Images and video with AI — a student’s map
What exists, what is free, what runs on your own laptop, whose licence lets you use the result, and what the law asks you to label. Every line links to the page it was read from, checked on 29 October 2026. Where no official page says something — a free tier, a clip length — we leave it out. The field changes monthly.
For the news itself: Ask AI → Read about AI → Images & video.
Sora is gone
Video in the cloud
Google Gemini — Omni Flash
Video from a prompt inside the Gemini app; Google says the Omni model replaces Veo (Veo 3.1 is the previous one). 10-second clips. Needs a Google AI subscription and an age of 18+.
Runway — Gen-4.5
Free plan: a one-time 125 credits. Gen-4.5 costs 60 credits per 5 seconds, so the free credits make about 10 seconds of video. Also offers partner models (Kling, Seedance).
Kling 4.0 (Kuaishou)
One of the most used video models; its own site and inside Runway and Firefly.
Luma — Ray3
Priced per 5- or 10-second clip; the pricing page shows paid plans only.
Pika
Its own Pika 2.5 plus other makers’ models. The free plan has no monthly credits and no commercial licence.
MiniMax Hailuo — H3
Chinese video model, available on its own site and through aggregators.
Images in the cloud
Google Gemini — Nano Banana
Nano Banana 2 is free in the Gemini app; 2.1, the newest, needs a paid plan. Images carry an invisible SynthID watermark and a visible one.
Adobe Firefly
Free daily generations without a plan (the number is not stated). Adds Content Credentials to every file. Adobe calls output from its own Firefly models safe for commercial use; output from partner models (GPT Image, Nano Banana, FLUX, Kling…) is your responsibility.
ChatGPT Images
OpenAI’s image model inside ChatGPT; version 2.0 launched in April 2026.
FLUX — Black Forest Labs
FLUX 3 Image and Video through the maker’s API and many apps; some FLUX models are also open (below).
On your own computer
Open models are free to download, and nothing you make leaves the machine. Read the licence before you publish: “open” ranges from anything-goes to non-commercial to not-in-your-country.
| Model | What | Licence | Note |
|---|---|---|---|
| FLUX.1 [schnell] | Image, 12B | Apache-2.0 — commercial use allowed | Fast few-step model; the classic starting point. |
| FLUX.2 [klein] 4B | Image, 4B | Apache-2.0 — commercial use allowed | The card says it fits in about 13 GB of GPU memory. (The 9B version is non-commercial.) |
| Stable Diffusion 3.5 Medium / Large | Image, 2B / 8B | Stability Community License — free, commercial use under $1M revenue a year | |
| Wan 2.2 TI2V-5B | Video, 5B, 720p | Apache-2.0 — commercial use allowed | The model card asks for a 24 GB GPU; ComfyUI’s guide says it fits 8 GB with offloading. |
| LTX-2 | Video, 19B | LTX-2 Community License — free commercial use under $10M revenue a year | |
| HunyuanVideo 1.5 | Video, 8.3B | Tencent licence — does not cover the EU, the UK or South Korea | If you study or work in the EU, this one is not licensed to you. |
Apps that run them
Draw Things
Free and offline on iPhone, iPad and Mac. Runs FLUX, Wan 2.2, LTX-2 and Qwen Image among others — the easiest way to try open models without code.
mflux
Command-line FLUX on Apple’s MLX, with 8-bit quantisation for smaller memory.
ComfyUI Desktop
The node editor most open-model workflows are shared in. macOS 13+, Apple silicon.
DiffusionBee
One-click Stable Diffusion for the Mac; about 30 seconds an image on an 8 GB M1 Air, by its own figure.
The rules
Label it — EU AI Act, Article 50
Since 2 August 2026, makers of generators must mark synthetic images, audio and video in a machine-readable way, and whoever publishes a deepfake must say so. For plainly artistic, satirical or fictional work the label may be light, but not absent. Systems already on the market have until 2 December 2026 for the marking.
Check it — Content Credentials and SynthID
Content Credentials (the C2PA standard) travel with a file and show how it was made and edited. Google’s SynthID is an invisible watermark; its detector also reads marks from OpenAI, NVIDIA and Kakao.
Own it? — copyright
The US Copyright Office (January 2025): a prompt alone does not make you the author. What you add yourself — your own material, a creative selection and arrangement, real edits — can be protected; using AI as a tool does not rule that out.
Charts are not pictures
Image models draw what looks right, not what the numbers say. A chart for publication comes from the data — the BullDozer app’s Edu tab has a map maker and a people-chart maker that print the source on every image.