Generative art, from prompt to pipeline
Image and video generation at three depths: no code browser tools, intermediate local setups like ComfyUI, and advanced training of your own models. Costs labelled honestly. The theory behind all of it lives on the Learn page.
Suggested curricula
First images, today no code
Anyone — no code, no GPU · a weekend
From zero to images you're genuinely happy with, using only free browser tools. The skill is seeing and describing, not software.
- Generate the same prompt in two tools — Gemini and Grok Imagine are both free — the differences teach you what models actually do
- Learn prompt anatomy — subject, medium, style, lighting, composition — Midjourney's docs teach it best, and it transfers everywhere
- Iterate, don't re-roll — change ONE thing per generation and watch what moves; ten deliberate iterations beat a hundred lucky dips
- Try text and typography in Ideogram — posters and titles are where most models fail; see what a specialist does differently
- Animate a favourite still — image-to-video in Grok Imagine or Kling — the fastest "wow" in this field
The ComfyUI path intermediate
Comfortable installing software; GPU with 8GB+ VRAM (or patience) · ~6 weeks at 4 hrs/week
Own your pipeline: unlimited local generation, community models, repeatable workflows. Node graphs look intimidating for about a week, then become the point.
- Install ComfyUI (desktop app) — the installer handles Python and models; git only if you like git
- Official first-generation tutorial — your first local image, and a map of what the default nodes mean
- ComfyUI Wiki basics course — learn what each node does so graphs stop being magic
- Pull two community models from Civitai — one photoreal, one stylised; compare against the same prompts from your no-code days
- Rebuild three official example workflows — img2img, inpainting, then a Wan video workflow — rebuilding beats downloading
Train your own style advanced
ComfyUI path done; 12GB+ VRAM ideal · ~1 month of evenings
A LoRA that reliably produces YOUR style, character, or subject. Dataset quality decides everything; the training itself is the easy part.
- Read the sanj.dev LoRA guide end to end — before touching a trainer — most failed LoRAs die at dataset prep
- Build a 20–40 image dataset — consistent subject, varied everything else; caption honestly
- First run in FluxGym — the simplest trainer that produces real results; iterate on the dataset, not the knobs
- Graduate to ai-toolkit — when you need newer models or finer control
- Understand what you just did — the HF course + the diffusion-theory shelf on the Learn page close the loop
Tools & guides
No-code image generation no code 7
Type a prompt, get an image — nothing to install. The assistants you may already use (ChatGPT, Gemini, Grok) now generate images well; the dedicated tools below give more control.
No-code video generation no code 7
Text-to-video and image-to-video in the browser. Clips are short (5–15s) and credits run out fast on free tiers — but the quality in 2026 is startling. (OpenAI's Sora was discontinued in April 2026; it's not on this list for a reason.)
Prompting & craft no code 3
Generating a good image is a skill: describing subject, style, lighting, and composition precisely. These teach the craft that transfers across every tool above.
Local & node-based — own your pipeline intermediate 7
Some setup required: a decent GPU, an installer or a git clone. In exchange: no credits, no content filters beyond the model's own, total control, and workflows you can save and share.
Train your own — LoRAs & beyond advanced 5
Teaching a model your style, character, or product. A LoRA trains in hours on a consumer GPU; the tooling below is what the community actually uses in 2026.