CyberFreezeDev- Home
- Blog
- R&D / Labs notes
- Self-hosted image generation for business: models and licences
Self-hosted image generation for business: models and licences
Open image models can run on a single GPU in your office, but several popular ones are not free for business use. Here is what the licences say, what hardware you need and when a hosted tool is simpler.
Self-hosted image generation means running an open image model, such as FLUX or Stable Diffusion, on your own GPU through a tool like ComfyUI or Fooocus, so prompts and reference images never leave your machine. It suits private product shots, steady volume and full control over the output. The catch is licensing: several popular open models are not free for business use, so read the model card before you build anything on one.
Which open image models can a business use?
All of these are diffusion models or close relatives. Here is what each model card said on 10 November 2026:
| Model | Size | Licence | Business use |
|---|---|---|---|
| FLUX.1 [schnell] | 12B | Apache 2.0 | Yes |
| FLUX.2 [klein] 4B | 4B | Apache 2.0 | Yes |
| FLUX.1 [dev], FLUX.2 [dev] | 12B, 32B | FLUX [dev] Non-Commercial License | Needs a paid commercial licence from Black Forest Labs |
| Stable Diffusion 3.5 Large | 8B | Stability AI Community License | Free under US$1M annual revenue, enterprise licence above |
| SDXL | CreativeML Open RAIL++-M | Yes, with use restrictions |
The [dev] licence trips people up. It says you may use the images for any purpose, including commercial ones, except training a competing model. But the same licence says use for revenue-generating activity is not a non-commercial purpose, and bars commercial or production use of the model itself. If a business runs it for paid work, plan on buying the commercial licence. The 9B version of FLUX.2 [klein] is also under a different licence from the 4B one.
SDXL's licence allows commercial use and even hosting as a service, but lists banned uses your staff must follow, such as giving medical advice or making automated decisions that affect people's legal rights.
Which tools run them?
ComfyUI is a node graph interface: you wire up loaders, prompts, samplers and upscalers as boxes on a canvas, save the graph, and rerun it. It supports SD 1.5, SDXL, SD 3.5, Flux.1, Flux.2 and many newer models, runs on Windows, Linux and macOS, and downloads nothing unless you ask. Licence: GPL-3.0. Its --offline flag switches off the optional paid API nodes.
Fooocus is simpler: a prompt box and a few presets, built on SDXL. Its README now says it is in limited long-term support with bug fixes only, and points users who want newer models such as Flux to WebUI Forge, ComfyUI or SwarmUI. Fine for SDXL work today; not where new models will arrive.
For anything you will run every week, ComfyUI's saved graphs are the better fit: the same graph, seed and settings give you a repeatable house style.
How much GPU do you need?
The project pages give a floor, not a comfortable working size:
- Fooocus: at least 4 GB of GPU memory on an NVIDIA RTX card, plus 8 GB of system RAM.
- ComfyUI: large models on as little as 4 GB of VRAM and 8 GB of RAM by streaming weights, which is slower than keeping the model on the card.
- FLUX.2 [klein] 4B: about 13 GB of VRAM, so an RTX 3090 or 4070 class card and above.
- FLUX.2 [dev], 32B: the model card's own example uses a 4-bit version on an RTX 4090 with the text encoder run remotely.
In our lab, open image models run through ComfyUI on a laptop RTX 3070 Ti with 8 GB, and images fit on that card far more easily than video does. Long batches heat a laptop GPU, so our jobs pause when the card passes 80 °C and resume under 72 °C. For a team generating images daily, a desktop card with more memory is the practical start; our GPU servers and AI hardware page explains how we size one.
Self-hosted or a hosted tool?
Hosted tools such as OpenAI's GPT Image models and Midjourney need no hardware and are quicker to start with. Their terms differ in ways that matter:
- OpenAI's terms of use say you own the output and OpenAI assigns its rights to you, but similar output may be given to other users, and you must not present it as human-made.
- Midjourney's terms of service say you own what you create, but a company with more than US$1,000,000 a year in revenue must be on a Pro or Mega plan to own it. Images are publicly viewable and remixable by default unless you use Stealth mode, and Midjourney gets a licence to your prompts and images.
Self-hosting wins when inputs are confidential (unreleased products, client photos), when volume is steady, or when you need a fixed pipeline. A hosted tool wins for occasional use and the newest model quality.
Who owns the images, and do you have to label them?
Ownership terms are one question; copyright is another. The US Copyright Office's January 2025 report concluded that purely AI-generated material is not protected, and that prompts alone do not give enough human control to make you the author. Human selection, arrangement or editing can be protected, case by case.
In India, the IT Rules as amended on 10 February 2026 cover "synthetically generated information": realistic images, audio or video that could pass for a real person or event. The amended rules require platforms that offer generation tools to label such content and embed metadata, and large social media platforms to ask users to declare it before posting. This is general information, not legal advice.
In practice: keep your prompts and edit history, never generate a real person's likeness without written consent, and declare realistic synthetic images when a platform asks.
Getting a pipeline running
Our video generation pipelines lab tests open image and video models through ComfyUI, and our generative video and image studio delivers campaign visuals on retainer, reviewed by a person and flagged where an AI disclosure is needed. If you want help choosing a model, a licence and a GPU for your own pipeline, book a consultation.
