Just tell it to make a comic, like specify how many panels, characters and subject. I made a pretty short prompt and it made an entire comic page filled with tons of dialogue

Hey! You should check out the bottom of this post - https://civitai.com/changelog?id=168 - cool stuff coming soon!

@KuroKenzo do you have telegram or any platform that you can message me?

You would not be able to run it anyways. It is a proprieatary model obviously you can not download it, wtf are you thinking? But you can generate images with that for free, since it only costs buzz, so more than 3 pro images per day using gemini free tier and no need to buy an API key. They offer you to use a model here, they do not force you and that does not mean the stop having open source models here. And when you see nothing special about nano banana you are just blind or ignorant.

@demonkingoftranny "Interesting how you completely ignored the actual point — which was about hostile Terms of Service and the normalization of non-downloadable models on an

open-source platform. Instead you explain to me what 'proprietary' means. Thank you, I was not aware of the dictionary.

You say I 'couldn't run it anyway' — a classic attempt to attack the person instead of the argument. Whether I personally can run it is completely

irrelevant to whether CivitAI should host models you cannot download, cannot discuss, and cannot audit.

'They don't force you' — neither did the frog notice it was being boiled slowly. The point is about platform standards and what precedents are being set,

not about individual choice in a vacuum.

And 'blind or ignorant' — the last resort of someone who has no actual counter-argument. You offered zero reasoning why non-downloadable, non-auditable

models belong on a platform historically built around open weights.

I'll leave my original point standing. It seems to have made you uncomfortable enough to respond without addressing a single part of it."

@pink0909 "you can not run it anyways" is not an argumentum ad hominem, since it hardware requirements are literally that high that no one with consumer grade GPUs can run it, it was not addressed to you only but to us all, you would need like H100 or H200 idk, definitivley more than a 5090. So my point was the abillity to download this model is not relevant for the vast majority, because the model is too huge, nobody can run it.

CivitAI did not reduce the ability of users to share and download open source models for free in any kind. They just offer you additionally to use a proprieatariy model for free, since when you generate locally and upload here you get blue buzz, for which you have no other use, so you can do much more pictures for free than with gemini free tier.

For anime pictures the illustrious finetunes are still the best, but for everything else like text render, photorealism, cartoon style, prompt adherence, concept knowledge, etc. is nano banana the best i ever tried, also i think it is pretty clear the best, that is why i called you ignorant, because you called it "not special", but it is in fact a very good model.

@pink0909I support you wholeheartedly, you are damn right!

@pink0909 Why you have defend yourself, stand to your opinion, even I disagree. Just think about you spent some millions to something and gift it too the world. Free to download? Yes would you do so?

@Silmas "First of all, many major companies do exactly what you think is illogical — offering free downloads. The reason it seems illogical to you is because you're conflating 'free to download' with 'free as in open source' or 'unprofitable.' Those are completely different things.

Companies can and do profit massively from freely downloadable models. There are several well-established ways:

Cloud API monetization — Meta released Llama for free download, yet businesses pay per API call to run it on Meta's infrastructure. Most companies don't want to manage hardware — they pay for convenience.

Enterprise licensing — The base model is free, but commercial use at scale requires a paid license. Stability AI does exactly this with Stable Diffusion: free for personal use, paid tiers for commercial deployment.

Ecosystem lock-in — Google gives away TensorFlow, Microsoft gives away VS Code. Why? Because once developers build on your tools, they naturally adopt your paid cloud services (Google Cloud, Azure). The free model is a funnel.

Fine-tuning and support services — Companies charge for custom model training on client data, dedicated support, and SLAs. The open model is essentially a loss leader that opens the door.

Talent and research reputation — Releasing a strong open model attracts top researchers who want to work with cutting-edge technology. That talent pipeline is worth tens of millions in recruitment alone. Mistral AI built their entire company reputation this way.

Data collection — When users run the model and report issues or use hosted versions, that feedback loop generates invaluable training data.

So the question isn't why would you give it away — it's why would you NOT, when the download itself is the marketing, the funnel, and the community-builder all at once. Google keeping it locked to their platform is actually the less proven business strategy here, not the safer one."

