First things first: Yes, the Daily Challenge or rather the Challenge system has problems.

I want to take the general AI Judge topic out of this right away. That is not part of this post or the proposed solutions, at least not in its entirety.

Why? The Challenges are designed as AI Challenges. If I personally do not want to participate in any event that is judged purely by AI then I join one of the Bounties, Community Events or other CivitAI events instead.

Challenges are so large and happen so often, in the case of the Daily Challenge literally DAILY, that in my opinion it is impossible to permanently include a human component here.

Especially because in the case of the Daily Challenge we are given 9,000 Yellow Buzz every single day. That does not even include the Blue Buzz. That is 270,000 Yellow Buzz per month distributed freely into the community. Roughly $270 in value.

Just putting that out there.

If CivitAI had to put a human behind this and presumably pay that person in some form then the value of this gift would probably multiply very quickly. That person would likely still be pretty badly paid if they had to go through thousands of images every day.

Yes, the AI Judge is not perfect and it definitely has a bias problem right now. I will come back later to how that could at least be reduced.

I only wanted to use this introduction to share my opinion about expectations. We are given 9,000 Yellow Buzz plus Blue Buzz every day and the system already prevents the same few people from permanently farming the winner rewards. There is a 7 day winner cooldown.

And no, I am not naive about this. I am very aware that this "gift" mainly benefits the platform itself. 😉

The Actual Problems ⚠️

1. The Bot Problem 🤖

Recently and especially since the release of Community Challenges there has been a significant new wave of bots and very questionable accounts from what I have seen. Presumably one of the reasons is to get around exactly that 7 day winner cooldown.

Using multiple accounts is obviously not allowed. But rules are only as good as their enforceability.

If I can simply bypass the cooldown by creating additional accounts and participating again with those accounts then the rule technically works but only partially solves the actual problem.

2. CopyCats 🧩

Users see an image that has been highly upvoted and that image or its concept gets copied.

Remixing itself has obviously always existed. In the past I at least had some control over it as the poster. If I hid my prompt then the option for direct Remix was hidden as well. That is still the case today.

The problem is that I can now simply select Edit Image or Animate anyway. And that is exactly the point I want to make.

Even though the poster hides their prompt and therefore does not enable the classic Remix option the image can still be transferred directly into the generator as a reference with one click. I do not need the prompt and I do not need the classic Remix function. I click Edit Image and the image goes directly into the generator as a reference.

With img2img I can very quickly create something extremely close to an existing image. The image is directly used as the starting point for a new generation even though the poster has essentially signaled through Hide Prompt that they do not want to enable direct Remix.

And I am not only talking about perfect 1 to 1 copies here. This is also about copying the concept.

An image works. People see it and notice that the Judge likes it.

Edit Image. The image goes directly into the generator as a reference. A few changes. Next submission.

And suddenly the same concept exists five, ten or twenty times.

3. Bypassing Model Requirements 🔧

And this is very important again:

The problem is NOT that CivitAI does not detect the model being used. Model detection exists and it existed before.

The actual problem is that the system can be manipulated.

I can first create an image through txt2img that matches my own expectations completely independently of whichever LoRA or Base Model the Challenge requires. After that I can run the image through img2img using the correct ecosystem and the correct LoRA.

At that point the required resource has actually been used and can also be correctly detected. But that does not mean that the actual image was really created with that LoRA.

If the original image was already 95 percent finished and I only run it through the correct workflow afterward then the technical requirement has been met. The creative requirement has not really been met.

With externally generated images this becomes even more problematic. As long as external images are allowed and their origin can only be judged based on the information available at the time of upload the system remains open to manipulation.

That is exactly why completely absurd combinations can appear where a Pony Daily Challenge LoRA was supposedly generated with a Krea 2 checkpoint for example.

The problem is not detection.

The problem is proving the origin.

4. The Same Concepts Keep Winning 🔁

Okay. Yes. This is an AI problem.

But not only.

It is AI plus problem number 2.

First, the AI Judge has a bias in certain directions. Second, people notice that and copy the concept.

The result is that the winners or the images near the top keep using the same or at least very similar concepts.

Imagine announcing an event where 95 percent of all submissions are cat pictures. The chance that a cat wins is:

VERY VERY HIGH!

That does not automatically mean the Judge is manipulated. But if everyone learns what the Judge likes and optimizes for it then the participants themselves keep reinforcing its bias.

None of these problems are really terrible on their own. The problem is that they all come together and show that this system is simply not protected well enough in several areas.

