LTX 2.5: Video Pricing, Local Setup and the 2.3 Comparison

The current API names are ltx-2-5-fast and ltx-2-5-pro. Fast prioritizes speed and cost; Pro is the higher-fidelity option. Both accept text, image or audio input and can produce landscape or portrait video up to 4K. Lightricks describes 2.5 as adding native multi-shot generation and an improved video decoder.

Keep the hosted API names separate from local checkpoints. The downloadable package offers dev and distilled transformers plus other components and pipelines. Choosing an API tier does not specify the hardware or local workflow you would need to reproduce its service.

These official rates, checked October 6, 2026, are in US dollars per output second for text-to-video and image-to-video. Portrait and landscape use the same rate at each resolution. The API bill depends on the selected version, tier and generated duration.

At 1080p, an eight-second text- or image-to-video attempt costs $1.04 on 2.5 Fast and $1.36 on 2.5 Pro. The same duration and resolution costs $0.48 on 2.3 Fast or $0.64 on 2.3 Pro. At 4K, the 2.5 examples rise to $2.40 and $3.12.

For a 30-shot project with eight seconds per shot, one 1080p attempt per shot totals $31.20 on 2.5 Fast or $40.80 on 2.5 Pro. If you generate three attempts for every chosen shot, that becomes $93.60 or $122.40. These examples cover generation only; editing or other operations have their own billing.

For that same three-attempt project, 2.3 Fast would cost $43.20 and 2.3 Pro $57.60. Trial representative scenes on both versions. Count the takes you can actually use rather than assuming the latest model will always justify its higher rate.

For text and image generation, choose the duration from the supported values below. API resolutions are 1280×720, 1920×1080, 2560×1440 and 3840×2160, or the corresponding portrait dimensions. Audio-to-video follows the input audio length instead.

LTX 2.5 audio-to-video uses the same Fast and Pro rates shown above, billed against input-audio seconds. For 2.3 Pro, audio-to-video costs $0.06, $0.10, $0.18 or $0.34 per input second at 720p, 1080p, 1440p or 4K respectively.

Hosted Retake and Extend use 2.3 Pro, not 2.5. The listed 1080p rate is $0.10 per second: Retake bills input-video length, while Extend includes generated extension and context frames, capped at 505 billed frames. A short extension can therefore bill more seconds than the new footage alone.

Create an API key in the LTX Developer Console and keep it in a server-side environment variable. Start with a fixed duration so the charge is predictable. This example requests eight seconds at 720p on 2.5 Fast, a $0.72 generation at the rates above.

The V2 API returns a job ID. Check the matching job URL at least five seconds apart until it completes, then download result.video_url before the link expires. For image-to-video, use the image endpoint and supply image_uri from a supported upload or reachable image URL.

On 2.5, sending duration: null lets the model choose the clip length from the prompt. Prepaid accounts temporarily reserve enough credits for the longest permitted clip at that resolution and frame rate, then release the unused balance. Use a fixed duration for a strict ad slot or tightly controlled test budget.

For example, a 720p, 24 fps Fast request can hold $1.80 for the 20-second maximum, even if the result is shorter. Automatic duration cannot be combined with a last-frame image condition.

The official Python quick start uses a 22B distilled transformer, the LTX-specific Gemma 4 12B text encoder, video and audio VAEs, and a spatial upscaler for the fixed-frame example below. Download size is separate from the VRAM needed while the pipeline runs.

Install the repository dependencies, obtain access to the LTX 2.5 weights on Hugging Face, and download the files named in the current quick start. Keep their directory structure: the command below assumes they are under models/ltx-2.5. Use the bundled, fine-tuned text encoder rather than substituting stock Gemma 4.

This fixed-duration example follows the official repository’s quick start: 121 frames at the default 24 fps, or about five seconds. It does not use a duration head; automatic duration is a separate setup. Local frame counts follow the pipeline’s rules rather than the hosted API’s duration menu.

Lightricks lists a starting configuration for LTX 2.5 of an NVIDIA GPU with at least 32GB VRAM, 32GB system RAM, 100GB free storage, CUDA 12.7 or later and Python 3.12 or later. Its recommended configuration is an A100 80GB or H100 with at least 64GB RAM and 200GB SSD space. Use these as the official baseline, then size the particular pipeline and output you will run.

Begin with DistilledPipeline, then measure generation and decoding at your intended size. The official memory options include FP8 casting and CPU or disk offloading. With BF16 weights, --quantization fp8-cast reduces the transformer’s memory footprint; --offload cpu or --offload disk changes where its weights wait between computations.

Rendering choices also matter. The diffusion video decoder spends more time and memory on detail; a lighter convolutional decoder is available. DFR (Diffusion Fidelity Rendering) adds generated keyframes and a detailing IC-LoRA to the distilled workflow. Its high-resolution decode can be the memory bottleneck, so a successful denoising pass is only part of a full test.

A named community report shows why decoding needs its own check. On August 12, 2026, Mike128 reported a ten-second image-to-video job stalling during decode on an RTX 3090 with 128GB system RAM. They confirmed that reducing the VAE Decode (Tiled) tile size from 768 to 512 fixed that run. LTX staff explain in the same discussion that smaller tiles trade decoding speed for lower VRAM pressure. If sampling completes but the video stalls at decode, inspect that stage and its tiling before changing the entire model stack.