@pink0909 Fair points, but you dodged the actual question about Google's model on CivitAI can't be downloaded. That was the whole discussion.

Just give me your honest, raw opinion in two or three short casual sentences. No numbered lists, no bold headers, no business frameworks — just talk to me like a normal person in a forum. What do YOU genuinely feel about it, in your own words?

@Silmas i dont know exactly what you mean i dodged the question, since i already laid out in previous replies why exactly is bad if its now downloadable. Also what i feel about it, can be estimated from that too. I think more than i feel about something, feelings cant be argued so much. I like to add arguments so the "feelings" (which are more like POVs) can be understood by others. You can see what happens when feelings are the primary drivers. My comment seems to be very controversial, and the funny sidefact is i got some Badges from Civitai (that was not my intention lol).

@pink0909 See, THIS is the reply I wanted. Real talk, your own thoughts, no polished essay. I'll be honest — your previous reply felt a bit... too structured, let's say. Like it came out of a template. But this one? This is you, and I actually get your point now. We still disagree, but I respect this way more. And congrats on the badges, that's a funny side effect of a heated discussion 😄

No.
Nothing special apart from inspiring detail, extraordinary range of styles and subjects, and a strong typographic ability. Probably the model Heineken would invent. But no, nothing special. Maybe you could try using it to improve the three images you've posted instead of berating users who have genuine points of view.
Some people.

@alulles maybe i underestimated the rise of people without own GPU to run locally (used cards with atleast 12gb start from ~160-250eur, 16gb at abtout 400). There are so many Model- families today (SD/SDXL/PonyXL/illustriousXL/Flux1+2/Chroma/Z-Image & Turbo/Qwen Image/and maybe i forgot some others) where most of them i tested and when used properly gave me the Conclusion i wrote initially.

@pink0909 you would need like a H200 cluster to run nano banana. I am totally a local user, but just because i prefer local, does not mean, that it is bad when CivitAI hosts proprieatary model via an api on their site. It is just an additional offer, not 100% of model on the site have to be local just because i prefer it, especially not models, bascially no local generation user could use, i bet the amount of people with a H200 or multiple of them is pretty low.

So, giving all the computing power to billionaires wasn't such a good idea after all.

@demonkingoftyranny demonkingoftyranny, I understood your argument the first time you mentioned it (that I can't run it locally). I didn't address it because, based on my experience with many models both large and small, I've come to the following conclusions:

1) Large commercial LLMs are not necessarily better than local ones. All the benchmarks are optimized and self-serving — companies essentially train their models to pass tests better (basically "teaching to the test"), but when it comes to real-life usage, the larger LLMs often get worse, not better. ChatGPT is a prime example: after 4.0 it improved on benchmarks but noticeably degraded in practice — more lying, more hallucinations, and lazier answers with less effort.

2) Image generation models — I was part of the evolution from the beginning. SD 1.4 / SD 1.5 launched, then the community took it to another level entirely. SDXL came out, same story — the community boosted it enormously. Then came Pony XL and Illustrious, again the community multiplied its value through fine-tunes, LoRAs, add-ons, and extensions. Then Flux, Qwen, Chroma, and so on.

I understood the jump in model size from SD 1.5 to SDXL (2 GB to 6 GB). Then a new generation arrived — Flux, Qwen, and that bunch — ballooning to 12–16 GB, which is too much for my taste. Thankfully, Q8 quantizations brought them back down to earth. 😉

So my questions to you:

What exactly can Nano Banana do that all the other models I mentioned — which run locally — cannot?

And if there truly is a meaningful difference (which I haven't seen while browsing the gallery extensively), does the VRAM requirement justify what might amount to a 0.0001% possible improvement you'd be hard-pressed to point to?

3) Video models — this is the one area where the gap between commercial and locally-run models is genuinely starting to close, thanks to Wan and LTX.

And not shift image by 1-2-3 pixels every time you edit :(

How do you work around degradation? I sometimes just erase parts with very soft eraser so edges between two layers are not visible, thus restoring higher quality parts where i need them from previous gens.