And then there are people's expectations:

HEY, THERE IS FREE STUFF! I WANT IT! I DESERVE IT!

And that gets mixed together with the actual problems:

HEY, THIS IS TOTALLY RIGGED AND BADLY MADE! HOW AM I SUPPOSED TO GET MY GIFTS?! FIX THIS SHIT!

Again, I am not trying to defend the system itself or its problems here. I just want to offer a little food for thought.

At this point the Challenges do not simply put money directly into CivitAI's pocket. Viewed in isolation rewards are distributed and infrastructure is used.

Across the platform as a whole they keep it incredibly alive. They drive generations, posts, model usage and people coming back every day.

And participation is huge. According to the Daily Challenge Dashboard there are regularly more than 2,000 submissions per day. Except of course when there was a massive platform problem.

So purely based on the numbers:

Why should they change anything?

Just asking the deliberately provocative question.

People keep participating.

But I am still completely with you on this. For fairness, for fun and especially for the USEFUL application of what is from a business perspective a brilliant tool there need to be changes.

But simply screaming RIGGED, UNFAIR, SCAM, I WANT THIS, CIVBOT IS SHIT does not get us anywhere.

So I am going to try this with a few ideas and hope I do not get completely torn apart for venting a little before people at least try to look at the issue from several sides. 🙂

The Good News ✅

In my opinion every one of these problems can at least be massively reduced. Sometimes that requires more development work and sometimes relatively little.

I will go through them one by one.

In a few days when I have properly collected the relevant parts of the code and the proposals on a code level I will also put them into a GitHub Discussion. Your upvotes and comments will be especially important there.

The more attention a discussion about a new feature gets the more likely it is that someone will deal with it. Not by fragmenting the discussion across 1,000 different places but by collecting it in one place where everyone can upvote the same things, raise objections and add their own thoughts.

1. The Bot Problem 🤖

The solution in my opinion is very simple.

Why are accounts that are only one day old even allowed to participate in Challenges? That practically invites botting.

Yes, using multiple accounts to abuse the system is not allowed. But rules are only as good as their enforceability.

That is why I think some kind of restriction or participation requirement would make sense here.

CivitAI already has various account information, moderation signals, strikes and a score or trust information visible to the user. Why not use those things?

Similar to Reddit Karma participation in Challenges, Events and Bounties could require certain minimum requirements.

That would not stop a new account from posting images. It could gain followers, comment, use models and build up an account normally.

It would only prevent an account from immediately starting to farm daily rewards five minutes after being created. And with Blue Buzz this would have another effect as well.

It would become significantly less attractive to create large numbers of accounts just to flood the Dailies, collect participation rewards and dilute the system with duplicates or similar images.

Of course this would not completely prevent botting either. But every additional step increases the effort and eventually the cost benefit ratio no longer makes sense.

2. CopyCats 🧩

In my opinion this has increased enormously since Edit Image and Animate became available in their current form.

Remixing itself has always existed. In the past I could hide my prompt as the poster and disable direct Remix at the same time. Hide Prompt still hides the classic Remix option today.

The problem is that I can still select Edit Image anyway. That directly transfers the image into the generator as a reference.

And that is the crucial point for me.

By using Hide Prompt the poster is basically saying: I do not want to share my prompt and I do not want this image to simply be reused through the Remix function.

The classic Remix option is then hidden. But right next to it Edit Image and Animate now provide other ways to transfer that exact image back into the generator with one click.

A large part of the original protective effect has therefore been lost.

I do not need to download the image. I do not need to take a screenshot. I do not need the prompt and I do not need to build my own complicated workflow.

I click Edit Image and the image is directly available as a reference in the generator.

Especially in Challenges this obviously makes CopyCats extremely easy. An image gets highly upvoted. Someone sees that this concept clearly works. They click Edit Image, make a few changes and submit it again.

Again, this is not exclusively about perfect 1 to 1 copies. The much bigger problem in my opinion is how extremely easy it has become to copy a successful concept.

The system with Swiss Judge and Ladder Judge is already known by now. I have already written my own articles about that and therefore do not want to explain again here how Swiss or Ladder work.

But in my opinion this direct comparison structure could be a good basis for additionally penalizing CopyCats.

Important: Swiss or Ladder obviously do not detect copies by themselves. That would require an additional similarity check.

If an image in a direct comparison uses practically the same concept as an already existing image it could receive a negative point for that. Older before newer.

