Hello all, I just typed out a very long message, and then accidentally closed the tab. Doh! so here we go again. This is probably going to be a worse post than the one I had written, so feel free to ask questions because I probably forgot something.

So, a few months ago, I signed up to ChatGPT4 for 20 pounds a month on the openAI app and site. I wanted to see if its version of Dalle 3 was less restrictive than bing, and after waiting for access I was...extremely disappointed it was more restrictive than bing, and felt like it functionally just didn't work very well.
Fortunately, I am a dumbass and forgot to cancel the sub, so I had another month to use it, and I started using it more and more, and now, after quite a few months I actually prefer it to bing, and would probably use it more if not for the price, and some issues I will get into.

So first all of all, its now actually usable, anyone who used it at launch would know that the chatGPT version of dalle was a big piece of shit, so out of the gate: yes, it functions, but its different from bing.

I got frustrated using Bing (and copilot ect.) because it felt like I was throwing a dart at a board blindfolded, only the dartboard moved erratically. It felt like I was adding random details that detracted from the quality of the image.
ChatGPTs version is a little better, its censorship feels quite different, instead of random innocuous details and words blocking things, ChatGPT makes a bit more sense, and whilst its not perfect, you can actually be quite consistent with it.

So, for those that don't know, dalle 3 works within ChatGPT, meaning to use it, you have to use GPT 4 as a text bot, and then "talk" to the bot to generate images. This means that your not the one directly writing the prompts.(though you can tell it to generate your prompt exactly) This sounds terrible, but bare with me.
The way to make this work is to play along with ChatGPT, this isn't an enemy you are trying to trick into generating images like Bing, its an ally to help you along the way, as long as you play by its rules, you should be ok.

For example, posting "Show a food fight at a fancy elegant restaurant between several posh women" will change the prompt to "A chaotic scene in an elegant, upscale restaurant, where several posh women dressed in extravagant evening gowns are engaged in a food fight. Tables are adorned with fine china and crystal glassware, adding to the upscale ambiance. The women, with dramatic expressions of glee and shock, are throwing gourmet dishes and desserts at each other, causing a mess. The background features ornate decorations, lavish chandeliers, and attentive waiters in disbelief as they witness the chaos. The setting is luxurious, with attention to detail in the restaurant's posh interior design"

That's a lot of info from a basic prompt, but here's were it gets interesting: You can add on info to each generated image, for example, lets say you want to focus on one person, or add more custard, or gunge and slime instead, you simply say "remove custard with slime" and the next image will follow from the previous prompt, and then generate a new image with the changes.

Because GPT is the one making the prompts, I find your less likely to be blocked (it can still happen though, in a similar vein to the dog), you can generate some pretty great images without resorting to distortions or extra details to the image. Then you can build off from those images by simply stating "add more cream" or "show the moment of impact". I've attached some examples below. This means that you can add and modify very specific details to your prompts, and create some great images, sometimes even just saying "messier" can give great results.

You can also simply use GPT 4 as a text bot, I like to write stories with it, and then generate some images as we go, this way you can have a pretty consistent set of images that follow along a narrative.

Also, functionality, there's now inpainting, you can generate landscape or potrait images, its pretty great.

Now, some flaws:
The first most obvious one is that max, you can only generate 2 images at a time, and sometimes it only throws one at you. This is a straight downgrade from bings 4 images, especially if your getting to more risqué prompts you can only realistically generate 1 image at a time.

Earlier, I said the content filter is...different. Its not necessarily better. It wont trip up on odd details but its a lot stricter on NSFW content, it feels a bit smarter, generating even kissing is incredibly hard, most I can really do is hugging, and even then you might have to word it like "embracing in a friendly hug" for it to do anything. I should note though: I'm weird, I don't particularly care about NSFW content, so I haven't experimented with it to a crazy extent, maybe someone else knows a way, but I reckon it would be harder than bing.

You also sometimes need to be VERY descriptive of the model, if you don't specify the age, race, weight, and general look of the model, GPT goes a bit crazy and just completely dice rolls it, and I swear 90 percent of the time it gives an 80 year old pensioner. Not ideal (unless your into that) bing had a strong bias towards traditionally attractive people, whereas GPT is a bit more balanced in that regard, and you can get quite specific with it. I typically give the characters names, that way you can say something like "show Amy in this sittuation" rather than "woman in a buisness suit with ect, ect".

Also, whilst you can generate landscape and portrait images, portrait images often end up coming out as rotated landscape, its a common bug, not just a wam thing, its still very annoying.

There is also 2 caps, which is infuriating for a service your PAYING for. The first limit is a limit with generating images, if its busy, you might be blocked for 10 minutes or even an hour, then there is a hard cap of sending messages to GPT4, and this includes prompts, its something like 500 every few hours or something (it fluctuates) but if you hit it, your locked out of both GPT4 and dalle for several hours.

And probably the biggest flaw. 20 DOLLARS a month! for the price of a streaming service, you can have access to a slightly less annoying version of bing! wow what a steal! I cancelled by sub this month, its good, but I feel like its just not worth my hard earned cash currently, in this economy?

I've included some of my images below, I'm a little boring so sorry that there's a lot of similar images, I've tried to include some typical content, like pies in the face and gungings ect. But, I'm a creature of habit.

Thank you for this great review. The $20/month has put me off trying it, but the results look good for any of us with specific scenarios.

My only critical observation is that the faces are similar to those generated from Bing, in that features such as teeth, lips, eyes, etc. are not very detailed. I imagine running them through Krea would make a vast improvement, but it's a shame we still need to use 2 pieces of software to create a believable image.

