The AI Video Explosion: 7 New Models You Need to Know (January 2026 Edition)

Okay, so this week has been absolutely wild in the AI video world. I'm talking blink-and-you-miss-it levels of announcements. We've moved way past those basic "type some text, get a janky video" tools. Now we're looking at systems that handle sound, 3D space, and editing in ways that honestly seemed impossible just months ago.

I've been testing these out, and here are the seven models that actually matter right now—what they do, and more importantly, whether you can actually run them on your setup.

1. UniVideo: The Do-Everything Model

Developer: Kling Team (Kuaishou Technology)

So UniVideo is basically what happens when someone asks "why do we need 10 different models?" It's a generalist framework that handles understanding, generation, and editing all in one package.

The tech is pretty clever—it runs two systems simultaneously: Qwen2.5-VL-7B handles the "understanding what you want" part, while HunyuanVideo-T2V-13B does the actual visual work.

Why you should care: This thing can do text-to-video, image-to-video, or crazy specific edits like "make this bomb explode" on a still image. You can even do localized changes—like turning an object to glass—without the model needing special training for glass materials. That's pretty wild.

The catch: It's massive. We're talking 95.2 GB here. If you've got enterprise hardware, no problem. For home users with an RTX 3090 or 4090, you can make it work on 24GB VRAM if you dial back the frame length to 12. I've seen people squeeze it onto 12GB cards, but honestly, go make lunch while it renders because you'll be waiting.

Where to get it: Code and weights hit GitHub on January 7. The official repo is your starting point, though ComfyUI support is still being worked on.

2. LTX-2: Video With Actual Sound (Finally!)

Developer: Lightricks

Here's something that's been driving me crazy—every video model gives you silent clips, and then you're stuck trying to add audio separately and hoping it lines up. LTX-2 just... fixes that. Audio and video generated together, perfectly synced.

The game-changer: Lip movements match the speech. Footsteps happen when feet hit the ground. Explosions actually explode. All native. Plus, it outputs at 4K and up to 50 FPS, which makes it one of the sharpest open models out there.

Control options: They added IC-LoRA support, so you get granular control over depth, pose, and edge detection.

Can you run it? Yes! The community's already done the hard work—there are GGUF quantized versions that'll run on 8-12GB VRAM. It's also fully integrated with ComfyUI through the ComfyUI-LTXVideo nodes.

3. DreamID-V: Face Swaps That Don't Look Terrible

Developer: ByteDance / Tsinghua University

Look, video face swapping has been rough. Flickering, weird lighting, masks that don't track right—it's been a mess. DreamID-V actually solves this by using proper 3D reconstruction instead of just slapping a 2D face onto bodies.

What's better: Extreme head turns? No problem. Hand covering the face? Still tracks. Complex expressions? Nailed it. And it maintains the target identity throughout instead of morphing into some uncanny valley creature halfway through.

The numbers: It's currently dominating the IDBench-V leaderboard, beating everything else including Face-Adapter.

Getting started: Code dropped January 5, models are on Hugging Face. The Wan-1.3B variant runs fine on 16GB VRAM, though 24GB is ideal for the full model.

4. DreamStyle: When You Want Style to Stay Put

Developer: ByteDance

Ever tried to make a video look like claymation or a specific art style, only to watch it randomly shift between different looks every few seconds? Yeah, that's style drift, and it's incredibly annoying. DreamStyle was built specifically to fix this.

How it works: Smart LoRA implementation that separates style from content. You can control style through text, a reference image, or even paint the first frame yourself and have it animate the rest matching your exact style.

The waiting game: Paper's out and looking promising, but the inference code is still marked "pending" in their repo. Definitely one to keep tabs on.

5. NeoVerse: Building Actual Worlds

Developer: CASIA & CreateAI

Okay, this one's exciting. We're not just making videos anymore—we're making worlds. NeoVerse uses 4D Gaussian Splatting to reconstruct scenes with depth AND time.

The Matrix effect: Take a regular video, reconstruct it into 4D space, then do whatever you want. Freeze time and circle around a water droplet mid-splash. Edit specific objects while leaving everything else untouched. It's basically that bullet-time effect but for any video.

Status check: GitHub's live with documentation, but the actual code isn't up yet. Star it and wait.

6. GaMO: 3D Models from Just a Few Photos

Developer: National Yang Ming Chiao Tung University

Most 3D reconstruction needs tons of overlapping photos. GaMO (Geometry-aware Multi-view Outpainting) flips the script—instead of making up new views from nothing, it "outpaints" from your existing angles to fill gaps.

Speed demon: About 25x faster than older diffusion methods. We're talking under 10 minutes. And it works with really sparse inputs—just 3 to 9 images can build a full coherent scene.

Availability: Research paper is up (arXiv:2512.25073), but we're still waiting on the public code release.

7. InfiniDepth: Depth Maps at Any Resolution

Developer: ZJU3DV (Zhejiang University)

Depth estimation usually hits a wall at your image resolution. InfiniDepth throws out the pixel grid entirely and uses Neural Implicit Fields instead.

Why that matters: You can query depth at literally any resolution. Got a 1080p source but need an 8K or 16K depth map? Done. It catches details other systems miss—individual hair strands, thin wires, stuff that usually just vanishes in standard depth maps.

Ready to try? Code's on GitHub (zju3dv/InfiniDepth) and ready for local testing right now.

So yeah, that's where we are in January 2026. Every single one of these would've been science fiction a year ago, and now they're just... here. The pace is honestly kind of insane, but if you're building anything in this space, these are the tools worth paying attention to.

TextToVideo

Related Topics

The AI Video Explosion: 7 New Models You Need to Know (January 2026 Edition)