ye, add them via the additional resources in the prompt menu

@BLESDBELO I'm asking about how good and accurate the results will be if used with the checkpoint and PonyXL LoRAs at the same time. /nm /npa

@urbanlegendwiki You can, its picky at times and deppends on the lora, sometimes you can get very good level of accurate, but there some other loras doesn't work at all,
Its like a roulette.

Firstly, use whatever works for you (unless you can't get something particular to work right that I can probably really help with).

I am happy to share my general approach to prompting though which - at risk of sounding arrogant, though I'm sure @LatteLeopard would attest - I am quite proficient at.

I virtually always use the following negative prompt, occasionally I will append some tags to it:

lowres, worst quality, low quality, large head, extra digits, bad eyes, monochrome, bad hands, bad feet, fewer digits, extra digits, deformed, mutant, signature, artist name, ugly face, mutated hands, blurry face, blurry, fat, chubby, jpeg_artifacts, gif_artifacts, grainy, wip, uncolored, censored, scan, action figure, miniature

Remove fat, chubby if that's your cup of tea.

As for positive, I adhere to a less is more approach. For instance if I want a male anthro rabbit sitting in a sauna with a towel on:

furry, male focus, solo, rabbit, brown fur, green eyes, nude, towel around waist, sitting, sauna

If this works well for the image I had in mind, I will append tags to refine it towards the final product I'm looking for... Say, you can see his dick a bit, he's confident, and glancing at you:

furry, male focus, solo, rabbit, brown fur, green eyes, sideways glance, smug, nude, towel around waist, penis peek, sitting, sauna

Take note of how I chose to place the additional tags; placing them close to other tags that have the same general contextual relationship.

Location tags (e.g. sauna, forest, bedroom, indoors, outdoors) always go after character tags in my prompts (unless they are the main focus on my image) - but that tends to not matter too much in practice.

Lastly, I always, place whatever default set of quality tags I am using currently at the end of my prompt:

furry, male focus, solo, rabbit, brown fur, green eyes, sideways glance, smug, nude, towel around waist, penis peek, sitting, sauna, masterpiece, best quality, highres, absurdres, newest, year 2024, aesthetic, very aesthetic, ultra-high resolution, 32K UHD

Hopefully this is informative. Feel free to ask follow-ups.

Edit: I wanted to clarify that this all evolves over time for me, isn't totally implausible that I could have given you a different answer next month. This is a very fluid hobby for sure.

You can see that in every gen of mine i follow the same prompt structure

-Quality tags-

-Character tags + personality and expresions-

-How body looks, shape, etc-

-Clothes-

-Scenery, setting, place-

-Angles and illumination-

They always gaves me the best outputs x3 ❤️

CFG 4.0, rarely 5.0

Steps 25 or 30

Euler A

And higher resolution before scaling (though this is only for local comfyui)

Enjoy sweetie! ❤️

My prompting is all over the place, because different prompt data sets tend to work with different models! Feel free to test out anything in https://civitai.com/collections/12250434 . Pretty much all I stick to is using Euler A and like, 4-5 CFG scale. Also, adding artist names to the prompt (like goonie-san, doxy, onta, meesh, etc) changes the final image HUGELY.

@r12x5s, @LatteLeopard, @SXCD2719, @GayPurpleSkunk, Yes, these are quite the answers I wanted. Furthermore, probably my final question, what would be the most optimal things to put in negative prompts? Here’s the negative prompt I’m working with now; (which I luckily grazed down a hundred or so characters)

“lazyneg, (human:1.2), (lowres:1.2), (worst quality, low quality:1.4), worst aesthetic, displeasing, downscaled, compression, emanata, text, words, signature, watermark, username, logo, twitter, patreon, old, early, simple background, picture frame, flattened image, irrelevant objects, (bad anatomy:1.4), malformed, mutant, ugly, extra digits, fewer digits, glowing eyes, blurry face, pale skin, pale body, blurry, bad hands, bad feet, tail between legs, multiple limbs, ugly fingers, fused fingers, large head, plunging neckline, neck fluff, pregnant, (muscular, abs:1.2) [delete if making skinny character] multiple navels, leaning forward, bowing, glistening body, sketch, high saturation, high contrast, toony, chibi, anime, thick outline, graphite, abstract, monochrome, particles, glitch, artistic error, jpeg_artifacts, gif_artifacts, grainy, wip, uncolored, censored, censorship, scan, out of focus !!, ??”

