AI models for image creation

AI models of the moment: what Axiris uses, for what, since when

What the studio uses today, with its date (Krea 2 for creation, Z-Image Turbo for fast local generation, FLUX.2 klein for editing by area, MiniMax H3 for video, DLSS 5 for finishing) and what has been published since.

Last updated: 11 October 2026. The image selection dates from August 2026; video, finishing and the autumn publications were added in October.

What we use today

The answer evolves, and that is intended. We do not keep an engine because it once held a major place in our workflows; we test, we compare, we replace when a solution serves our way of working better.

This page never says "the best". It says what we use, for what, since when, and it has to be able to age: a model very present today may leave it in a few months; its family page will remain in the Models section.

For images: three engines, three roles

They do not perform the same function. That is precisely why we use them together: we are not looking for a single model able to do everything, but for a pipeline where each engine steps in where it is most relevant.

Krea 2: creation, aesthetics, exploration

Krea 2 is the family of image models developed by Krea, presented by the publisher as its first foundation model developed in-house, geared towards aesthetics, creative control, styles and the use of visual references, with RAW and Turbo variants.

We use it in the creation phase, when we are looking for the material, the aesthetic direction, the visual richness, the character of an image, leaving the model visual freedom while keeping a precise direction. Krea 2 has been one of Nova's engines since 28 June 2026: Nova grows: Krea 2 and Qwen-Image 2512.

Official sources: Krea 2 · Krea 2 Technical Report.

Z-Image Turbo: fast local generation

Z-Image is a family of models developed by Tongyi-MAI; Z-Image Turbo is its distilled version, designed to produce an image in very few steps. Its official model card, read in August 2026, states a 6-billion-parameter model, an image in 8 inference steps and execution in about 16 GB of graphics memory in its reference configuration. Those are the publisher's figures.

In our local workflows, speed is not a comfort: fast generation lets you test more, compare more, change direction without breaking the rhythm of interactive creative work. That is the role of Z-Image Turbo.

Official sources: Z-Image Turbo · Z-Image.

FLUX.2 klein: editing and retouching by area

FLUX.2 klein belongs to the FLUX family from Black Forest Labs, which positions it as a compact model for generating and editing quickly: iteration, prototyping, fine-tuning, local execution depending on the variant.

At Axiris, its value lies in image editing and interventions by area: you select what has to change, you apply the inference there, the rest does not move. That is the logic of retouching by area and of our way of building the image by areas: the box, the mask, the frozen layer.

Official sources: FLUX.2 klein · FLUX.2 Image Editing.

For video: MiniMax H3

Since 16 August 2026, video has been an axiris.ai service, built on the MiniMax H3 model. A film is composed shot by shot in a timeline: each shot carries its continuity, its intention, its camera move and its start image; it is launched on its own, judged, redone if it does not fit, and the edit is made only once everything is validated. The news item gives the measured times of a nine-shot sequence generated locally: Video: MiniMax H3 and the timeline.

For finishing: DLSS 5

Since September 2026, NVIDIA's DLSS 5 neural rendering has been a finishing pass on our video shots and a tool of the studio. It does not change the resolution of the image: it rebuilds the material (skin, stone, fabric, metal) from the image and the motion. Axiris has published its ComfyUI pack and what it understood by measuring it: DLSS 5 in ComfyUI (10 October 2026).

A pipeline, not a recipe

A project can chain Krea 2 to find a visual direction, Z-Image Turbo to multiply local iterations, FLUX.2 klein to edit and rebuild by area, MiniMax H3 to animate, DLSS 5 to finish, with ComfyUI, Photoshop and the Axiris tools to orchestrate, select, correct. Depending on the project, we use only one of these engines, we change their roles or we pick another one.

The workflow decides the model. The model does not decide the workflow.

What we have published since

What has come out and what we follow, on the date we wrote it: FLUX 3 announced in Early Access (25 July 2026) · NVIDIA SoL-Refiner: 4K video changes method (6 October 2026: installed, first renders compared, measurement under way) · Ideogram 4.5: image editing enters a logic of precision (6 October 2026) · Axiris is preparing its AI router (6 October 2026).

A selection drawn from our workflows

This page is not a general ranking. Other creatives have different constraints and preferences. We evaluate models in our context: graphic designers' workflows, image generation and transformation, ComfyUI, visual references, LoRAs, retouching by area, building by areas, art direction, video, and the integration of Nova into production.

To choose in yours: How to choose an AI image model? and Comparing AI image models.

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.