The part of the ecosystem with the worst licences
Image generation has the largest open community and the most confusing terms. A single release can ship three variants with three completely different permissions, and the names do not tell you which is which. This lesson is mostly about not shipping something you were not allowed to ship.
The families
Stable Diffusion. SD 1.5 and SDXL are the models the community built on for years; the ecosystem of LoRAs, ControlNets and workflows around them is enormous. They ship under CreativeML OpenRAIL licences, which permit commercial use but attach use restrictions. Later Stability releases moved to a community licence with a revenue threshold above which a paid licence is required.
FLUX (Black Forest Labs). This is the trap worth memorising, because it catches people every week:
- FLUX.1 [schnell]: Apache 2.0. Genuinely open, commercially usable, fast, four-step generation.
- FLUX.1 [dev]: non-commercial licence. It is the one everybody uses and posts about, and it is the one you may not use commercially without a separate agreement.
- FLUX.1 [pro]: not released at all; API only.
Three variants, one family name, three different answers to may I ship this. The output images from the non-commercial variant are also covered by the terms, so generating with it and selling the picture is not a workaround.
Qwen-Image and several newer releases ship under Apache 2.0 with permissive terms, which is why they have become the default for anyone who has been bitten once.
Video. Wan, Hunyuan Video, LTX-Video and Mochi are the open video models. Licences vary from Apache to bespoke. Video is heavy: a few seconds at moderate resolution takes minutes on a good GPU and is not a laptop activity, though smaller distilled variants are appearing.
The free stack
Everything here runs on free software.
- ComfyUI is a node graph editor and the tool the community standardises on. It runs every model above, exposes each step, and its workflows are shareable JSON files. Steep at first and then the only thing you want.
- Automatic1111 and Forge are the older form-based interfaces, easier for a first hour.
- InvokeAI sits between the two and has the better canvas for editing.
- Krita with the AI diffusion plugin puts generation inside a real painting application, which for retouching and inpainting is a better fit than any of the above.
Memory is the constraint. Roughly: SD 1.5 runs in 4 GB of VRAM, SDXL wants 8, FLUX at reduced precision wants 12 to 16, and video wants 24 and up. Quantised and distilled variants shift these down at a quality cost, and CPU generation works but is measured in minutes per image rather than seconds.
If you have no GPU: Google Colab’s free tier, Kaggle’s weekly GPU allowance, and Hugging Face Spaces running someone else’s public demo will all generate images at no cost. For a small number of images the sensible answer is often to use a Space rather than to set anything up.
LoRAs, and what they are here
The adapter idea appears in image models too, and it is where most community work happens. A style LoRA is a few tens of megabytes trained on twenty to a hundred images, and it teaches a look, a character or an object. The licence of a LoRA is separate from the licence of the base model, and a LoRA trained on a living artist’s work to reproduce that artist’s style is the practice at the centre of the strongest objections to this technology. That is a genuine dispute about consent and livelihoods, not a settled question with an obvious answer, and it is worth being clear with yourself about which side of it a given piece of work sits on.
Provenance and the honest limits
Two things belong in any production use.
Label synthetic media. C2PA content credentials embed signed provenance in the file. Metadata is strippable and this is not a security control, but it is the emerging norm and several jurisdictions are moving to require disclosure of synthetic content.
Do not generate people who did not consent. Photorealistic images of identifiable real people, and sexual imagery of anyone without consent, are the two uses that have caused the most concrete harm from this technology. Every open model can produce them and no licence has ever prevented it. The constraint has to be yours.
Every lesson here is open. Your progress is recorded once you pass the Maths You Actually Need exam.
Check your understanding
A studio generates marketing images with FLUX.1 [dev] locally, arguing that since the weights were downloaded freely and the images are their own creation, commercial use is fine. What is wrong with the reasoning?
Pick the one you would defend. Your answer stays private.
Stuck on something?Ask Anjali, your AI tutor