Updated September 2026 · Services, versions and prices change often: always check the official pages. This is part of our AI and ComfyUI tutorials.
AI-made video is what has surprised us most over the last two years. We've gone from blurry, silent clips a few seconds long to believable videos with natural motion, audio and synchronized dialogue. And for anyone doing video design, like us, the question becomes very concrete: is it better to pay for an online service or to make them on your own computer?
Let's tell you straight away where we stand: for video the choice matters much more than for images. Online you pay for every second generated; locally you pay nothing, but you need a suitable graphics card and a bit of patience. Here we compare Google's Veo, Kling, Seedance and Runway with ComfyUI and open models such as LTX-2.5 and Wan, with pros, cons and real costs.
In brief
| Online | Local (ComfyUI) | |
|---|---|---|
| Quality | The highest available today | Very good with LTX-2.5 and Wan 2.2 |
| Speed | From tens of seconds to a few minutes per clip | A few minutes and up, depending on the graphics card |
| Cost | Paid credits or seconds, can climb fast | Free once you've bought the hardware |
| Hardware | Any computer | 16–24 GB graphics card or more, 32–64 GB RAM |
| Privacy | Your material sits on the service's servers | Everything stays on your computer |
| Control | Prompt, starting image, a few tools | Total: LoRAs, pose and depth controls, workflows |
| Continuity | The service can shut down (see Sora) | The downloaded model stays yours |
The best online video generators
- Google Veo 3.1 and Gemini Omni — Veo 3.1 is among the highest-quality models, with synchronized dialogue and audio. In May 2026 Google added Gemini Omni alongside it, which lets you create and edit videos conversationally. You use them from the Gemini app with Google AI plans and from the Flow tool.
- Kling 3.0 — excellent value for money, clips up to 15 seconds, lip sync in several languages and free daily credits to try it out.
- Seedance 2.0 (ByteDance) — very fast and popular for animating images and for reference-based videos; also available in ComfyUI's paid nodes.
- Runway Gen-4.5 — the richest in control tools (motion brushes, consistency across scenes), designed for editing and production.
- Hailuo (MiniMax) — MiniMax's online service, from the same company behind the open H3 model.
What about Sora? OpenAI shut down the Sora app on April 26, 2026 and the API on September 24, 2026: anyone who had built their work around Sora had to switch tools. We discuss it in the editorial AI of the week.
Pros and cons of online services
- ✅ Top quality and fast turnaround, even from any old laptop.
- ✅ Advanced features ready to go: dialogue, lip sync, clip extension, 4K.
- ❌ Costs that climb fast: videos are paid by the second or in credits, and it takes many attempts to get the right clip.
- ❌ Filters on content and faces, sometimes very restrictive.
- ❌ Dependence on the service: prices, rules and even the product's existence are out of your hands.
Local AI video with ComfyUI
2026 was the breakthrough year for local video. Lightricks' LTX-2 family (up to August's LTX-2.5) generates video with synchronized audio, multiple linked shots and 4K, with a license that's free even for professional use below $10 million in revenue. Wan 2.2 remains the benchmark with a fully open license. Be careful, on the other hand, with MiniMax H3 (and all its evolutions, such as Singularity: we cover them below) and HunyuanVideo: their licenses exclude the European Union. You'll find everything in our handbook on video models in ComfyUI.
MiniMax H3 and its evolutions, up to Singularity
Among open video models, MiniMax H3 is the one that made the most noise in 2026. It was created by MiniMax, the same company behind the Hailuo online service: announced on July 31 with weights published on August 3, it generates video with stereo audio in the same pass, clips from 4 to 15 seconds, vertical too. It comes in two versions: FL2VA (from text, with first and last frame) and Ref2VA (from references: up to 9 images, 3 videos and 3 audio tracks). ComfyUI supports it natively.
The problem, originally, was its size: 33 billion parameters and about 19.5 GB for the model alone in the INT8 version. In the months that followed the community worked on it a lot, and today there's a small ecosystem around H3:
- DMAD (Texas A&M and ByteDance, October 2026): two LoRAs that take H3 from 50 down to 4 steps, with a ComfyUI conversion already available. Project on GitHub.
- FreeVideo: a ComfyUI plugin, with a Windows installer, that runs H3 with 8 GB of video memory and 16 GB of RAM, loading the model in pieces as it works. It's a preview project that changes every day. FreeVideo on GitHub.
- Veda: a ComfyUI node that speeds up H3 with a "sparse attention" technique, meaning it only computes the parts of the image that really matter. Veda on GitHub.
- ControlNet Union 2.0 (Alibaba PAI): eight controls in a single file (pose, depth, edges, scribbles and more), plus video-to-video and video inpainting.
