover moves if volume is spiky, because fixed infrastructure is priced for your peak while metered billing is priced for your average. Our breakdown of LTX API pricing has the full rate picture.
Which route should you pick?
Route selection resolves cleanly once you know which of five situations you are in.
• Low or bursty volume: the LTX API. Fixed infrastructure cannot compete with zero fixed cost at low utilization.
• High, steady volume: self-host. The crossover arrives faster than most teams expect.
• IP-sensitive material: self-host. Settled before cost enters the conversation, and what air-gapped deployment exists for.
• You need to fine-tune: self-host, because adapting the model to your own style or footage requires the weights.
• No GPU and no ops capacity: the LTX API, or a third-party platform if you already bill through one.
Embedding video generation into a product you sell? Check the licensing position first, because the revenue threshold governs that independently of which route is convenient.
How fast is self-hosted LTX?
Self-hosted latency depends on pipeline choice and GPUs, and both are levers you control. DistilledPipeline is the fastest documented path, running with 8 predefined sigmas, 8 steps in stage 1 and 4 in stage 2.
The repository also documents multi-GPU inference for running a single generation across several GPUs to reduce latency, including sequence parallelism, distributed VAE and distributed text encoding. That lever does not exist on a metered route.
One constraint applies to duration planning on every route. The Video VAE requires frame count to satisfy (F-1) % 8 == 0, so valid counts at 25fps include 25, 33, 41, 49, 57, 65, 73, 81, 89, 97 and 121. Not every duration in seconds is achievable, which is worth checking before you promise a clip length.
Summary
LTX-2.5 runs on the first-party LTX API, on your own hardware, or through third-party platforms, with Hugging Face serving as the distribution point for the weights. The LTX API bills per second of generated video with no request or per-asset fees, from $0.09 per second for ltx-2-5-fast at 720p up to $0.30 at 4K.
Self-hosting carries no per-generation charge and is free under the LTX Model License below $10M annual revenue, but targets NVIDIA GPUs with 80GB or more of VRAM, a CUDA stack tested at 12.7 or newer, and roughly 66 GiB of weights on disk. The crossover is a function of utilization: metered billing wins at low or bursty volume, fixed infrastructure wins at high steady volume. Self-hosting is the only route that permits fine-tuning on proprietary material.
Still choosing a model family rather than a route? Our open versus closed comparison is the better starting point.
