Why AI Posters Look Bad, and What the Good Ones Do Differently
A Hacker News post in September 2026 titled *AI-generated posters don't have to be bad* drew one of the largest comment threads of the month. The observation underneath it is hard to argue with: AI posters have a look. Give the same prompt to three models and you get three images that share an unmistakable family resemblance — the same lens flare, the same over-saturated gradients, the same confident typography that spells nothing.
The usual diagnosis is that the models are not good enough yet. That diagnosis is wrong, and the proof is that the work being produced with the same models can look genuinely good. The difference is in what the tool is being asked to produce.
Generators produce a frame. Posters are a system.
An image generator is optimised to produce one plausible rectangle. A poster is a system with constraints: a hierarchy that tells you what to read first, a type treatment that has to survive being read from across a room, a colour relationship that holds across the whole sheet, and a composition that works at both thumbnail and full size.
When you ask a generator for "a poster for a jazz festival," you are asking for all of that at once, and the model satisfies the request by producing the average of what it has seen. The average of a million jazz posters is a generic jazz poster. That is not a failure of capability. It is what averaging means.
The work that looks good comes from a different structure entirely: someone decides the system — the grid, the type scale, the palette, the hierarchy — and then uses the model to fill slots inside it. The model is a texture generator, not an art director.
The specific tells, and what causes them
Four patterns account for most of the badness.
Typography that does not exist. Diffusion models render text as visual texture. It looks like lettering at thumbnail size and becomes nonsense at print size. Anything with words in it needs those words composited as real type afterwards, not generated.
No hierarchy. A generator has no concept of primary and secondary. Everything in the frame gets equal attention, so nothing leads. Human design spends most of its effort on what to de-emphasise.
Palette drift. A poster needs a constrained palette. A generator will happily use forty colours if the prompt implies richness. The result reads as noise at distance.
Perfect composition. Real posters have a focal point and deliberate dead zones. Generated images tend to distribute interest evenly across the frame, which reads as a stock photo with a headline bolted on.
None of these are fixed by a better prompt. They are fixed by changing where the model sits in the pipeline.
What actually works
The reliable workflow is compositing, not generation.
Generate the *elements* — a texture, a background, an illustration, a figure — and compose them in a real design tool with real type. Midjourney and Leonardo AI are excellent at producing usable texture and illustration layers; neither is good at producing a finished poster, and treating them as the former rather than the latter is most of the improvement.
Use Adobe Firefly when the elements need to be commercially safe, since its training data is licensed. For anything that has to survive legal review, that property matters more than raw quality.
Then the constraint that does the most work: build a template, not a prompt. A grid, a type scale, a fixed palette, defined slots for image and text. The prompt becomes "fill this slot with a texture that reads as X at this crop," which is a much easier request to satisfy well. Iterate on the template once and reuse it, rather than re-prompting from scratch every time.
Character consistency is the same lesson applied to people. Leonardo AI and Stable Diffusion with a trained LoRA can hold a face across frames; a plain text prompt cannot, because the model has no persistent identity to return to. If a campaign needs the same figure in six posters, you need a reference the model can condition on, not a description it can reinterpret.
A different idea entirely: simulate the medium instead of approximating it
One October 2026 launch is interesting precisely because it abandons generation. Stillwet is a gallery where AI models write every brushstroke as code, and a simulation renders that code as oil paint or watercolour. There is no image generator anywhere in the pipeline. Each painting ships with a session you can replay to watch it being made, plus a written journal of what the model decided to change between passages.
The results look different from generated images — not better in every case, but different in a specific way: they have the behaviour of a medium rather than the appearance of one. Brushstrokes accumulate, edges vary, and the mistakes are the kind of mistakes a painter makes rather than the kind a sampler makes.
For anyone making posters or editorial work, the transferable idea is not the tool. It is the architecture. Constrain the model to a structured, inspectable output — code, layers, a template slot — and let a deterministic renderer produce the pixels. You get consistency, replayability and the ability to change one thing without regenerating everything.
Where the assets end up
A practical note that catches people out. Once you have good elements, finding them again is harder than making them.
SCM solves this for anyone on a Mac with a large media folder. It runs deep local search across every photo and video frame using CLIP, SigLIP or EmbeddingGemma embeddings, plus OCR in 36 languages and Whisper transcription for anything with speech. You can find the exact frame where a texture appeared, or the clip where someone said the thing you half-remember. Everything runs locally — no account, no cloud, no upload.
For poster work specifically this turns a folder of two thousand generated textures from a liability into a searchable library, which is the difference between a reusable asset and a one-off.
The short version
AI posters look bad because a generator is being asked to be an art director and is answering by averaging. The fix is structural, and it is not waiting for a better model.
- Use the model for elements, not the finished piece. Texture, illustration, figures.
- Composite real type. Never generate words you intend to print.
- Build one template and reuse it. Constraints are what produce a look.
- Constrain the palette deliberately. Fewer colours reads better at distance.
- Make your assets searchable. Reuse beats regeneration every time.
If you take one thing: the good work is not the work with the best prompt. It is the work with the best structure around the prompt.
FAQ
Why can't AI models render text properly?
Because they produce pixels by sampling, not by typesetting. Text is generated as visual texture that resembles lettering, which breaks down at print resolution. Words that must be legible need to be composited as real type.
Which model is best for posters?
None of them, for a finished poster. The useful question is which produces the best *elements* for a compositing workflow, and that depends on whether you need commercial safety, character consistency or specific stylistic control.
Does using a template make the work look generic?
The opposite, usually. A template constrains the wrong degrees of freedom and frees you to spend attention on the image and the message. Hand-prompted one-offs drift more, not less.
Is local media search worth setting up?
If you have more than a few hundred generated assets, yes. The first index build is slow, but after that finding "the blue texture with the diagonal light" takes seconds instead of scrolling.
