AI Image Generator for Beginners: Models, Settings and LoRAs Explained

Every AI film starts with its cast. Before you can build a storyboard or animate a scene, you need clear images of your heroes, any creatures they meet and the world they live in. In this lesson you will learn how the Proxima Art image generator works, which models exist and what they are good for, and how LoRAs change the look of your images. You do not need any prior experience: we will go through every setting you see on the screen.

Where to find the image generator

Open the top menu Create and go to the generator. You will see Generate Images & Video with three tabs: Image, Video and Magic. Stay on Image. The page has four main parts, from top to bottom:

1. Choose a Model — the AI model that draws your picture.
2. Mode and quality settings — they change depending on the model.
3. Prompt — your text description of the picture.
4. Number of images, LoRAs, aspect ratio and Expert Settings, then the Generate Image button with its price.

Under the button you always see two balances: Daily Free Credits and Credits. Every account gets 100 free credits every day, and generations use the daily free credits first.

Which AI model should you choose?

A model is the “artist” that turns your words into a picture. Proxima Art offers a wide selection, and each one has a different taste. Click the model field to open the list: every entry has a short description, and badges show if it is New or Premium. Explore All Models opens a page with examples for every model.

Prices are in credits and can change; the Generate Image button always shows the exact total before you click.

Premium models and censorship

Models marked Premium need a subscription. Several of them — for example GPT Image 2, GPT Image 2.5, Flux 3 and Grok Imagine 2 — also carry the note “Strict censorship in place”: they refuse some subjects, such as nudity or graphic violence. If a model keeps refusing your prompt, switch to a model without that note.

Qwen 2.1 settings, one by one

This tutorial uses Qwen 2.1, so here is what each of its settings does:

• Lite (marked “lower cost”) — an alternative way of generating with Qwen 2.1 at a lower cost, 15 credits per image. Results differ from Creator and are sometimes even better.
• Creator — full-quality generation with maximum creative control and LoRA support, 20 credits per image.
• Native 2K Resolution — renders directly at about 2 megapixels for sharper detail and text. It costs more and allows up to 2 images per click.
• Enable Upscaling and Refinement — enlarges the finished image (to 2K or 4K) and adds detail. With Native 2K on, only 4K is offered.
• Number of images — up to 4 per click. Two is a good habit: you can compare and keep the better one.
• Aspect Ratio — 1:1, 16:9, 9:16, 3:4 or 4:3. Use 16:9 for wide, film-like frames and 3:4 for full-body portraits.
• Expert Settings — seed, CFG, steps and sampler. Beginners can leave them as they are.

How to write a good prompt

A prompt is your description of the picture. The model only knows what you write, so be concrete. A reliable order is: shot type → who → appearance → clothing and props → setting → light and mood. In the video the first heroine starts like this:

Full-body character portrait of a beautiful 25-year-old blonde warrior with European features, standing on a remote desert planet. She wears practical nomadic clothing: a sand-colored tunic, layered weathered fabric, leather belt and a headscarf… An ornate dagger hangs from her belt, while a large translucent crystal sword rests at her side…

Notice that it describes one character in detail and says exactly what she carries. Those details matter later: the sword and the dagger become part of her identity in every scene.

What is a LoRA, and how do you use it?

A LoRA is a small add-on model. It does not replace the main model; it teaches it something extra — a visual style, a type of lighting, sharper details or a specific concept. Think of the base model as a painter and a LoRA as a set of special brushes and a style guide you hand to that painter.

Adding LoRAs

Open Apply Lora (Styling). You can search by name or open groups such as Premium Styles. Click a LoRA to add it; it appears at the top with its strength in percent. To remove all of them, use the bin icon next to “Selected Loras”. The list only shows LoRAs that work with the selected model — LoRAs are trained for one model family, so Qwen 2.1 has its own set.

A few LoRAs available for Qwen 2.1, to give you an idea:

• GPT style (Premium) — luminous celestial-fantasy style with gilded detail and cinematic glow. Used for our heroines.
• Fantasy Film Still Realism — pulls painterly fantasy toward believable cinematic photography; works best at about half strength.
• Cinematic Photo Lighting — dramatic, directional light and film-like colour.
• Epic Fantasy Illustration — majestic, storybook fantasy art.
• Qwen 2.1 Fix — repairs small mutations such as hands and fingers; it works best on its own because it weakens style LoRAs.

Choosing the strength (weight)

The slider under each LoRA sets how strongly it is applied. Each LoRA starts at a sensible default; the description sometimes recommends a range (for example “works best at about half strength” or “keep the weight below 60”). Some LoRAs are sliders that also accept negative values to push the effect the other way. In the video, two LoRAs are used: one at 90% and one at 60%.

How to pick LoRAs

• Decide on the look first (realistic, cinematic, fantasy illustration, anime) and choose one LoRA that matches it.
• Add a second LoRA only for a clear reason, such as better lighting or detail.
• Read the description: some LoRAs are made for editing photos rather than generating new images.
• Premium LoRAs need a subscription; without one they are switched off automatically.

Tips for a consistent cast

• Keep the same settings for everyone. In the video the second heroine, the beast and the world are made with the same model and LoRAs, so they belong to one film.
• Make references, not final shots. The beast and the world only need to show what they look like; the storyboard will place them in scenes later.
• Describe one subject per image. A clean image of one character makes a much better reference than a crowded scene.
• Generate two at a time and keep the best. Your images are saved to your library, so you can pick them later in Character Asset Pack.

Common mistakes

• Too many LoRAs at full strength — the classic cause of melted faces and strange artifacts. Lower the weights or remove some.
• A vague prompt such as “a warrior girl” — you get a random warrior every time. Describe age, hair, clothing and props.
• Switching models between characters — the cast stops looking like one world.
• Fighting a censored model — if a Premium model refuses your idea, change the model instead of rewriting the prompt again and again.

Next, you will turn one of these images into a consistent, reusable character with Character Asset Pack.

How we used this in Sleeping Sands of Kharuun

The opening workflow creates four kinds of reference: Iraya, Ashara, the two-headed beast and their desert world. At 01:14 the tutorial introduces the image model and LoRAs; at 01:57 it continues with the companion, beast and environment. These are source images for the later character and storyboard steps, not the final animated shots.

Keep each reference readable: a character image should show the face and clothing clearly; a creature reference should make its two heads unmistakable; a world reference should establish the desert look. The later storyboard combines these ingredients into scenes. The examples above explain the workflow, rather than promising that a fixed seed or LoRA weight reproduces the exact film.