Understanding an AI model: what it is, what you can ask of it
An AI model is the engine that produces the result; it is neither the application that runs it nor the workflow. What its weights are, why two models answer the same instruction differently, and where to read its documentation.
An artificial intelligence model is the engine that analyses what it is given (a text, an image, several references) and produces the result; it is not the application that runs it, and ComfyUI, for instance, is not a model but an environment that loads several of them. Each model has its own capabilities: photorealism, language understanding, text in the image, editing, references, speed, memory. At Axiris, understanding a model serves to know what you can reasonably ask of it within a workflow, not to name the best one.
An artificial intelligence model is the engine that analyses what it is given and produces the result. In imaging, it is designed to generate a visual from a text, transform an existing image, interpret several references, animate a shot or finish a material.
At Axiris, understanding a model serves one purpose: knowing what you can reasonably ask of it within a workflow.
The model is not the application
The model and the interface that runs it are two different objects. The same model can be used in ComfyUI, in a specialised application, behind an API, in a web service or in an automated workflow.
ComfyUI is not an image model. It is an environment that loads models and chains operations; the ComfyUI section explains it. Nova, the Axiris AI, is not a model either: she is the assistant who turns your intention into instructions for the chosen engine. See Nova in the workflow.
The model's weights
When a model runs locally, it relies on files that hold the parameters learned during training: its weights. Depending on the technology, several files contribute to the result: the diffusion model, the text encoder, the VAE, a checkpoint, LoRAs, control components.
Downloading a model is therefore not always enough to get a working workflow: you also need the components it expects, and a machine able to load them. See How much VRAM to generate AI images?.
Each model has its capabilities
Two models react very differently to the same instruction. They differ in photorealism, language understanding, composition, text in the image, character consistency, editing, visual references, styles, speed and memory consumption.
That is why Axiris avoids the question "which is the best model?" and prefers: which model matches the work we want to do? The answer is built in How to choose an AI image model?.
Architecture, training, behaviour
A model's behaviour depends on its architecture and on the data and methods of its training. You do not need to master the mathematical research to use it. But knowing its main characteristics (what it was trained on, what it accepts as input, what it can edit) lets you understand its strengths and limits before putting it into production.
The official documentation, then our tests
When we document a model, we first look for its source: the publisher's site, the official documentation, the repository, the model card, the technical paper, the licence. We then keep that information apart from our own observations.
The documentation describes the model. Our tests show how we can really use it. The method is described in Comparing AI image models; the section AI models for image creation gathers the families we follow.
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.