<About The Model>

@LatteLeopard Thanks.

I’m going to frank, out of all the models I ever mixed, this is the one I hate the most. It feels bland, uninspired, sloppy, and subpar. A product of pressure and frustration, because I felt like I needed to update the model after four months of nothing. I’m deeply dissatisfied with it and I feel like I missed the mark, and the bus. If it had come out three months ago, then it would be decent but it didn’t, and there’s already dozens of models that do what this model sets out to do, but better. I spent a month testing v1.0 before releasing it, I only spent a couple days with v1.5. I didn’t even want to release it but I didn’t want people to wait another two-three more months while I made the ‘perfect’ merge. I might release a ‘better’ version, more focused on what’s it’s supposed to be, 2D, and maybe even finish working on the update to TTT, but that’s far off and after that, I’m done. With all of it. Semi-real, realism, 2D- I’m done. I’m throwing in the towel. It’s too demotivating to spend time making things no one will ever use. Nova exists, Willy’s exists, and there’s a dozen 2D models better than this one. My niche has always been realism and semi-real, and that motivation died totally when a model author bitched that I didn’t portray their Noob realism slop model in the best light.

Tl;dr: I’m done, after these next two model releases. It used to be fun, now it isn’t, and I simply don’t have the time to dedicate to this anymore.

I appreciate the support, past and present, and hope you find more use out of this model than I did. And I’m certain you can make a better furry realism model than any of the shit I posted.

@LatteLeopard Thanks for the response, and for taking the time to even look at this model.

I'm going to go on a rant about the realism side of furry models and why nothing ever happens.

The sheer lack of development and the shift in ethos in furry realism models is genuinely infuriating. MPJ and TTT was my response to Willy's and the obvious short-comings of the model. It's why I put it in F.M.M.R., because they're specifically made to help people produce better 'realism' furry models, but no one cares about what those models bring to the table because those people only care about 'muh-photorealism' and maximum resolution and detail output. If a model is incapable of doing basic things, it doesn't matter because they use Comfy and run multiple models, slap on half a dozen Loras, inpaint the fuck out of it then upscale it to 4k.

I try not to be harsh with criticism but after watching Thermal spin their tires in the mud of realism merging and falling into the same pit as most people who has ever merged furry realism models for Pony, just to restart, it's hard not to come off as overly harsh.

Many of the issues with Willy's is because of the over-focus on 'photorealism' and trying to maximize raw image resolution over basic quality, subject accuracy, and conceptual knowledge. Illustrious, and by extension Noob, wasn't trained to produce 'photorealistic' results. The solution to that was to train datasets onto the model bases or to simply merge Loras into custom finetunes. Datasets and Loras primarily comprised of people, photorealistic human people. This can work, but you have to sacrifice one of the three (quality, accuracy, knowledge) when it comes to using Loras or making a model 'photorealistic'. This is my guess as to why Willy's had tinting issues early on, because of a badly trained 'photorealism' lora that was merged in or was a part of another model that was used in the merging process.

Another issue is that the further you go into 'photorealism' the more rigid your model becomes.

I'm going to state something that should be obvious but apparently isn't for a lot of these people.

The further you stray from the content and style that the model was trained with, the more you lose.

You probably know this intrinsically. It's a balancing act most people merging realism aren't aware of. If you're merging mostly photorealistic models and Loras into a model, your model may be able to produce a photorealistic Renamon or some random anthro wolf but it won't be able to generate a whole lot of anything else without the help of a Lora- just like every photorealistic human model in existence. The float point hasn't magically gained the ability to cram more data into the same size, it's just mixing the weights. And if you dilute the weight of the furry model with more and more photorealistic human Loras/models, surprise, suddenly your furry model isn't so furry anymore. A good and fast way to judge this is the models ability to prompt for styles that isn't realistic. It's something everyone in the furry realism model merging scene fail to realize, the inherent compromise between realism, knowledge, and accuracy. It's something I realized long ago when I was merging DisgustinglyReal and some other models I unpublished and trying to get 'realistic' Pokemon to output accurately. It produced high quality images in a semi-real style, and besides Pony Realism at the time, there wasn't all that many floating about. It's why Hayden used it as a base for his furry Realism models. But the problem was it was rigid and inflexible, by the very nature of Pony. There's only ever been one PonyV6 and everything else is a derivative. Base Pony has terrible quality but high knowledge (except artist). Remerging base Pony into the realism model would destroy the model's quality but reinforce the model's output knowledge. That's the main thrust of the problem, the further from base, the more it loses. Accuracy and knowledge collapsed as quality and realism increased and improvement of this simply wasn't possible, so I abandoned Pony after Calypso. When I started merging with Illustrious, I centered my focus on accuracy and knowledge, and developed the models from the mindset of maintaining focus on those. It's why I use a Zorua as most of the main cover images for the models I post- it's basically a barometer for model accuracy. Something with a simple design that is routinely generated incorrectly. I test knowledge through the model's ability to prompt in different styles and ability to output different concepts.

