Yeah, I went through some of the older versions, and it's become annoying how badly the more recent ones work with style tags or even complete style loras. 17 in particularly egregious in that regard. 15.5 was decent, but pales in comparison to for example 10.

Some of the tags plainly don't work anymore, being unrecognizable even at higher weights. Even some of my style loras plain out refuse to give recognizable results with 17 that had no issues with older iterations. Which, thank goodness, can still be downloaded.

Back to the roots, I'd say.

@FrivolousBand10 I couldn't agree more. Trying to work with 17 literally inverts everything you know, leaving you basically starting from scratch like it's your first foray with image generation.

i think its why i've been so iffy with nova for so long. im tired of all these updates for changes i hardly even notice upon testing/comparing the models...

its exactly why i use the older versions because it DOES things better than what the newer versions do. whenever i hear a new version of a model, i want to see exactly just how better it can do X and Y. but nova for some time now has kinda been the opposite of that. it's just the past model but it does things slightly better or, wow, the image doesn't look AS burnt before! blah...

also i really would like if he provided more variety in his example images, i want to see more of its capabilities outside of 'making previously inferior image better' yknow...

@gibbledrip If you check the artstyle of the example pictures, it seems to have drifted towards a more eastern, kemono-style default with more recent versions. Prompt adherence hasn't improved by much as far as I can tell, and according to some, even gotten worse with 17. I can only vouch for the style tags, they've been pretty much unusable in 17.

The annoying bit is that there is a Nova Kemono for folks who want that sort of thing.

Maybe this LoRA can help?: https://civitai.com/models/2456317/better-feral-canine-male-ixl

So NAI is a version of Illustrious 0.1 with slightly more data trained. Its outdated compared to Illustrious 2.0(has all of NAI data + more)

[Oldest] Illustrious 0.1 -> NAI -> Illustrious 2.0 -> Chenkin Noob [Newest]

This is pretty much the pipeline of trained checkpoints that add new data trained onto illustrious.

But this model uses NAI + Illustrious 2.0 merged together as a base.

So it should be very versatile and have the lattest data.

I imagine he hasnt added Chenkin Noob yet, because its really not good enough to finetune at the moment. But in the future it might be more used.

Cool story bro.

@BabyBearButtz ah yes, my points are completely invalidated now, you've won, pat yourself on the back

@Kaltingale Care to share better models then? Or any alternates you find superior? It's easy to complain like this, but at least provide something for others to take besides the sourness.

Cool story bro, care to share the superiour setup? Or you just spilling hot air?

Nova Illustrious V2 was probably one of the best models I've used, and even though it's a bit outdated now, it's still very good, and even if I struggle with the more recent ones, I come back to it from time to time because it remains a reliable go-to. So if you have better models, feel free to share, I'm curious too.

I make images focused on transformation, which is quite tricky, especially back then, here are the results I was able to get with Illustrious V2.

https://www.deviantart.com/humanitizer/art/Cow-Transformation-28-1245175646
https://www.deviantart.com/humanitizer/art/Yuzuki-Beneath-the-Moon-1194838256
https://www.deviantart.com/humanitizer/art/Mizutsune-Transformation-1-1164722119

For artist tags, NovaFurry 10 was the last version that worked properly with them. Prompt adherence is decent enough, so that has become my go-to version as of late.

Everything that came afterwards needs higher weights to get a visible effect (1.3 weight and up) and with the more recent version, also tends to work poorly with style LoRAs.

@dasrael944 I literally mentioned banana and molkeun in my post. Can't you read?

@rickybobbyjoe any of the banana models, molkeun's deepblue versions and deepnavy v4, wolfmondai. This isn't a gotcha you think this is. Maybe you wish to compare my creations to yours as well? I'll be thrilled.

@Humanitizer sadly I can't view whatever you are making, I don't have a DA account for it. My personal go to choices are MolKeun's and Banana models. They worked perfectly for everything I needed and both are super reactive to artist tags in my experience, helping make some unique styles thanks to it.

I'm sorry to be a bother but would you be able to give me a link to the banana model you talk about? There are many models like here on civit ai and I don't know which is the best.

@Plakson Here you go: 

https://civitai.red/models/1893003/banana-kiwi-xl

https://civitai.red/models/1261882/banana-splitz-xl-new-xxl

@Kaltingale In the post, you’ve mentioned nova being “banana & molkeun slapped together into an amalgamation”, which does not infer that you consider those to be good models. Thanks for elaborating

@FrivolousBand10 thank you for giving them a link for me. I don't visit CivitAI too often, so I haven't noticed their comment until now

@FrivolousBand10 in what sites you can use the earlier versions?

@jmfmagnum32959 No idea, really. I generate locally, so I just downloaded the older versions from Civitai and haven't looked back (or is that forward?) ever since.

As long as you have a somewhat decent 8GB GPU (Video Card), you can run ForgeNeo or a similar frontend and run the models locally.

Look for Stability Matrix if you need an installer to handle the dependencies.

@FrivolousBand10 forge neo, thanks

it was better, recent versions of it became disappointing

@Kaltingale you have 6 images on your profile here on civit, are those the images im comparing against? I'll post a couple quickies if thats your best. Im assuming not though. Where can I see what you got? I just posted them, not sure how long it takes to show up.

I slapped those together with some 75ish word positive and negative prompts. 2 with different style lora and one cropped and no lora. I would like to see what you have..maybe its wildly better and I can convert to your stuff.

@FrivolousBand10 Thank you. You recommend any specific versions of these models? 

@Plakson https://civitai.red/models/1261882/banana-splitz-xl-new-xxl has a nice cartoonish look I sometimes use when going for funny, but I have no special preference here. I use mostly NovaFurry10 and 15.5 these days, along with MolKeun Mix for furry stuff.

For more "realistic" stuff, WickedIllustriousMix and PerfectDeliberate9

It's a matter of taste mostly. They're all based on Illustrious and perform roughly similar. I also have Zimage Turbo and Qwen 2.8 at hand, but those can't do yiff and are heavily centered around "realistic" imagery, which is kind of the opposite of what I do.

If you want some suggestions, chat me up, this isn't exactly the place for an in-depth discussion.

I can only fully agree on that!

I can fully agree on FurorSaxonicus agreeing to kronokin.

It's defo a nightmare.

I second this. Im struggling hard to keep dark brown off my legs and arms T_T . (Love the model though and grateful for the work put in. so please dont be discouraged. Looking forward to the next update)