Lightricks Releases LTX-2.5 Open Weights for Custom LoRA Training, Free Under $10M in Revenue

Lightricks released open weights for LTX-2.5, its video and audio generation model, so studios can train LoRAs on their own characters, settings and visual styles. Productions can fine-tune and self-host the model on their own hardware, which Lightricks pitches as a way to avoid per-generation billing and API dependency.

• Weights are on Hugging Face, training code is on GitHub. The LTX-2.5 Hugging Face repo hosts the gated weights, which Lightricks describes as built for local execution and fine-tuning. The LTX Trainer code sits in the official GitHub repository.
• There is an official character LoRA recipe. LTX's character LoRA guide walks through building the dataset, choosing text-to-video or image-to-video mode, deciding on voice, validating, and picking the right checkpoint.
• Multishot generation holds identity across cuts. According to Lightricks' Hugging Face model card, LTX-2.5 generates multiple connected shots in a single pass while holding character identity, environment, lighting, voice and visual style. Previous versions produced a single continuous shot.
• Commercial use is free below $10M in revenue. Companies under that threshold can use the model in production at no cost under the LTX-2.x Community License, though transferring fine-tunes may require a paid license.
• Prebuilt LoRAs target finishing work. The LTX open-source library includes workflows for 4K and 8K output, alpha mattes without a green screen, and SDR-to-HDR conversion.

Custom LoRAs train through the LTX Trainer on a 22B model you can run locally

The official GitHub repository packages Python inference alongside the LTX Trainer for LoRA training. LTX-2.5 generates synchronized video and audio from text, image and video inputs, which is why LTX's character guide includes a step for deciding on the character's voice.

The quick-start download includes a 22B distilled transformer, a Gemma 4 12B text encoder, separate video and audio VAEs, and a 2x latent spatial upscaler. Together they come to roughly 66 GiB.

For tighter GPU memory, the README points to fp8 quantization plus CPU or disk offloading. LTX's open-source page separately advertises local generation on 12GB of VRAM.

Access is gated. Users accept the model terms on Hugging Face and log in with a read token before the download will start.

Custom LoRA training is reaching other models too; we covered fal adding LoRA training for MiniMax H3, which started with an open-source realism people LoRA.

LTX's own LoRAs cover restoration, alpha mattes, HDR and 8K output

Beyond the base model, Lightricks publishes ready-made LoRAs and workflows for the LTX model family in a community library. The LTX team built these for LTX-2.5:

• Native Resolution. A two-stage Tiled Fusion workflow runs IC-LoRA effects at Full HD, 4K or 8K on a single GPU.
• Layout to Render. Turns a 3D viewport animation or blocky playblast into a finished shot, keeping the camera path and object placement while a first frame and prompt set materials, lighting and palette.
• Alpha Gen (beta). Generates a frame-aligned grayscale alpha matte from ordinary RGB video with no green screen, mask or prompt.
• HDR Native and SDR to HDR. The first brings EXR/HDR plates through the model and returns color-managed EXR and HLG output. The second converts 8-bit SDR footage into 16-bit HDR as ACEScg EXR frames plus a BT.2020/HLG master.
• Restore and Refine Details. Cleans and colorizes damaged archive footage, and rebuilds texture, edges and grain in soft or upscaled material.
• Slow Motion Control. Turns an image and prompt into video with motion speed set from real time to 40x slow motion through a single parameter.

Some community entries in the same library, including a Sci-Fi Cinema style LoRA and a lens-flare remover, list LTX-2.3 compatibility. Check the version tag before building one into an LTX-2.5 pipeline.

Free commercial use stops at $10M in annual revenue

Under the LTX-2.x Community License, entities with less than $10M in annual revenue get commercial and production use at no cost. Revenue counts across the whole entity, including subsidiaries and affiliates under common control.

Above that line, a paid Commercial Use Agreement from developer Lightricks covers full weights, engineering support, LoRAs and flexible deployment options.

The license also states that transferring fine-tunes may require a paid license. Studios planning to share fine-tuned models should check that clause before training.