I’ve seen things like “Conjoined” (which I assume is for multiple characters?)

Bodily (neg) prompting is the worst, I really can’t find anything that fits.

Also, I’ve seen what @SXCD2719 is doing with “(thickline work, 2d, flat coloring, flat shadows):1.4,” Which I also assume leads to more of a 3d design, which is what I’m sort of aiming for, whilst not being full on realistic. A style somewhat like the works of aforementioned @SXCD2719, and hell, even your style as well. As well as @GayPurpleSkunk.

Thank you all for your time and help. :)

@M1Carbinemann You'll find many mixed opinions on negative prompts...

As SDXL models have developed over time the need for them has greatly diminished, and a lot of the issues you see like "conjoined" characters are commonly the result of poorly constructed positive prompts.

Frankly, with a well trained model like BSv6, you could probably get by without one at all; and doing so might help you gain a better understanding of how your positive prompting influences an image.

As for optimal, I personally recommend against emphasizing tags in negative prompts, and sticking to a minimal set of tags that consistently appear to work to your benefit.

@r12x5s Yes, I see. I assume that would mean weighting negative tags wouldn’t be necessary? Furthermore, should negative prompting be structured like positive?

For example, having basic terms at first (low quality, low res, text, wording, watermark, etc) then having bodily (disfigured [something] then everything else [chibi, anime, high contrast]

Also, I really indubitably thank and appreciate you and the other genners for helping a newbie like me.

@M1Carbinemann one thing i've learned through try and time, is that the negative prompt must keep simple and short

Neg prompt is insanely powerful, the less you put there, more freedom you gave. Ngl, my negative prompt is literally what @LatteLeopard used back on the V2 of this checkpoint, but with a few tags added for personal taste

About "(thickline work, 2d, flat coloring, flat shadows):1.4" line isn't really for 3D outputs, it helps in 3D loras or using 3d tag with only checkpoint ofc, but i use mostly for that "full render" look of professional furry art, you know, gorgeous shadows, illumination, colors, etc.

Best wishes! ❤️

@M1Carbinemann I am using the term emphasizing to mean weighting, so, yes; weighting tags in negative prompts should not be necessary.

Generally speaking, the model should prioritize tags that come earlier in a prompt, over tags that come later. That same general principal will apply to both positive and negative prompts. But as models have advanced this typically matters more in theory than in practice.

Edit: Appending to @SXCD2719's comments a bit...

Think of your negative prompt like a counterweight of sorts. The positive prompt will push the scales in one direction, while the negative will attempt to rebalance them some. Too much counterweight, and you break the scale.

@r12x5s Yes, thank you, I’ll post some gens using information given, Ergo, a reformed negative prompt. Furthermore, what would be nessacary tags in negative prompt? (Last question I’ll ask seriously) :)

@M1Carbinemann As I said suggested earlier... With a quality model like BSv6, there aren't really any necessary negative tags.

That being said, these probably won't hurt to have:

lowres, worst quality, low quality, extra digits, bad eyes, bad hands, bad feet, fewer digits, extra digits, deformed, signature, artist name, grainy, wip, uncolored, censored, scan

@r12x5s Yes, thank you once more! If I’d ever need help, I’d be sure to go to you! (You can check my images with the new prompting if you please)

It takes a really really really long time for civitai servers to do this, though. I need to test on my PC soon to note how much longer it takes there. Sometimes when I do this on civitai it exceeds the generation time limit and refunds automatically lol

Are you talking about the lora trainer? If so, I would ignore estimates.

