Gpt-image-1 deprecation may break the visual identity of my game project

[Bliggfang]

Hi everyone,

I am a solo indie game developer working on Fusiomon, an online trading card game where AI-generated monster artwork is a core part of the game.

Over a long time, I built and refined a very specific visual style with gpt-image-1. This style is not just a nice extra. It defines the card pool, the monster identity, the fusion results and basically a huge part of the game’s feel.

Now gpt-image-1 is scheduled to be shut down, with gpt-image-2 listed as the replacement.

I understand that OpenAI has to clean up and maintain its API/model lineup. That makes sense in general. But I honestly struggle to understand why there seems to be no stable path for projects that were built around the specific behavior of an existing model.

For my use case, gpt-image-2 is not a drop-in replacement. I spent hours testing and iterating prompts, trying to recreate the old style. The results are technically impressive, but they look clearly different. More like a new generation of assets than a continuation of the same art direction.

That means I may have to regenerate my entire existing card pool to avoid a hard visual break between old and new monsters. And even then, the same fear remains for the future: what if I build the live game around another image model and that one gets deprecated too?

For image-based products, the model output can become part of the creative identity. A newer model can be “better” in general, but still break a specific established art style.

So I wanted to ask:

Is there any possible way to keep using gpt-image-1 after the shutdown date, maybe through legacy access, higher pricing, limited access, an enterprise option, or any kind of compatibility tier?

Are other developers facing similar problems with model deprecations breaking the creative consistency of their products?

Has anyone found a reliable way to recreate a gpt-image-1-specific style with gpt-image-2?

For my project, this is not a minor inconvenience. It puts a central part of my production pipeline and visual identity at risk.

I would really appreciate any thoughts, workarounds or official guidance on this.

[PaulBellow]

Not likely an older model will return after being dropped.

Maybe we can help fine-tune your prompts for the new model? We’re happy to help.

Good to see you back after a year! Looks like you’re making great progress otherwise.

[jeffvpace]

In addition to what Paul said:

@Bliggfang

But I honestly struggle to understand why there seems to be no stable path for projects that were built around the specific behavior of an existing model.

Join the club. Unfortunately, this is the case for many of us, depending on the model.

I would really appreciate any thoughts, workarounds or official guidance on this.

From my experience, the only thing to do sometimes is to bite-the-bullet and power through it the best you can.

I have used gpt-image-1, 1.5, and 2 extensively, And while gpt-image-2 is far superior, there are nuances that are problematic.

[Bliggfang]

Paul hi!

I actually remember your welcome message from back then. Seems like you’re one of the old hands here.

Thanks for the offer, I really appreciate it. I might come back to that.

The tricky part is that I dont really want to post my exact prompt setup publicly, since the art style is part of what should make Fusiomon special. It’s not just a prompt detail for me, but part of the games identity.

Still, sad to hear there probably wont be a way to keep using the old model. And honestly, a bit worrying when thinking about tomorrow.

[Bliggfang]

Hi jeff and thanks for sharing that.

Its kind of reassuring and sad at the same time to hear that other developers see it the same way.

I guess Im not the only one who finds this “bite the bullet” approach to model deprecations more than questionable.

And yes, I fully get that gpt-image-2 may be better overall. But “better” unfortunately doesnt always mean “compatible” when a project depends on a very specific style.

[windysoliloquy]

I understand that prompts are secrets, but those secrets only take a couple of turns in ChatGPT to unravel at the current level we’re at with AI.

Have you considered taking the approach where you have generations of cards?

That way, your styles change slightly or get better, just like real trading cards did… it also leaves you open to future upgrades and not having to feel like you were left behind.

It seems to me to plan for the future like that is a more reasonable approach for a business model such as yours.

[Bliggfang]

Thanks, I appreciate the suggestion and yes I have considered the idea of treating this as different “generations” of cards, and realistically I may not have much choice anyway.

The difficult part for me is that with GPT-Image-1 I was able to get a very specific, unique visual identity for Fusiomon. With GPT-Image-2, even after a lot of careful prompt engineering and many days of iteration, the results still tend to drift into a much more generic AI-art look.

It is also harder in my case because Fusiomon is built around visual monster fusion. Two existing monsters go into a fusion with one established art style, and then the new fused monster comes out looking clearly different. That kind of mid-system style break is much harder to explain than a normal card generation shift.

[windysoliloquy]

Grab up a bunch of your images and load them into a ChatGPT prompt

Ask the model to tell you how to request the common styles shared by the images you’ve provided (the more the better, generally, 10 is probably enough)

You’ll be prompting for :

“How can i recreate this style using a prompt? What should I specifically ask for? I am not concerned as much about the creatures in them, the setting, but rather the style in which these cards are created. I would like to be able to reproduce that style using the current image model”

[Bliggfang]

@windysoliloquy and @PaulBellow and everyone else, I would like to invite you to try this directly:

Pick any of the attached gpt-image-1 monster images and try to recreate this shared art style with the current image model.

No reference image, only a standalone prompt to gpt-image-2.

Not the exact creature, but the visual identity: softness, rendering, composition, and card-creature feel.

If someone can get consistently close to this style, I would genuinely like to see the prompt approach.

