Text-to-Image
Type a description, get a picture that didn't exist — the capability that made AI visible to everyone, and the one with the most unresolved argument underneath it.
When not to use it
- When you need a specific thing. "An image like this" is easy; "this image" is not. That gap is where professionals still live.
- When provenance matters and training data is undisclosed. For commercial media, you're accepting a risk on someone's behalf.
- For text in images, counting, or spatial relations. These are known weak spots rooted in the text encoder, not fixable by prompting harder.
- When a stock photo would do. Sometimes the licensed image is faster, cheaper, and legally clear.
Reach for something else instead
- Licensed stock — clear provenance, no argument, unremarkable.
- A human illustrator — for anything where specificity or intent matters.
- Models trained on licensed data — the provenance answer, at some quality cost.
- Image editing (inpainting, conditioning) — usually what you actually wanted rather than generation from nothing.
Sources & further reading
- Rombach et al. (2022), High-Resolution Image Synthesis with Latent Diffusion Models — Stable Diffusion; moving diffusion into latent space is why this escaped the datacentre.
- Radford et al. (2021), Learning Transferable Visual Models From Natural Language Supervision — CLIP; the alignment that makes text guidance possible.
- Carlini et al. (2023), Extracting Training Data from Diffusion Models — memorisation is real, rare, and legally consequential.
Primary sources, listed so you can check the claims on this page rather than take them on trust.
Where people go wrong
- Cranking guidance scale to force prompt adherence, and getting oversaturated, rigid images. 7-8 is the range.
- Blaming the image model for prompt misunderstanding. It's usually the text encoder — attribute binding is a known weakness.
- Treating prompt folklore as technique. It's empirical, model-specific, and expires with each version.
- Assuming output is automatically clear of the training data. Memorisation happens.
- Expecting consistency across generations without conditioning. Same seed, same prompt, same image — that's the only guarantee.
At a glance
FieldGenerative AI
Pipelinetext encoder → latent diffusion → decoder
Key knobguidance scale, 7–8
Weak attext, counting, spatial relations
Blockerprovenance, not quality
DifficultyBeginner
Flashcards for this concept
Question
—
Answer
—
1 / 4
Often compared with
Text-to-image vs. stock photography — one is instant and of contested provenance, the other is licensed and generic. Right now that's the actual trade.