A = Reduced Noise, Better Posing

B = Flat Coloring

Sounds like A is the better pick!

I looked down the comments and saw this:
"The result is IL v6B: extend from IL v3.0, which has more flat coloring rather than volumetric coloring
IL v6A is same as IL v5.0 but it has improved posing and reduction in noises
Both of them are good and compatible with Illustrious / NoobAI LoRAs"

@Craz3D Much obliged

@Crody yea use that, but running locally i cant find the best, schedule, i use mostly Karras for example. thank u for answering by the way.

I use Euler A - scheduler: Automatic
At 25-35 steps and CFG rather low around 4. Higher and the colors become to harsh, at least on my older v6 model, will update to v8.

For a basic node set-up: Load Checkpoint -> CLIP Text Encode (Advanced) w/ A1111 + Empty Latent Image w/ desired resolution -> KSampler (Efficient)

Any kind of other error would be from other nodes conflicting with those, so wouldn't know exactly what else is causing that error as I don't believe CLIP Text Encode should have those errors.

Thanks!

If you wanna see how big of a difference "masterpiece" makes then do try adding "(masterpiece:5)" to your promp and you should see a huge difference in shading and composition. It's more like for best results you should have them but otherwise works fine enough without.
They do still have their uses, even if not necessary. 👍

@ermtest I'm not sure that is a practical or even useful A/B test since anything pushed that hard in weight is going to have an impact on the sampling just from conditioning pressure. I'm talking more intrinsic and pragmatic value of the terms alone. I've done A/B testing dozens of times using different control prompts and I couldn't tell any obvious quality differences. At least not with this model and the size/specificity of the prompts I regularly use.

@Uncanny_Undyne At least from my testing I do see tangible differences when using those terms, just having them at a higher weights lets me see how each of them impact the images at their extremes so I can apply them normally and test for redundancy with a fixed seed. "masterpiece" does seem to have its own latent style from comparing generated with it, without it, and with other terms selectively one by one. It seems that you get the most stable and highest prompt adherence without the prompting tags but using the positive tags can get 'better results' at the cost of being less stable and lower prompt adherence.

@Uncanny_Undyne 
Here's a link to a collection of different tests of the positive prompting tags:
https://civitai.com/posts/17929758
It seems "masterpiece" and "best quality" do have notably different effects on the image generation at their extremes, just having them at a normal weight won't have that same extent of the effect but would still obviously lead it closer to those results.
So, unnecessary but useful. Better to have access to them than have nothing, in my opinion at least.

Generally between 20 to 50 steps and CFG around 6 is fine for most models. The most effective prompt enhancers are actually being more specific and detailed with what you want to see rather than most "quality enhancers" general purpose tags. The more the models are familiar and trained well, the more impact your tags and weights become. I recommend just starting as simple as possible for what you want to see and iterating on what comes out, using a fixed seed. Use both your negative and positive prompts frugally and explicitly.

@Uncanny_Undyne alrighty! I’ll be sure to use my microscopic brain to understand this! Thanks.

@Uncanny_Undyne My two cents, CFG of 6 is way too high. Generally you'll want to be at or between a CFG of 3.5 to 4.5. Anything higher then 5 tends to over weight LoRAs and images start to look stiff and less organic.

What is a feral? Different from furry or anthro?

Are you trying to create images greater than 2000 or more in size? Make sure your settings, whether you use the size increase or not, don't go over like 2400 or else generation will take longer.

I have 3070, on the 7b model, I have one image generated in 31 seconds. Euler A, 30 steps, 5 cfg, 1024x1536

@Crody well you did a bloody good job on every last one of em, Have a darn excellent day.

@Crody Oh nice, must be a difficult time crunch to do…

A couple tips could be to use 'CLIP Text Encode (Advanced)' to use A1111 weights which I believe CivitAI uses, and to use recommended supporting tags in the description of this model.

@Crody Thanks, I’m still an idiot.

@M1Carbinemann Even though now I use mostly Supermerge & untitled merger.

 Without his guide and tools(Chattiori Model Merger) I would have never made Banana Splitz XL, iffymix, Yiff in Hell and Stormtooth XL.

So besides just the nova models, he's responsible for influencing quite a bit of other models.


P.S. If you are looking to make a model. Dont start with Chatiori, its pretty advanced.

I would start with something that has a GUI like supermerger. It runs as a tab in A1111/Forge.

Also I highly recommend looking through this.
https://rentry.co/BlockMergeExplain

This explains how the unet works and how block merging works.

Its honestly very simple once you realize how it works.

I don't think so bro, what do you base that on? Did you really compare them with the others? I have to know which one is really the best😭

@Chris198 No. I just selected v8a for shits and giggles.

@Chris198 8A is my favourite since 6A. I've tried most versions.

@Crody Hello! Thanks for your reply!
But I have a few more questions, please...
I am a humble indie solo developer. I am creating a dream game that will probably never get released. Yet, I don't want to violate anyone's rights...

If I edit the images created by this model through inpainting (changing clothes, colors, etc.), are these changes sufficient?

Can I use the images created by this model to create 3D meshes and UV textures for them?

PS. Of course, I will credit the use of this model in the game's credits if it ever gets released!

@Crody Thank you so much!