Introducing LTX-2.5, the world model the world builds on. One of the biggest upgrades yet to the model already powering film, robotics, and real-time workflows. Higher pixel fidelity, multishot scenes that hold together across cuts, and a pretrained foundation built to be finetuned across domains. Plus: Diffusion Fidelity Rendering, a new rendering approach built for pixel quality that holds up frame by frame, even on a cinema screen. This isn't a tool you rent. It's a foundation you build on, open and yours.

been training an LTX 2.3 IC LoRA for editing it's not totally there yet, but lowkey good already! training on a synthetic dataset that is a bit outdated so it's actually a nice signal, will imrpove the dataset and re-train 🤞🏻

Really happy to share this one 🙌 I trained this render-to-real IC-LoRA for LTX-2.3 on 3D data I made myself + a few synthetic scenes. It's the first LoRA of its kind, so still a bit prototypal — it shines on a single scene, and for multiple scenes it's best to generate them separately. Biggest tip: nail your first frame. The reference first frame should closely match the opening frame of your video — that's what makes the result clean and coherent. Hope you all get great results with it ✨ Run on fal: fal.ai/models/fal-ai/ltx-2.3… Model on Hugging Face: huggingface.co/fal/LTX-2.3-3…

My read on LTX-2 is different from the technical coverage I’m seeing. This is a $1.8B company that just gave away their crown jewels. And it might be the smartest thing theyve done. Here’s the math on why open-sourcing makes strategic sense. Lightricks has 6.6 million paying subscribers generating roughly $80M in annual revenue from apps like Facetune. Their core business model is consumer subscriptions on mobile. But the AI video market is consolidating FAST. Sora, Kling, Veo, Runway… all closed systems with proprietary weights. The big players are racing to lock in developers and enterprise customers through API contracts. Lightricks can’t win that game at their scale. So they flipped the table. By open-sourcing LTX-2 with full training code, they’re making a bet on ecosystem dynamics. If ComfyUI workflows and indie developers standardize on LTX-2, Lightricks becomes the de facto foundation layer for audiovisual generation. The real product becomes LTX Studio. Which charges for the UI, collaboration features, and enterprise support. It’s the Red Hat playbook. Give away the engine. Sell the gas station. What makes this interesting is the timing. They dropped this at CES 2026 alongside NVIDIA optimizations that cut model size by 30% and double inference speed on consumer RTX cards. That NVFP8 quantization means a 19B parameter model runs on hardware people already own. Compare that to Sora, which still has limited access and undisclosed pricing. Or Kling, which requires cloud API calls. The constraints are flipping. Access used to be the bottleneck. Now compute is. And Lightricks just positioned themselves on the right side of that shift. The failure mode here is if the community fragments. If developers choose other open models or the proprietary players drop prices agressively enough to make local compute irrelevant. But my guess is that’s a 2-3 year window minimum. And in that window, Lightricks has a shot at becoming the default standard. The 18x faster inference stat buried in the paper matters more than the 4K resolution. Speed determines iteration cycles. Iteration cycles determine adoption. This is a product strategy disguised as a research paper.

Fast matters. LTX-2 is the fastest open-source model in its class, up to 18x faster than competitors.

A fun Tear FX LoRA I trained on the recently open sourced LTX-2 by @LTXStudio If you want to tear some object, grab the LoRA here : huggingface.co/oumoumad/LTX-… LTX-2 provides unprecedent control and flexibility comparing to anything else in the space of video models today

LTX-2 is natively supported in ComfyUI on Day 0 🎬🔊 The next chapter in controllability for open-source video generation. - Open-source audio-video foundation model - Generates motion, dialogue, SFX, and music together - Canny, Depth & Pose video-to-video control - Keyframe-driven generation - Native upscaling and prompt enhancement Everything in a single pass.

🚀 Long shot generation in LTXV-13B Now supports up to 60 seconds of video! • Auto-regressive generation (up to 60s) • Standard generation (up to 20s) • Streamable on H100 with low latency • Compatible with control LoRAs (released last week) • Time-varying prompts supported

🚀 A new way to control AI video generation - by concatenation of control signals and LoRAs trained on just a few samples We’re releasing 3 control LoRAs for LTX-Video: open-pose, depth, and canny edges. Plus: training code so you can build your own types of control. 🧵

🔥 Following the amazing feedback we got for LTX-Video-13B, we’re now releasing: ⚡ 13B v0.9.7-distilled – HD videos in 10s instead of 50s – Low-res preview in 3s – Supports LoRA trained on the base model ⚡ 13B v0.9.7-distilled-fp8 (quantized) ⚡ 13B v0.9.7-distilled-lora

🚀 We just released LTX-Video 13B – our most powerful open-source video generation model yet. – 13B parameters – Multiscale rendering for better fine details – Better motion and scene understanding – Still fast - runs on local GPUs – Supports keyframes, camera/character motion & multi-shot sequences