- Community LoRAs: for example Character Swap, which replaces a person in a video with a reference character while keeping the scene and camera moves.
MiniMax H3 Singularity is the most ambitious of these evolutions. It's not an official MiniMax model: it's a "merge" created by the AIGC-Singularity community, published on Hugging Face on September 5, 2026 and updated several times. It combines several H3 variants into a single file, then refined with three more days of training. The stated goals are concrete: less blur in fast motion, sharper faces in medium and wide shots, less "shiny" skin, more energetic movement in fight scenes, better effects and particles, and more precise response to camera moves. The authors say prompts and workflows made for H3 keep working.
In ComfyUI you load it in place of the original H3 model, and it works from text, image, references and video. There are several versions: v1.3 "pruned" in INT8 (about 21 GB, the recommended one), an even more compressed w4a8 version (about 12 GB) and the new v2 in INT8 (about 34 GB). To work faster, the authors recommend pairing it with a 4-step "turbo" LoRA. You'll find files, workflows and examples on the Singularity page on Hugging Face.
The honest note, which for us is decisive: the Singularity page lists the license as Apache 2.0, but it's built on H3's weights, and those weights fall under the MiniMax H3 Community License. This explicitly excludes the European Union, the United Kingdom, South Korea and the United States. In plain terms: from Italy, H3 and all its evolutions (Singularity included) can't be used locally without a license requested directly from MiniMax. We follow it with interest because it's technically remarkable, but for working from Europe the open choice remains the LTX-2 family; online, on the other hand, you can use MiniMax's Hailuo service, which has its own terms of use. The license text is on MiniMax's official page.
Pros and cons of working locally
- ✅ Unlimited attempts: video takes lots of tries, and locally they cost nothing.
- ✅ Director-level control: first and last frame, poses, depth, LoRAs for characters and styles, videos made to talk over an audio track.
- ✅ Privacy for client projects and confidential material.
- ✅ Integration with the rest of your work: images, upscaling, interpolation and editing in the same workflow.
- ❌ Demanding hardware: video models are the heaviest, ideally 24 GB of video memory and 64 GB of RAM.
- ❌ Longer render times than the big online services, especially at high resolution.
- ❌ Steeper learning curve than a simple prompt.
The hybrid workflow: the best of both worlds
Many creators take a middle road:
- Ideas and tests locally: shots, style and characters with ComfyUI, at no cost.
- Key images locally with an image model, to animate later with image-to-video.
- Online only for the final clips that need top quality, paying for a few seconds instead of hundreds of attempts.
- Finishing locally: upscaling, interpolation and editing.
ComfyUI also lets you use some online models (Seedance, Veo and others) inside your workflows through paid "partner nodes", so everything stays in a single program.
Which route to choose
| If you want… | We recommend |
|---|---|
| To try it without spending | Online with free credits (Kling, for example), or locally if you already have a good graphics card |
| A few high-quality clips a month | Online |
| Shorts, Reels and frequent videos | Locally with LTX-2.5 or Wan 2.2, possibly with a few online clips |
| Music videos, commercials, client projects | Hybrid workflow |
| To really learn AI video | Locally with ComfyUI |
For the math on hardware versus subscriptions, read how much it costs to create images and video with AI.
The graphics cards we recommend for AI video
| Prodotto | Scheda video | Memoria video | Link |
|---|---|---|---|
| ASUS Dual GeForce RTX 5060 Ti OC 16 GB | NVIDIA RTX 5060 Ti | 16 GB | Vedi su Amazon |
| ASUS Prime GeForce RTX 5070 Ti OC | NVIDIA RTX 5070 Ti | 16 GB | Vedi su Amazon |
| MSI GeForce RTX 5090 32G Gaming Trio OC | NVIDIA RTX 5090 | 32 GB | Vedi su Amazon |
| PC Ryzen 7 5700X + RTX 5060 Ti 16 GB | NVIDIA RTX 5060 Ti 16 GB | 16 GB | Vedi su Amazon |
Frequently asked questions
Can you make AI video with audio locally?
Yes: LTX-2, LTX-2.3 and LTX-2.5 generate synchronized video and audio in the same pass, directly in ComfyUI, with a license that can also be used in Italy.
What's the best online AI video generator?
For pure quality, Google's Veo 3.1; for value for money, Kling 3.0; for speed, Seedance 2.0; for control, Runway. It's worth trying two or three with the same idea.
Is a 12 GB card enough for video?
For learning and making short, low-resolution clips, yes, with the reduced versions of the models. To work consistently you need at least 16 GB, ideally 24 GB or more.