If the new image is actually better executed, cleaner, has fewer artifacts or develops the concept creatively then it can still gain positive points in other areas. The penalty for strong similarity would remain.

That would make an almost identical copy significantly less worthwhile.

For true 1 to 1 copies the system could go even further. A normal file hash only works reliably if the exact same file data is being used. As soon as an image is saved again, encoded again or minimally changed the hash can change.

For visually identical or almost identical images you would therefore need a perceptual hash or another form of visual similarity comparison.

That could then be used to say for example: The first image is judged normally. A later practically identical copy remains visible as a submission but is excluded from winning or never enters the actual judging process in the first place.

The most important thing in my opinion is that the poster gets this control back.

If I enable Hide Prompt or select an additional option for it I want to be able to decide whether my image can be used directly for Remix, Edit Image or Animate.

Of course this does not prevent copies completely. Someone can still download the image, take a screenshot or use other methods.

But that is not the point.

CivitAI does not need to offer the easiest possible way to copy something with a single click if the poster explicitly does not want their image to be directly reused.

It adds one more step someone has to take in order to cheat. And every extra step can be one step too many for someone.

It is the same as with software in general: Nothing is unhackable. But the harder you make it for an attacker the less attractive the target becomes.

Cost benefit ratio.

3. Bypassing Model Requirements 🔒

Oof.

This is actually completely solvable. But it definitely will not be popular with everyone.

And again: We do not need better model detection here. The model is already detected.

We need provenance that cannot be manipulated.

And in my opinion there is only one way to truly guarantee that:

Challenge images must be created exclusively through the CivitAI Generator AND posted or submitted to the Challenge directly from that Generator.

Everything else remains open to manipulation.

In my opinion this requires several clear rules.

First: img2img is forbidden for Daily Challenge submissions. Such images are not accepted in the first place or are rejected.

As long as I can use any arbitrary starting image nobody can guarantee how much influence the Challenge LoRA actually had on the image.

Second: Externally generated images are forbidden for Daily Challenges.

A lot of people will hate that. I understand why. But this is the only way to ensure that everyone participates under the same technical conditions and that manipulation of external metadata no longer plays a role.

Third: Even an image that was originally generated through the CivitAI Generator cannot first be downloaded and then manually uploaded again.

It has to be posted directly from the Generator and remain connected to the server side generation job that CivitAI already knows.

Only then does CivitAI have complete control over the full origin.

The server then knows: This image was generated here, at this time, with this workflow, with this model and with this LoRA. And this exact generated image is the image that is now being submitted.

Then there is no need to rely on whatever information might exist afterward in some metadata.

Everything else can be manipulated.

Of course this also creates another side effect for CivitAI: Pure Blue farming using ten locally generated images would no longer be possible.

But there is obviously also a downside or an additional component that needs to be considered:

Generation costs versus Buzz rewards.

It still has to be worth it for users.

One possibility would be for models or LoRAs that are part of the Daily Challenge on that day to receive lower Generator prices. Maybe even a small number of discounted generations specifically for the Challenge.

That could also be automated but it would definitely require more testing and development time.

Together these three solutions would in my opinion already massively reduce botting, CopyCatting and cheating.

4. The Part That Will Probably Cause the Most Work and Controversy 🧠

CivBot and CivChan.

The bias is real. There is no point pretending otherwise.

GPT 5 nano is currently used for Daily Judging. But even if another model were running there tomorrow some kind of bias would always exist.

In my opinion the problem is also not just the model itself. The rules, categories and weightings that the Judge receives every day matter just as much.

If I keep giving a Judge the same or very similar rules then naturally similar preferences will keep emerging. And as soon as users recognize those preferences they will optimize for them.

That takes us directly back to problem number 4 from above.

That is why I would not simply replace the model. I would make the judging itself more dynamic.

I see two layers for that and in the long run I would use both.

First Layer 🎲

A random list. Hidden and unknown of course.

Certain motifs, subjects or objects receive a penalty for that day. That would mean there are different terms or motifs every day that CivBot does not particularly want to see.

That alone would already create some dynamics.

If the list is large enough and the random selection is designed intelligently with cooldowns for certain terms for example then every Daily becomes a little less predictable.

Then nobody needs to enter the Challenge every day with the same Toilet Vader or another concept that has already succeeded a hundred times.

Of course those things will eventually come back. Just not every day.

I know that a hidden list would also be controversial. People could obviously say that they do not know why their motif was judged worse on that particular day.

That is why I would not use something like this as a complete exclusion. I would use it more as an additional light weighting against motifs that have dominated too strongly recently.