@Lostcut When I have it edit it within a Gemini or AI studio thread where it was made I ask it to not make any other changes or move anything after my prompt and its like 50 / 50 whether it moves it or not, so I usually roll until I get it. When uploading the image into a new thread the degradation and movement problem seems worse for some reason (More like 10/90 in that case). Worst case scenario, I do some very careful resizing using transparency to line them up in paint so I can erase them in.

Really, that's happening with ChatGPT Images now where each edit makes the artifacts worse

Ask Google

They have some kind of weird integration with Gemini in it. Probably wouldn't work without it.

@Meta_Void So it'll be just as limited as is through Gemini?

@bernlarsen616 The only difference using this through Gemini directly is that it will be a little bit more strict about the rules and you will have a watermark.

@Meta_Void I found a page to remove the watermark from gemini with no quality loss. How would I use it here? I don't think I've ever used the civit generator.

@bernlarsen616 Well if you buy some buzz on here you just press the create button on this page.

@bernlarsen616 I forgot to mention; Gemini is also throttled so it might take awhile to gen based on how many people are using it. If you use a paid service it will gen a lot faster.

bro lol

it was only for generation. there is no downloadable model.

do you have a H200 cluster to run it?

@demonkingoftyranny It would nowhere near need an H200 cluster. This could easily fit on one Blackwell.

@WolfpackJac3 no it will not fit on any "gaming blackwell", so on none of the RTX 50XX cards, but sure the enterprise blackwell B100, B200 and B300, are even better than a H200. But sure you might not need a whole cluster.

@demonkingoftyranny There is no such thing as a "gaming" Blackwell. Proof you dont know what you are talking about. It WILL fit on an Nvidia RTX Pro 6000 Blackwell. No, It does not need a B100. This model is NOT over 90gb.

@WolfpackJac3 Nvidia 50XX generation has Blackwell architechture. It is just the successor of ada lovelace. But sure i do not know what i am talking about while you mention a whole gpu architechture which even the 5050 is part of. And Nano banana is fully proprietary you can not know how large it is. Considering that it is better than Flux2 max which is already over 90GB nano banana2 will probably be significantly larger than 100GB, but who knows what google has done, it is proprietary. point is it is so large that no one of us can run it, so a download would be useless anyways.

@demonkingoftyranny "has blackwell architecture"... do you really think that ANYONE talking about a 5090 is going to call it a Blackwell card when there are literally BLACKWELL CARDS. Nonsense argument. Plus quantization allows Flux.2 FP16 to run on anything more than 50 gb. "None of us can use it" Also nonsense. There are plenty of people with engineering rigs and CAD workstations that could absolutely run ALL of these models. So no, a download would NOT be useless.

@WolfpackJac3 Blackwell is the name of the architechture, when someone says blackwell i think of this gpu generation. I actually never talked with anyone before who says blackwell but does not mean the blackwell architechture but only some cards of that architechture where they also put it in the name XD. It does not matter that nvida puts the name of their new architechture in the name of some cards, does not make them more blackwell than a 5090. Again you can not know how big nano banana, also nano banana is significantly better than Flux.2 Max, so i guess it is also much larger. "plenty" is a strong word for i small minority. 4090/5090 users are already not much, but people with enterprise GPGPUs are very very few. Why would i care about them? Would be still "generation only" for most people even when it could be downloaded, makes nothing better for them when a extremely small group of people can use it locally.

@demonkingoftyranny I'm done discussing this with you. My point still stands that nobody has described a 5090 as a Blackwell card when they're talking conversationally. It's just a stupid thing to say. When I say a Blackwell card, I mean the literal card named Blackwell. And you thinking better means more gigabytes is also a fallacy. That is not always the case.

@WolfpackJac3 no your point does not stand. i literally said that you are the first one i ever talked to who means only these blackwell cards where they decided to put the architechture name also in the card name XD. Everybody i have talked to has described every 50XX as a blackwell card since they literally are. Sure it does not have to be bigger. it is a proprietary model it can be anything. But i think it is very likely that it is bigger than Flux.2 Max.

@demonkingoftyranny dude... shut up

It would make for good whale vore images. I recommend trying it out.

So good they deleted their account.

@kftiger They were so excited that they transcended existence