LTX Video 2.5 Fast

• 🏢 Lightricks' open-weights LTX video family, fast image-to-video tier
• 🧠 LTX-2.5 built as a 22B dual-stream diffusion transformer
• ⚡ Distilled checkpoint runs in a few steps, not twenty
• 🎯 Animates a still image into video with synchronized audio
• 🆕 Replaces LTX-2.3 Fast; LoRAs need retraining for new latents
• 🔧 Available via Diffusers and the official inference package
• 📚 Open weights, permissive terms below a revenue threshold
• 🌐 Pairs with a Pro tier for finished, production-quality renders

Lightricks

Lightricks is a creative technology company founded in January 2013, widely recognized for its popular selfie-editing app Facetune. Originally focused on mobile video and image editing tools, the company has expanded significantly into generative AI,…

LTX Video 2.5 Fast is the speed-optimized image-to-video tier of Lightricks' LTX-2.5 generation, released in 2026. You supply a still image plus a prompt, and the model animates it, producing synchronized audio alongside the picture rather than dubbing it afterwards. It sits alongside LTX Video 2.5 Fast for prompt-only generation and LTX Video 2.5 Pro for higher-fidelity finals.

Architecturally, LTX-2.5 is described as a 22-billion-parameter asymmetric dual-stream diffusion transformer paired with a Gemma-based text encoder and separate video and audio autoencoders. The "Fast" designation corresponds to the distilled checkpoint, which Lightricks describes as built for iteration and running in a handful of denoising steps where the full model needs roughly twenty.

Compared with its predecessor LTX Video 2.3 Fast, Lightricks positions 2.5 as the recommended starting point for new projects, particularly where workflows depend on multi-shot continuity, editing existing footage, or fine-tuning to a specific domain. The 2.5 line adds capabilities such as retake for regenerating a single stretch of a clip, dialogue replacement, keyframe interpolation, and in-context LoRA adapters for pose, depth, inpainting, relighting and upscaling. One migration caveat: the latent space changed, so custom LoRAs trained on 2.3 must be retrained.

The weights are published openly with training and inference code, under terms permitting commercial use below a stated revenue threshold.