MOSS-Image lands in NOVA: Axiris's new favourite
18 octobre 2026 · Axiris
MOSS-Image is an open image generation model published on 16 October 2026 by OpenMOSS. Its family of models rests on a 5-billion-parameter backbone, with Base and SFT versions under the Apache 2.0 licence. At Axiris, MOSS-Image is already installed in NOVA: our first generations in BF16 impress with their skin rendering, their lighting and their understanding of very short prompts.
There are the launches you watch go by. And the ones you install immediately to see what they really have under the bonnet. MOSS-Image belongs to the second category.
MOSS-Image is already up and running in NOVA
Barely released, the new OpenMOSS model has joined our creation environment NOVA, thanks to the work of our development agent LeCode.
From the interface, we can select MOSS-Image as the generation engine, type a prompt, choose the ratio and the resolution, use a random or fixed seed and select a precision such as BF16. The first renders were obtained directly from this integration.
So this is not a simple rehash of the launch visuals: we are running the model inside our own ecosystem.
The first surprise: striking skin rendering
The first portrait was a control test run by LeCode during installation. Nothing like a long optimisation session, and yet the result immediately draws the eye.
The skin has a remarkable presence: fine grain, natural variations in pigmentation, slight irregularities, subtle highlights and a soft transition between light and shadow. The face is alive, with an organic material far less artificial than what one regularly meets in image generation.
For a simple start-up test, it is a very fine opening.
"cyberpunk portrait": two words are enough
Second image, change of universe. The prompt is two words: "cyberpunk portrait".
Result in NOVA: a vertical 896 × 1536 portrait, generated in BF16, with a green and cyan neon mood asserted at once. Hair, gaze, light, technological details and reflective materials compose a strong image, with no endless description.
This second test reveals a quality much appreciated in creative work: the model grasps an artistic direction and translates it into a coherent composition, even when the starting intent is minimal.
What does MOSS-Image offer technically?
OpenMOSS publishes two complementary variants:
- MOSS-Image-Base: the base model, interesting in particular for adaptation, fine-tuning and LoRA creation.
- MOSS-Image-SFT: a version tuned for image generation that can be used directly.
The team also highlights several capabilities: text generation inside the image, prompts of up to 4,096 tokens, and Layout JSON to control the placement of objects and typographic elements. The project also offers a ComfyUI integration.
For a creation studio, the prospects go well beyond the portrait: art direction, advertising compositions, posters, editorial visuals and future custom styles. These advanced functions are among the possibilities announced by OpenMOSS; not all of them have yet been evaluated in our NOVA trials.
Z-Image? A new favourite has just arrived
We liked Z-Image a great deal. But discovering MOSS-Image is already stirring that small thrill you feel when a model opens up new creative possibilities.
The full comparison will come as productions go on. For now, no point spoiling the pleasure: MOSS-Image has clearly just earned a place in our hearts, and in NOVA.
The verdict of these first Axiris trials: an exceptionally seductive start. And that was only the installation.
Official sources
- OpenMOSS: official MOSS-Image repository
- MOSS-Image-SFT: weights and documentation on Hugging Face
- MOSS-Image-Base on Hugging Face
- MOSS-Image: ComfyUI support and installation
First Axiris renders produced in NOVA on 18 October 2026. The observations on the image and on skin rendering come from our own tests, not from external benchmarks.