[LarisaHaster]

Hi @Bliggfang ,

Fun challenge and I of course wanted to test. Don’t know what you think about this, but this was the closest I got. Here’s my prompt:

Create a small adorable fantasy fire-dog creature in a soft painterly collectible-card illustration style. The creature should have a round body, fluffy fur, oversized glossy dark eyes, a tiny nose, and a warm friendly expression. Its fur should glow with orange, amber, and golden fire tones, with a fluffy flame-like tail and little flames around its paws. The overall feeling should be cute, magical and polished, like a premium fantasy creature card illustration. Use soft cinematic lighting, gentle atmospheric glow, subtle background blur and a simple dark enchanted background that helps the character stand out. Center the creature as the clear main subject, with a slightly oversized head, appealing proportions, soft rendering, rich detail and a cozy magical fire-element mood.

quality:low size: 976x704

Also.. I´ll attach screenshot too:

[DysTopia]

I was thinking for this sort of aspect, would be nice to have a proposal for LTS Long-Term-Service like Models. While not the same as Code, Frameworks or OSs, it does have an impact on how a model is implemented in any given industry, from creative to research. And while these models are versatile, us humans can only do so much. Learning how a model works is ok. But a Best practices and standard to lead the industry would be cool. Gives more resilience to implementations that are depending on X,Y or Z model

[windysoliloquy]

I just explained to you the only current process to get what you want requires reference images man

I’m not going to chase after something that doesn’t work man.

[jeffvpace]

Did an image analysis and asked gpt-5.5 to create a prompt:

Prompt

Create a square 1024x1024 image of an adorable baby polar bear cub sitting front-facing in the center of a glowing blue ice cave. The cub is small, round, and plush, with fluffy crystalline fur that looks like soft frost, snow, and tiny ice shards. Its fur is bright cyan, turquoise, and white, with spiky frosted tufts around the head and cheeks. It has oversized glossy black-blue eyes with bright white reflections, a tiny black wet nose, a subtle sweet smile, round fuzzy ears with darker teal inner shadows, and small front paws visible at the bottom with dark curved claws.

The scene is a magical arctic ice cavern with tall translucent icy walls on both sides, a cold reflective frozen floor, and soft snow patches. Use a vibrant monochromatic blue palette with cyan highlights, deep teal shadows, and glowing icy reflections. Lighting should be soft and luminous, coming from above and slightly in front, creating sparkling highlights on the fur and ice. The background should be slightly blurred for depth, while the bear is sharply detailed. Style: ultra cute fantasy creature, high-end 3D animated film look, detailed digital painting, soft cinematic lighting, whimsical, clean, no text, no watermark, no extra characters.

Here is what it produced:

Maybe play with the prompt a bit?

Bottom line is that you are gonna have to accept the fact that you may not be able to replicate your characters exactly using gpt-image-2.

[jeffvpace]

Or, you can use the gpt-image-2 images endpoint and recreate your character images exactly:

You can also use the gpt-image-2 images endpoint to modify your gpt-image-1 images without changing the style:

So I quess you don’t have to bite-the-bullet afterall,

Note: This was done via the gpt-image-2 API and not ChatGPT.

[Bliggfang]

Thanks for the attempts, I really appreciate it.

But I think this also shows the core issue: without a dedicated reference image, and even then only with a very explicit copy instruction, it does not really reproduce the same unique style with the new gpt-image-2 model that I was able to get from the old gpt-image-1 model.

For me, this becomes an even bigger problem once monsters are fused together. I attached a few more examples of real in-game, on-the-fly fusions created by actual players in my game. With gpt-image-1, I was able to build a very specific visual identity around that.

With gpt-image-2, interesting and impressive results are definitely possible, but everything tends to look much more glossy and AI-generated and overall less unique.

Maybe some people here have experience with what it means to work with AI in gaming and especially in game dev community, where you already have to be careful not to get chased down with torches and pitchforks.

@DysTopia Yes, exactly. That kind of LTS / long-term model availability would at least be an important direction to consider, and it would not put developers like me in this exact situation. But of course I also understand that this is particularly difficult in the AI sector, with new models and approaches emerging every day.

Realistically, I have accepted, or rather I have to accept, that I may not be able to preserve this perfectly. My current plan will probably be to not throw away or regenerate my beloved (and lovingly crafted) existing card pool, but mark it as Gen1 and hopefully make players understand why there is a sudden style shift.

[DysTopia]

What I’ve been observing is that the new image model has a strong tendency to enrich and add detail to simple concepts. It’s quite good at that.

So that’s an aspect to consider.

Observations:

1. The verbosity of the model is really high. Enriches basic semantic altering the meaning.

2. The second observation is that the visual reasoning logic is not an iterative progressive growth through mode. So the outcomes of GPT 1.5 and 2 are quite different, because they are labeled different.

This is, thankfully just for personal usage, but the needs of consistency is evident,

Example of a Webtoon with ad-verbatim prompt:

GPT Image 1.5

GPT Image 2

Edit: I don’t want to send the bad message to the NASA workforce. You guys are awesome! Comic was done during a casual talk. Wondering what we would miss from the Flyby event.

[windysoliloquy]

@Bliggfang

but everything tends to look much more glossy

you can also prompt for time-worn, aged, or weathered textures.

(and much more)

to get rid of the gloss effect

[merefield]

Plus 1 for magic ducks​

(Hope you find a workable solution )

[me108]

Have you tried Llama and huggingface Muapi/gpt-image-1-style-flux?