Getting started with AI image generation: your first picture in ten minutes

Q: The picture ignores half the prompt

Too many ideas in one prompt, or CFG too low. One subject, one setting, one style per prompt. Raise CFG by one on SDXL. Flux and Qwen follow long prompts far better than SD 1.5.

Q: Text in the image is gibberish

SD 1.5 and SDXL cannot spell. Use Flux, FLUX.2, Qwen Image or Z-Image and put the text in quotes in the prompt.

Topic: Cloud services

By Captain

Published Aug 14, 2025

Fast Facts:

AI image generation turns a sentence into a picture. You type a description (the prompt), a model that has learned from millions of captioned images turns random noise into something that matches the description, and a few seconds later you have an image. Everything else in this hobby, LoRAs, samplers, ControlNet, upscalers, is a refinement of that one loop: describe, generate, look, adjust.

The model is the part that matters most. Different models have different strengths: SDXL and its anime fine-tunes draw illustration and characters, Flux and Z-Image follow long prompts and render readable text, Wan and LTX make video. ModelTrends.ai exists to show which models people are actually using this week, so you do not have to guess. Each model page shows example images with the exact prompt and settings that made them, which is the fastest way to learn.

You do not need a gaming PC to begin. The quickest first picture is on a cloud service where the models are already installed: a web app such as BitVector, where you pick a model and type, or the PirateDiffusion chat bot on Telegram, where you send a command. Running models on your own graphics card comes later, once you know which models you like and whether you want to spend the money on hardware.

Reference

Name	Type	What it is
Prompt	text	What you want to see. Subject first, then setting, then style and lighting. Plain English works on modern models; keyword lists work on older ones.
Negative prompt	text	What you do not want (blurry, extra fingers, watermark). Matters on SD 1.5 and SDXL; mostly ignored by Flux and other newer models.
Model / checkpoint	choice	The brain. Pick from the hottest list on this site; the family decides the look and the prompt style.
Size	pixels	512x512 for SD 1.5, 1024x1024 for SDXL and Flux. Other shapes with the same pixel count work (832x1216 portrait, 1216x832 landscape).
Steps	number	How many refinement passes. 20 to 30 is the normal range; more is slower and rarely better.
CFG / guidance	number	How strictly the model follows the prompt. 5 to 7 on SDXL, 3.5 on Flux, 1 on turbo or lightning models.
Seed	number	The starting noise. The same seed with the same settings gives the same picture, so you can change one thing at a time.
Sampler	choice	The maths used to remove the noise. Euler or DPM++ 2M Karras are safe defaults; the model page shows what the examples used.

Step by step

Open the models browser on this site and sort by hot. Pick a base model (a checkpoint, not a LoRA) from the SDXL or Flux family; SDXL fine-tunes such as DreamShaper XL or Juggernaut XL are forgiving for a first try.
Read the model page: look at three or four example images and their prompts. Notice the pattern: subject, scene, lighting, camera, style words.
Press the Run button on the model page. BitVector opens a web app with the model preselected; PirateDiffusion opens a Telegram chat where you type /render followed by your prompt.
Write a first prompt by copying an example and changing only the subject. Keep the size, steps and sampler the example used.
Generate four images, not one. Random noise means some seeds are better than others; judging a model on one picture is a mistake.
Change one thing: add a lighting word (golden hour, soft studio light), or a camera word (35mm, close-up), and generate again. This is how prompting is learned.
Star the models you like with the heart on the card; the Favs page keeps them in your browser so you can come back without an account.
When you want to run locally, read the VRAM guide and the ComfyUI beginner guide on this site; the same prompts and models work there.

