Here are a few pics I created in Gemini. The image quality I feel is slightly better than Image-fx. The downside is only generates one image per turn.

You know it's the same engine? Imagen4. Same as Sora uses the same engine as GPT4o.

But it seems to follow my instructions better.

There will be variations in the prompt understanding. Gemini and GPT4o reinterpret your prompts before generation. Sometimes that's beneficial othertimes not so much but the engine is the same.
With Imagen or AIStudio you will need to prompt more precisely but have better control over the outcome.
The disadvantage to Sora and Imagen for me is that the CFG is too low. You get very little variation of the same prompt between generations. Image quality though.. the same if you prompt it well.

Maybe so, but it does seem to handle text better. The 'Win £100' signs for example, with GoogleFX it produces random non-English characters that I didn't ask for, and looks very clearly AI.

It still looks good. In fact it wasn't so much the WAM elements but the totally realistic cars in the first one that caught my eye.

It's the same issue that's been recurring since Dall-e/Bing Image creator was released. Knowing how to prompt is 100% the key. Each model has nuances. Using Gemini or GPT4o amends what you ask.

I took an educated guess at your prompt so the nuances may differ but These were generated directly with Imagen4.

What tool you use doesn't really matter but at least try and understand the reasons why each differ and how it can help you rather than spreading incorrect conclusions.

Image-Fx is Imagen 3 and Gemini and Whisk are Imagen 4, so there is a big upgrade from Image-Fx.

Perhaps you are correct and I stand corrected. That said, my point remains the same. Learn to tailor your prompts per engine. Also.. Why use use something significantly worse when better exists on the market. We've had image generators for over 2 years now and still have the same responses.
You can also use https://aistudio.google.com to access imagen4 and imagen4 ultra directly if that's your jam for free.

Which generators are you referring to as worse/better? Which one do you think is the best?

Honestly, there is very little choice between public models. Sora with custom instructions or Imagen4/Ultra. In some cases I do prefer Imagen3 over 4 depending on the context.
Outside of censorship, the biggest problem with Sora is that people generally don't prompt away the yellow hue's. It can be removed quite easily. The other thing is overloading the prompt. There is a finite level of detail available for a resolution. Over complicating the prompt can kill the quality. Background people look terrible, distortions etc.

I'd avoid any of the other models. I'm a big fan of ideogram for graphic design but it's terrible for anything wam related. Plus it's censorship is terrible. People seem to keep going back to Krea but their models were basically SD models, then moved to Flux and now something separately trained. Either way, it doesn't have the complexity or context as Sora/Imagen.

Open Models. Custom Flux or Hidream. Wan can also do decent images as well video. A lot of the other image sites popping up as basically SD/Flux models despite what they may say. I've seen my custom loras I posted on CivitAI ripped and used on other sites without my permission. SeaArt a notorious one for it.