Second Layer 🧩

In my opinion this mainly makes sense if we have a better system for new Daily LoRAs.

Specifically by connecting the Community Challenge system and the Daily Challenge system more closely.

Nobody knows a LoRA better than its creator. At least that should be true with a good LoRA. 😉

That is why it makes the most sense to me for certain key points of a Challenge to come directly from the LoRA creator.

That could easily be added to the Model Card or the corresponding creator settings.

The creator could say:

Yes, my LoRA may be used for Daily Challenges.

If I enable that as a creator I am then required to provide certain key points.

What aspects of my LoRA are important? What should the Judge pay attention to? What is the actual purpose of this LoRA? Which criteria should be weighted more heavily or less heavily?

That information could then influence the Challenge criteria or the Judge prompt.

And this is exactly where a truly dynamic Daily Challenge system could develop.

Why should I insist on humor for a pure aesthetics LoRA? Why should a Meme LoRA be judged using exactly the same numbers as a Character LoRA? Why should a Style LoRA have the same priorities as a Concept LoRA?

Every LoRA has its own focus.

That is why fixed standard judging numbers for every Daily LoRA make no sense to me. The weightings need to be individually matched to the model.

And who knows that focus better than the model creator?

The fact that the Challenge infrastructure already supports different categories and weightings is perfect for this. That existing part would only need to be connected in a useful way with the model creators and the selection process for Daily LoRAs.

I still need to think more carefully about what this would look like in detail, how it could be protected and how the system might even automatically prevent broken or badly functioning LoRAs from being selected for a Daily Challenge.

But I still believe this is the right direction.

What About the Judging Itself? 💸

There is another point that often gets forgotten in this entire discussion:

Cost.

From my own experience I can say yes, a fully featured good AI Judge for images is possible.

BUT:

Building one properly requires a huge amount of time, development and continuous computing power.

In my own tests for a fair AI judging system I burned through almost $250 in two months. If my system from back then had been used permanently the judging of around 50 images would have cost about $6.

My dashboard has regularly shown around 600 judgings for 2,000+ Daily Challenge submissions per day for months.

If I simply extrapolated my old system to 600 judgings that would be around $72 per day.

Just for judging.

The current system using GPT 5 nano is in a completely different league in terms of cost.

Based on my own research and calculations one call is roughly around $0.0001215. That would be around $0.12 per 1,000 judgings or not even $0.08 for 600 judgings.

Of course this is only an estimate. The actual amount can vary depending on resolution, depth of analysis, actual token count, contract type and technical implementation.

I do not know CivitAI's internal terms.

But that is not really the point.

The point is the difference in scale.

On one side there is a very cheap automated system.

On the other side there is a significantly more complex Judge that can quickly cost many times more.

That is why I also think "Just use a better Judge" is far too simple as an answer to the problem.

Of course you could build a much larger and more complex judging system.

The only question is whether the additional quality justifies the additional price.

That is exactly why I would first try to use the existing system more intelligently.

More dynamics.

Individual criteria matched to each LoRA.

Creator input.

Swiss and Ladder.

Additional weighting against constantly recurring concepts.

And the other measures against botting, CopyCats and manipulated submissions described above.

It can always be refined afterward.

But in my opinion these changes alone would already make a massive difference without having to replace the entire judging system with something that suddenly costs a hundred or a thousand times as much.

Conclusion 📝

As I said these are my general opinions about the Daily Challenge and some more detailed thoughts on how the problems could be solved or at least significantly reduced.

Everyone will have a different opinion on this and everyone is just as free as I am to write an article about what bothers them and how they think it could be done differently.

I do not expect general agreement or disagreement here. These are simply my thoughts on the topic.

I will post an update here once I have opened the feedback Discussion on GitHub. That will also include some more detailed findings about the code and implementation.

Anyone who likes my ideas or feels exactly the opposite is encouraged to leave their opinion or simply their vote there on GitHub.

Because only when feedback is collected in one place and many people react to the same ideas there do you get the attention needed to potentially trigger a change.

Do not fragment the same topic across 1,000 different places.

Collect it in one place. Discuss it. Upvote it. Raise objections. Improve it.

If you made it this far thank you very much for your attention. 🙏

Happy generating and posting.

All the best 💕

-- Moonbear 🐻

One Last Disclaimer 😄

Before a bunch of people DM me again to ask... no, I am not a CivitAI employee. I definitely would have announced that very loudly by now. ^^