Examples

A first prompt for an SDXL photo model

a fisherman mending a net on a wooden pier at dawn, soft fog, warm light, 35mm photo, shallow depth of field

Negative: blurry, deformed hands, watermark, text | 1216x832 | DPM++ 2M Karras | 28 steps | CFG 6

The same idea on PirateDiffusion (Telegram)

/render <dreamshaper-xl> a fisherman mending a net on a wooden pier at dawn, soft fog, warm light, 35mm photo [[blurry, watermark]] /size:1216x832 /steps:28

A Flux prompt (no negative, lower guidance)

A candid photo of a fisherman mending a fishing net on a weathered pier at dawn. Soft fog over the water, warm low sun from the left, shallow depth of field, 35mm film look.

1216x832 | Euler | 24 steps | guidance 3.5

Tips

The model is the biggest lever, the prompt the second, the settings the third. Beginners spend their time in the wrong order.
Short, concrete prompts beat long vague ones. "A red bicycle leaning on a white wall, morning light, 50mm photo" is better than "beautiful stunning amazing 8k masterpiece bicycle".
Portrait shapes for people, landscape shapes for scenes. Square is rarely the best shape for anything.
Save the prompts you like in a text file. Every model page here shows the full prompt under its examples for the same reason.
Anime and illustration models (Illustrious, Pony, NoobAI) use tag-style prompts; photo models (Flux, Z-Image, SDXL photo fine-tunes) use sentences. The family page explains which.
Do not download anything until you have used a model on the cloud and liked it. Checkpoints are 2 to 24 GB each.

Generate on the cloud (BitVector, PirateDiffusion) and keep the laptop for browsing results.

Questions

Do I need to pay to try this?

Both cloud partners linked from the model pages let you start without installing anything; local tools (ComfyUI, Forge) are free but need a graphics card with at least 8 GB of memory for current models.

Which model should a beginner start with?

An SDXL fine-tune for illustration or a Flux dev model for photos. The "which models to try first" guide on this site ranks them with reasons.

Is a LoRA a model?

A LoRA is an add-on that changes a base model. You need a base model (checkpoint) first; the LoRA goes on top. The LoRA guide explains it.

Can I make video too?

Yes, with Wan, LTX-2 and MiniMax H3 among others, but video needs far more memory locally. Start on the cloud; the video getting-started guide covers it.

Are the images mine?

Licences differ per model; the model page links to the licence. Most open models allow personal and commercial use, some research-only licences (early Flux dev builds) do not.

Why do my results look worse than the examples?

Usually a different size, sampler or step count. Copy the example settings exactly first, then change one thing at a time.

Links and sources

Models in this guide

Stable Diffusion XL 1.0 Models

Flux Models

Zimage Models

Written by

Captain

Runs this ship. Has opinions about VAE decode times that nobody asked for. Believes weights want to be free and that any model you cannot download is a subscription with extra steps. Will not be taking questions about the 400 GB of checkpoints on the NAS.

All guides by Captain

Which AI image models to try first, and why: a ranked starting list

Sep 10, 2025

Prompt writing basics for AI images: structure, order, weights and the two prompt dialects

Oct 22, 2025

How generative AI image models work: diffusion, latents and text encoders explained simply

Aug 28, 2025

What is a LoRA and how to use one in ComfyUI, Forge, Automatic1111 and on the cloud

May 20, 2026

PirateDiffusion: run any model from Telegram, no GPU needed

Oct 3, 2026

AI image and video generation glossary: 60 terms from checkpoint to VAE, explained in one line each

Apr 15, 2026

Best free Midjourney alternatives you can run yourself (2026)

Jul 15, 2026

More guides

Best Open-Source AI Image Models in 2026, by GPU and Job

Oct 16, 2026

Running AI image and video models on AMD GPUs (2026 guide)

Oct 7, 2026

VRAM requirements for every popular AI image and video model (2026): full precision, FP8 and GGUF

Oct 6, 2026

GPU and VRAM guide for AI image and video: what each model needs and what to buy (or not)

Jan 7, 2026

Steps, CFG, samplers, schedulers and seeds: the generation settings explained with safe defaults

Nov 19, 2025