Today I tried my hand at generating a few inspirational posters. I’m sick of all the clichés you get online but wanted something to keep me inspired and sane.
I started by asking ChatGPT (5.4 thinking) and Claude (4.7 Opus) to generate me 50 positive affirmations. The aim is to motivate me in the mornings and keep me going throughout the day. I asked to generate these based on previous writings and conversations. They were good.
I then wanted to test out OpenAI’s new image generation model (GPT Images 2) that was reported to be better than Google Nano Banana model. To do this I started brainstorming with the following prompt:
Given [the generated positive affirmations] do you have any ideas for an inspirational poster that is mainly visual and not words but that is not cliched new-age or inspirational fluff that I can look at in the morning to get me going? Let’s brainstorm some ideas then after I approve one we can generate an image. One idea is maybe in the form of a movie poster, one in the form of a bus shelter advertising poster, another is maybe those stylised Visit X posters from the UK. We can always iterate on our top N ideas and generate them all.
ChatGPT then returned a series of ideas A to M. I asked to provide these as a Canvas I could run through and add some notes to. I did that – a couple of sentences for each idea, trying to communicate my initial thoughts and feelings on each one. I asked to reduce to the top diverse and orthogonal 5 ideas and then swapped one or two around to aim for different styles and ideas.
The next stage was the generation of detailed JSON prompts for my favourite three. This via the prompt:
Let’s generate detailed and tighter JSON style prompts for 2, 4 and 5. We’ll start with those.
This returns the initial JSON here – https://gist.github.com/benhoyle/494816a354a8806809466d99b52aafc8#file-initial-inspirational-poster-json-prompts
I then asked to generate a slightly tighter set of prompts with some tweaks. This returned the JSON here – https://gist.github.com/benhoyle/494816a354a8806809466d99b52aafc8#file-second-tighter-set-of-json-prompts
I then copied the JSON for each poster into a new chat window for both ChatGPT and Gemini so I could iterate with a clean context for each poster.
With ChatGPT I did need to edit 2-3 times. It performed relatively well on those limited number of edits – making the updates I wanted. If you find that the chat no longer follows your instructions very well, the best idea is to ask for updated JSON and then start another iteration in a new chat window. The JSON format works well to enable you to port specific instructions from one provider to another. You don’t need the specific fields – just get the LLM to generate those. You can find decent templates online if you want more specific control. But remember sometimes the LLM just ignores you anyway!
The final JSON prompts for each poster can be found here: https://gist.github.com/benhoyle/494816a354a8806809466d99b52aafc8 . I also added a fourth idea to a “Morning Shift” style poster that ChatGPT decided should be inspired by a Mike Leigh aesthetic. I’ve added the ChatGPT new image model versions and the Gemini versions so you can compare.
Same prompt. ChatGPT on the left, Gemini on the Right.
I found that for several of them I preferred to personalise with my own face (if I use me then I give permission to use me!) – this made them more resonant. To do this I just added one to three images of myself to the prompt. I used head shots of myself with the background removed that I have on file. This helps the image models concentrate on just me and not the background, or get confused by other people.
I think I actually do prefer the new ChatGPT model. I’m not keen on the default “slightly” speckled image properties but you can reduce this with some default prompt steering. The Gemini Nano Banana versions are also great, but often they look simplistic and a little basic compared to the ChatGPT generations. Amazing how quickly technology moves!
I really like the little details in the new ChatGPT model – the fraying and wear effects are very good. The ChatGPT adds lots of cool little details that are absent from the Gemini models. The ChatGPT images look less “perfect” and less typical “AI”.
Funnily enough, doing this exercise I don’t think “creatives” are out of a job. The “defaults” from the AI are averages, “good enoughs” but not “knock out” amazing. I’ve seen much better and funnier stuff on social media. Also you need to have a visual sense to really design a good poster. AI can generate you something in a particular style but to be distinctive or really inspired you need a trained creative mind. It really just raises the bar on daily life and often shows the continual need for creative talent.
Give it a go with your own ideas based on the workflows and linked prompts!
Robotic Ape 