I never used Willy's in any of my merges. It's not because it's a bad model, it's actually a pretty decent model. It's just not a very versatile furry model. It has difficulty doing basic things. And with such a small pool of 'realism' furry models, merging off of it won't bring anything meaningful. Thermal has already centered their focus for their model on 'maximum photorealism' and has already fallen into the same issues that's befallen every single person that's focused on maximum quality, maximum realism (the very same issues you stated), which sorta makes the model dead in the water for me. Even with the 'restart' they're doing, it's probably not going to achieve much in the long run if they do the same thing again. I could probably write a 20,000 word essay on this topic alone, but I'll cut it there.

Ethos

My ethos when it comes to model merging has always been the same since I started merging two years ago:

The best with the least.

Doesn't that sound corporate?

No upscaling, no hires fix, no adetailer, no embeddings, no loras. None of it.

Just raw gen.

If I NEED to use embeddings, loras, or have to upscale, use adetailer, or hires fix to get good outputs, it's trash. Full stop. Everything I make needs to be 1:1 reproducible on A1111, with the exact settings I provide. It needs to be able to produce quality images as often as possible with as little as possible for the people who aren't able to local gen and don't have access to the full toolset everyone else has. It's why I hate V1.5, I feel it doesn't follow that ethos and is more of a placation while I continue to merge and test. It doesn't feel substantial, it doesn't feel like an upgrade. It's actually a downgrade in some regards, which goes against the very point of an update. Everything needs to be an 'upgrade' and be constantly updated to be noticed at this point. It's something I simply don't have the time to do, I only get one day off a week and most of that day is spent crunching on getting the next version out.

I usually have to hunt down feedback for my models, and so far this model version has achieved nothing in that department. No use outside the sites, no bids, no conversation, nothing, like most of my models. The only feedback I ever got on my models is through comments on the model page, and the very rare worded review. Every thing else is heard second-hand or through searching. No one has directly told me anything, what needs to change, what could be improved, or even what people even use the model for. The only feedback I ever got was the model had 'Nova noise' on the Renamon on the beach image, and I had to look for that complaint. Very cool. My most successful resource is a group of embeddings I re-uploaded that someone else made for SD 1.5, that everyone slaps onto their images even though it doesn't do anything. Nothing I post achieves anything. No one reads anything, no one says anything, it's beyond discouraging.

As for Nova, there's not much I can say about Nova, I've never used it.

Apologies about the long-winded response, I just felt I needed to illuminate some things when it comes to furry realism models, while I'm still around to give the information, in the event you ever decide to take another crack at it. Not very technical in the slightest, but I hope you find some use from reading all of that. I'm going to quit wasting your time now, thanks again.

@LatteLeopard  It’s no issue. Black Days is a Pony model, which is SDXL. But Pony models are pretty cooked when it comes to outward mixing, you can only really successfully merge them with other Pony models. Despite the architectural similarities of SDXL 1.0 and Pony, (both are SDXL) merging them together is basically an impossible task. When it came to mixing Pony and Illustrious together, I’ve had some minor success with TTT, which has Calypso mixed in it in very small increments, but it scales very rapidly so amounts need to be very small. There was only ever really three-four SDXL 1.0 furry models ever made, Echo Alpha Conceptualizer is one of them. Illustrious, Noob, Pony- it’s all SDXL, so theoretically merging them is possible.

Black Days is still floating around, just unpublished on Civit. The links to it should still be in the F.M.M.R. description, if you wish to try it. It’s a very unstable model however. I fixed that but seeing the complete and total lackluster response, I never released the fix. Take mind of that if you wish to attempt merging it into any of your models. If you’re merging with TTT, you should adjust the input and output noise amounts to get cleaner high CFG outputs if the merge outputs look too overly detailed, splotchy, or crunchy. Baking in a VAE with each merge is also a necessity to avoid wash out with realism merging. There are many good options, I usually default to SharpSpectrum but there are other SDXL VAE’s to experiment with.

@Feem Thanks for the support, it means a lot to me!