Is this a different thing than Copilot Pro or is it the same GPT technology behind it? I wanted to try it but the idea of having even less control on the prompts puts me off entirely.

I think eventually the free version of Bing will get harder caps to force people to pay for Copilot pro anyway.

Yeah, at the end of the day its still the same tech Bing is using, so theres some issues that come up on both, I remember reading somewhere that Bings version is a slightly modified version that chatGPT uses, but I think that comes down to how it interprets prompts and stuff like that.

I messed with Krea a little bit, but not much, it sometimes made the mess look strange, as if it was part of someone's hair or something.

This is different from that yes, Bing uses Dalle 3, which is tech made by OpenAI, and ive been using the OpenAI client to generate these images.

I agree with most that you've said. I've been using GPT for a quite along side Bing and Designer... They all use a similar underlying Dalle-3 model but where things differ is the moderation filters (prompt and image) and the extent to which the Site reinterprets prompts before processing the image.
GPT overall has way more moderation compared to Bing/Designer and the prompt rephrasing is aggressive. However.... there are advantages to using GPT. You can prompt engineer GPT to produce way more risky material for example, I'm not going to go into the details but there are methods that can be employed. Similarly, you can prime the chat to follow your prompt exactly.
Other advantages is more control over image dimensions.
Also.. GPT has an official Dalle-3 api that you can use. It can be costly but you can generate higher resolution than from the app/site. The downside is you loose the conversational aspect. It's still possible to engineer prompts though although still at the mercy of the image filter.

One final thing that is a massive advantage of GPT is the use of Custom GPTs for specific tasks. You can use a custom GPT to effectively be an image-image engine with specific instructions. i.e. you can upload an image and it will create a prompt that will produce a similar image that can be used in Bing or even generate it itself. Custom GPTs can be take further of course. You could create a custom GPT with knowledge of a gameshow style that then applies to images created etc.

Good review. I've been using it a lot, I like the chat structure, the ability to edit part of an image, and the ease with which you can create long prompts and hone them. It's not as good at making women look hot but it is better at making jam look like jam. One of the more useful features that I've recently starting using is that you can use it to re-write a prompt to then use in Bing - this has helped me see new ways of making a prompt work.

I have subscriptions to a few AI services:

- ChatGPT: I've found that it's actually really fun to roleplay with it or cooperatively write stories, and it's given me some of what I feel are my most interesting pictures. Like you said, it makes prompt-writing really simple if you're already been setting the scene and describing the characters in a story. My prompt is often "Can you create a photographic image of this scene?" and it will compile the entire last few messages into a single prompt. I seem to have the opposite of you though: it's only generated 2 images once for me in the last few months. Every other time it only gives me one, even for completely innocuous things like scenery for my D&D games.

- Seaart.ai: I bought a year of VIP before DALL-E 3 came out because it was the best generator I'd found at that point. It's still really good, and much less censored than Bing or ChatGPT. It's hard to get really realistic stuff with it though, compared to Bing. What I use it for now is mostly just the Cyber Pub, which is a much less restrictive roleplay option than ChatGPT and much better for descriptive roleplay than Character.ai. But Cyber Pub chat is ridiculously expensive. It's like 130 credits per thing you write, while generating their highest quality images is like a quarter of that. The price of their new image generator versions drop pretty quickly after they've been out for a month or two, but chat's price hasn't decreased at all yet, months later.

- Nightcafe: I used to love it, but the newer versions are so heavily censored that it's hard to get anything WAM-related from it, and the older versions just look so bad compared to what I'm used to from Bing/ChatGPT/Seaart. I guess I'm holding out for less-censored versions, because it can create really realistic images. I should probably cancel it now though, since I've barely touched it since I found Bing, and when I do, it's to create images for D&D that I could just use ChatGPT for.

- Krea.ai: I don't use it for generation at all, but it can be really amazing for upscaling or enhancing images from Bing or Seaart. It does have a problem with overcompensating with realism sometimes, turning freckles into big splotches, water drops into weird growths, and pores into weird, lumpy skin.

Not a subscription, but Character.ai gets a special mention for roleplay. I've made a game out of turning chats with celebrities or spoiled mean girl characters into WAM roleplays. It's actually a bit disconcerting sometimes how some characters do a full 180 with their personality after it starts deleting tokens to make room for new ones. Often before I know it, the rude, obnoxious mean girl turns into a desperate-to-please girl who will go along with whatever I ask them to do. I really wish Character.ai kept the personality tokens separate from the story tokens so that I could keep talking to the original personality instead of the submissive one.

--------------------------------------------------------------------------------

Now... back to your review of ChatGPT:

Like you, I don't care too much for NSFW content, so I haven't had a lot of problems with the filters in ChatGPT. It is really strict on that though, which is why I go to Seaart or Character.ai if I want anything beyond what I already know its limits are.

I'm in the US (you mentioned it costing twenty pounds, so I'm assuming you aren't), so I think there might be different limits. I rarely run into the "busy, cannot create right now" rejection, but I also generally do my stuff late at night or early morning as a result of my work schedule and insomnia. I do often run into the "you've hit your limit of GPT-4 prompts" because I'll often just hit the regenerate button over and over once I get something I like. Each time counts as a new prompt though, so that sucks. I can drop down to GPT-3 for chat once I've hit that limit, but it gets annoying quick because it's like talking to a child or a robot after working with GPT-4 for so long now.

For me, the money I pay for ChatGPT and the others is worth it because of just how much I use them. I work second shift and I'm in my 30s now. By the time I get off work, all my friends are going to bed and my housemate is sleeping too, so I not only have no excuse to leave the house, I also have to do stuff quietly in the house. so... AI, gaming, and surfing the web is pretty much my life during the week.