The trainer’s 32GB low-VRAM configuration and 80GB recommendation describe training. For inference, size the complete workflow separately. For a purchase or rental decision, run the complete inference workflow with your chosen quantization, frame count, decoder and offload settings, then record peak GPU memory, system RAM and elapsed time.

Lightricks maintains ComfyUI-LTXVideo for visual workflows. Use a 2.5 example graph with its matching weights and adapters. This is useful when you want to inspect image inputs, conditioning, generation and upscaling as separate stages.

Start with ordinary text- or image-to-video before adding controls. Keep an accepted low-resolution take as the reference for an upscale, and inspect small product details afterward: a generative upscaler can synthesize new texture and structure.

Local workflows and hosted endpoints have separate capability lists. For example, the local repository includes retake and video-to-video pipelines, while the hosted 2.5 API supports text, image and audio generation. Choose the route that implements the operation you need.

Alpha Gen (Beta) is a separate IC-LoRA 0.9 adapter for the distilled LTX 2.5 base. It takes RGB footage and produces a frame-aligned grayscale alpha matte for compositing, including soft edges such as hair or smoke. Combine the matte with the original footage to replace its background. This is a local video-to-video workflow, separate from the Fast/Pro generation endpoints priced above.

Use the official general single-stage video-to-video IC-LoRA ComfyUI workflow, select ltx-2.5-22b-ic-lora-alpha-gen-0.9.safetensors at strength 1.0, and leave the prompt empty. The model chooses the foreground; it cannot select a particular person or object from a text prompt or input mask.

For predictable results, use 121 frames or fewer. Frame counts must follow 8n+1; the supported ceiling is 145 frames and 1920×1088. Dimensions must be divisible by 32, so pad 1920×1080 footage rather than stretching it. For full-HD, 145-frame jobs, the model card calls for H100/B200-class hardware; reduce resolution or clip length on smaller GPUs. Split longer shots and inspect matte edges and cuts before compositing.

Choose a 2.5 trial when you need multi-shot generation or want to evaluate its newer decoder and distilled model. Keep 2.3 in the comparison if single-shot output already meets your brief: its text/image API rates are lower at every listed resolution, and 2.3 Pro supplies hosted Retake and Extend.

For local migration, 2.3 checkpoints bundle components into a single file and use a separate Gemma 3 encoder. Version 2.5 uses the split component layout and its LTX-tuned Gemma 4 encoder. Keep those base-model components paired. Lightricks says most 2.3 LoRAs and IC-LoRAs run on 2.5, with some exceptions: check each adapter’s supported base, loader and workflow, then test it with fixed settings before production use.

For an existing API integration, use the explicit 2.3 or 2.5 model IDs with V2 endpoints. Original ltx-2-fast and ltx-2-pro IDs have been removed. Lightricks has scheduled retirement of the legacy V1 generation endpoints for October 26, 2026 at 23:59 UTC; its upload endpoint remains on V1.

LTX weights use the LTX Community License. The repository maps 2.5 releases from August 11, 2026 to LICENSE-2_x; original LTX-2 and the specified earlier 2.3 releases use LICENSE-2. Select the agreement distributed with your checkpoint.

Both set a paid-license threshold at annual entity revenue of at least US$10 million, counting affiliated and commonly controlled companies. The threshold concerns the entity’s revenue, not just sales from LTX. Below it, the community grant remains subject to the agreement’s use and redistribution conditions.

For entities at or above the threshold, the newer 2.x agreement explicitly exempts qualifying non-commercial testing, evaluation and R&D in non-production environments, without direct or indirect payment. Revenue-generating work, end-user-facing or end-user-impacting use, and training or adapting models for commercial use fall outside that exception. The older agreement has no equivalent explicit exemption.

If you distribute modified weights or adapters, preserve the required notices, agreement and usage conditions. Include the appropriate model-license review in a self-hosting budget before comparing GPU cost with hosted API spending.

Measure the full batch on the hardware you would actually use. Divide its billed rental time, storage and other costs by the number of accepted clips. Include model loading, failed attempts, decoding and idle time between jobs. On owned hardware, include power and the share of the purchase cost attributable to the workload.

Compare that result with the API price for the same clip duration and resolution, using your observed acceptance rate on each route. A local pipeline can be attractive for repeated work and custom controls; occasional jobs also have to absorb the initial setup and large model downloads.

For another local audio-video option, Kandinsky 6.0 provides five-second Lite and Pro models with MIT licensing and published timings on several NVIDIA GPUs. Its shorter clip length and different model controls make a representative shot test more useful than comparing parameter counts.

Is LTX 2.5 free? The weights are downloadable under a conditional Community License. Running them consumes compute, and some organizations require a paid model license. The hosted API charges for video generation.

Can I turn off generated audio? Yes. Set generate_audio to false in a text- or image-to-video request when you want silent output.

Can I request 20 seconds on Pro? Pro supports 6, 8 or 10 seconds for text/image generation. For a 20-second request, use Fast at 720p or 1080p and 24 or 25 fps.

Can I copy the API’s 4K dimensions into local DFR? Local DFR requires dimensions on its 64-pixel grid. Its documented 4K configuration is 3840×2176, whereas the API accepts 3840×2160. Follow the dimensions for the route you are running.

Should I start by buying a GPU? A small API trial can establish whether the model handles your actual scenes. If you then need local controls or sustained throughput, benchmark a matching rental configuration before making a hardware purchase.