LTX-2.5
An open-source video generation model released by Lightricks, characterized by relatively low VRAM requirements and fast generation speed. It serves as an active base model in the open-source video generation community and supports extensions such as LoRA for image quality enhancement.
Timeline (last 90 days)
Unsloth v0.1.907-beta supports generating videos with synchronized audio from text or a starting image via LTX-2.5.
LTX-2.5 is an open-weight video model launched by LTX, a company spun off from Lightricks, and has garnered 1.58 million downloads and 5,900 likes.
LTX-2.5 generates a 10-second 720P video from a single image in just 6.8 seconds, replaces the traditional VAE with a diffusion video decoder, supports native multi-shot generation, uses the 2B version of Gemma 4 as the text encoder, and supports 4K HDR and RAW workflows.
LTX-2.5 is free for commercial use by teams with annual revenues below 10 million USD, requires a minimum of 16GB VRAM, and ComfyUI provided Day 0 native integration.
The video enhancement LoRA requires that the faces in the input video must already be properly generated; if the original face is already blurry or smeared, the tool cannot repair it.
Creator CQdesign shared an updated video enhancement LoRA for the LTX-2.5 model on Reddit.
The author states that the enhancement LoRA's results for objects such as vehicles are touch and go, i.e., inconsistent.
The new version of the LTX-2.5 video enhancement LoRA eliminates the visual artifacts (ghosting) present in previous versions.
When upgrading the enhancement LoRA, you cannot simply replace the model file; you must also download the new workflow, otherwise artifacts will still appear in the video.
The model and workflow have been published on a Hugging Face page for the community to download and use for free.
All Facts (8)
Unsloth v0.1.907-beta supports generating videos with synchronized audio from text or a starting image via LTX-2.5.
50%UnverifiedLTX-2.5 generates a 10-second 720P video from a single image in just 6.8 seconds, replaces the traditional VAE with a diffusion video decoder, supports native multi-shot generation, uses the 2B version of Gemma 4 as the text encoder, and supports 4K HDR and RAW workflows.
50%UnverifiedLTX-2.5 is free for commercial use by teams with annual revenues below 10 million USD, requires a minimum of 16GB VRAM, and ComfyUI provided Day 0 native integration.
50%UnverifiedThe new version of the LTX-2.5 video enhancement LoRA eliminates the visual artifacts (ghosting) present in previous versions.
50%UnverifiedThe author states that the enhancement LoRA's results for objects such as vehicles are touch and go, i.e., inconsistent.
50%UnverifiedWhen upgrading the enhancement LoRA, you cannot simply replace the model file; you must also download the new workflow, otherwise artifacts will still appear in the video.
50%UnverifiedThe video enhancement LoRA requires that the faces in the input video must already be properly generated; if the original face is already blurry or smeared, the tool cannot repair it.
50%UnverifiedThe model and workflow have been published on a Hugging Face page for the community to download and use for free.
50%