AI models for image creation

Which AI model to use for which job (image, video, finishing), on what hardware, locally or through an API, and under which licence. The families the studio follows, what it uses today, and how it tests.

Visual artificial intelligence models are the engines that generate, transform or finish an image or a video. This section is written for the person who has to choose one: a graphic designer, an art director, a photographer, or a company bringing AI into its production.

At Axiris, we are not looking for "the best model". We are looking for which model serves which job, under which constraints, on which hardware and in which workflow. The model suited to a photorealistic advertising image is not the one that modifies an existing image without destroying the rest, nor the one that animates a shot, nor the one that finishes the texture of a render.

The model is a building block of the workflow, not the workflow itself. That is the Axiris thesis: we no longer generate an image, we build it, and what is validated no longer moves. Any engine, whichever it is, joins this method; it does not replace it.

Three families of engines

For images. The families we follow, each with its own page: FLUX (Black Forest Labs), Qwen-Image (Qwen), Gemini Image (Google), GPT Image (OpenAI) and Stable Diffusion (Stability AI). Krea 2 and Z-Image Turbo, which we use in production, are presented in Models of the moment.

For video. Since August 2026, video has been an axiris.ai service, built on the MiniMax H3 model: a film is composed shot by shot, each shot is judged and redone on its own, and the edit is made only from what has been validated. The news item of 16 August 2026 gives the measured times: Video: MiniMax H3 and the timeline.

For finishing. A validated render then goes through engines that do not invent the image but finish it: upscaling, and since September 2026 DLSS 5 neural rendering, which rebuilds the material (skin, stone, fabric) without changing the resolution. Axiris has published its pack and what it understood: DLSS 5 in ComfyUI (10 October 2026).

What is dated, and where it lives

A family page describes what a model is; it ages slowly. What changes quickly lives elsewhere: Models of the moment says what Axiris uses today, for what and since when, with its update date at the top; the news says what comes out, on the day it comes out. When an announcement interests us, we install it, we set it against what is already running, we measure, and the result becomes a guide.

How we test

An Axiris test always reads the same way: same intention, known conditions. The model and its version, the hardware, the settings and the references are stated; the result is dated; what the maker announces is kept apart from what we observed. Whenever possible, the test shows a before / after comparator that you drag yourself. The method is described in Comparing AI image models; the tests themselves are in the guides and the news.

What Axiris does with it

For a creative, choosing the model is part of building their workflow: How to choose an AI image model? is the starting point, and AI workflow for creatives the next step.

For a company, two ways of working with us: on a machine in Switzerland, rented by the hour and delivered ready to use, you choose and install your own models; in the Nova service, Axiris chooses the engine for each task, local, Swiss or cloud, according to the level of confidentiality you have set. The direction is explained in the news item of 6 October 2026: Axiris is preparing its AI router.

Contact request: a first conversation, free and with no commitment. Every practice is different; it is by listening to you that we propose what suits you.