LTX-2.5 — Lightricks Just Open-Weighted the AI Video Model You Can Actually Run: Native Multishot, 4K HDR, and the Same Character Across the Cut

Hook

The same week the AI world lost its mind over MiniMax H3, a quieter release did the one thing H3 refused to: it let you run it. Lightricks dropped LTX-2.5 — a 22-billion-parameter, open-weights audio-video model that fits on a 16 GB GPU, ships no watermark, and generates a 10-second clip faster than the clip is long. And it finally cracked the hardest problem in AI video: keeping the same character alive across a cut.

The Story

AI video has been stuck on a wall for two years. You can generate a gorgeous 10-second shot. Ask for a second shot of the same person — a close-up after a wide — and the model hands you a stranger. Different face, different jacket, different room. Every “AI film” you have seen is really a pile of single shots, glued by luck and a lot of re-rolls.

LTX-2.5 attacks that wall head-on with native multishot. One generation produces several connected shots — wide, medium, close-up — that hold the character, the environment, the lighting, and even the voice across the cuts. This isn’t a stitching trick bolted on afterward. Under the hood it’s a 22B asymmetric dual-stream diffusion transformer with a Gemma 4 12B text encoder, generating picture and sound together through bidirectional cross-attention. Sound effects, music and dialogue are modeled in the same forward pass as the pixels, not layered on later.

Two more features earn their keep. Auto duration lets the model decide how long a shot should run based on the action you describe — no more guessing frame counts. And Diffusion Fidelity Rendering (DFR) is a new pass aimed squarely at real production: it renders against real footage, keeps results editable, and can output cinema-grade EXR. That last detail matters more than it sounds. EXR is the language of compositors. A generative model that speaks it isn’t a toy — it’s a plugin for the pipeline you already have.

Why You Should Care

Here’s the part that makes this a Lab story and not just a spec sheet. LTX-2.5 is genuinely open. Weights are on Hugging Face. It’s free for any organization under $10M ARR. It runs locally on any GPU with 16 GB+ VRAM. No per-clip fees. No watermark. No cloud gatekeeper reading your prompts.

Contrast that with the week’s louder headline. MiniMax H3 also open-weighted — but its Community License explicitly excludes the United States, the EU, the UK and South Korea from local deployment. In other words, most of this blog’s readers can admire H3 and cannot legally self-host it. LTX-2.5 has no such wall. For a working creative technologist, “the model you can actually run, on your own box, on your own footage” beats “the model with the flashier demo reel” every single time.

And it’s fast. Lightricks clocks a 10-second clip in roughly 6-8 seconds with the distilled checkpoint on NVIDIA superchips — faster than real time. Even off the halo hardware, the int8 and NVFP4 quantizations exist precisely so mortals with a single consumer card can join in. Native 4K HDR output rounds it out, with HDR pipelines baked into the official workflows.

Try It / Follow Them

The good news for makers: ComfyUI support landed on day one. Three official templates ship out of the box — text-to-video, image-to-video, and a first-frame/last-frame blend (FLF2V) — plus text-to-audio and audio-to-video modes.

• Weights: huggingface.co/Lightricks/LTX-2.5 — base and distilled, plus separate video/audio VAEs and 2× upscalers.
• ComfyUI guide: docs.comfy.org — LTX-2.5 workflows.
• Release notes: ComfyUI Wiki write-up.
• Model page: ltx.io/model.

Grab the distilled int8 transformer, the Gemma 4 text encoder, and the two VAEs, load a template, and start with image-to-video on a still you already like. Then try the thing that matters: describe a second shot of the same subject and watch whether the face survives.

IK3D Lab Take

We’ve covered a parade of video models this year — Kling, Seedance, HappyHorse, the lot. Most were bigger, prettier, or louder. LTX-2.5 is the one we’d actually put on a workstation. The multishot consistency is the real headline: it’s the difference between “AI clip generator” and “AI shot list.” The EXR/DFR angle is the quiet tell that Lightricks is aiming at pipelines, not TikTok.

Be honest about the caveats: “open” here means open weights, not open training data, and the fastest numbers come from datacenter silicon you don’t own. Multishot will still drift on hard prompts — this is a wall being pushed, not demolished. But for a tool you can download today, run offline, and fold into ComfyUI next to your splats and your 3D generators, LTX-2.5 is the most exciting video drop of the month. The Lab’s verdict: clear a weekend, pull the weights, and see how far you can carry one character.

Sharing is caring!