Unless something has changed significantly since I last trained a lora onsite, as long as you aren't training something massive, or have really extreme settings... It should take ~1-2 hours on a good day.

@r12x5s Ah, foolish me, I meant a base model itself.

@M1Carbinemann I'm not entirely sure I follow you. A "base model" would be a new Checkpoint.

To the very best of my knowledge, it is not possible to train a new Checkpoint on Civit, nor is it possible to create a new Checkpoint merge like Banana Splitz.

Perhaps you could elaborate some on what exactly what you're trying to train (i.e. what is the intended use case?)

@r12x5s @LatteLeopard Yes, foolish me again, I DID mean a merge.

@M1Carbinemann Yeah, no way to do a merge on site.

Merges are on a very basic level, are not totally dissimilar to you opening a text file, changing some text, and then saving it.

The actual file changes happening when merging models is much more involved than the text file analogy, but the general concept of merely modifying existing data holds true. Due to that, the execution/processing time of a merge workflow can be lightning fast with a simple workflow.

ComfyUI will probably be the most commonplace tool you'll find used for SDXL merges, and getting started is relatively simple (though if you've never used Comfy, you'll have an initial learning curve that can be somewhat intimidating).

I've pulled some references for you to take a look at if you're interested:

1) https://roblaughter.medium.com/a-first-timers-guide-to-cooking-up-a-fine-tuned-sdxl-model-446652d018f4

-- A sort of primer on model merging. It is more surface level than a true technical guide, but I think it will help situate your understanding and maybe give you something to build from.

2) https://civitai.com/articles/3897/merging-4-models-in-comfyui-easy-to-understand-workflow

-- This is a very basic merge workflow that merges two base models, and applies some loras. Please bear in mind that merge workflows can be highly complicated, but they don't have to be for great outcomes.

Otherwise, search articles on Civit, you will likely find many that approach the general concept of model merges.

@r12x5s Eh, what I'm trying to do is upload a model (I didn't make it, but it was deleted) I assume it still won't work lol. But I will take this into consideration. Thanks.

@M1Carbinemann I'd need more details to assist with uploading (like a particular issue you are facing), but as long as you have suitable files for a model you can technically upload any model you want (though you'd need to consider the Terms of Service as to if it would be allowed).

See: https://education.civitai.com/using-civitai-uploading-content/#the-basics

@r12x5s No particular issue. Just the wait time, Last I saw, It said 18 or so hours, if that's commonplace then I get it.

@r12x5s @LatteLeopard. So sorry for bothering you guys again, but just one more thing-

The model (clarified, A merge) I'm trying to upload is bloody 6.6 GB's (Probably not that big lol). And the common upload waiting time is around 8 or so hours. Is there a difference between pruned and full? (I am currently doing pruned) And what is this "Floating Point"? (Doing 16 since that's what seems to be the standard) I have no issue with waiting such a long time, but is this the standard? Thank you, and once more my apologies for speaking on this near week old comment, instead of using the chat feature.

@TheGarandMann I'd need to know the exact name of the model you're referring to really answer this.

In regards to pruned vs full. Pruned is essentially an optimized model, there are instances where that can mean it yields less quality images than the full model; but in a lot of cases this is negligible.

FP (Floating Point) is the a maths thing, and can get very complex in its actual influence on a particular model.

FP16 should technically be faster than FP32, at the cost of some quality. But similar to pruned/full, that can be a negligible difference.

~8 Hours upload time for a 6.6GB file on Civit is not unexpected, uploads on this site can sometimes be very sluggish.

It's pretty much impossible to get two characters to do exactly what you want on any model - the only things that have worked for me are trial and error with multiple versions of the same prompt, or inpainting on my PC. Also, adding specific tags can help put two characters in the image; for a gay scene, I might add 'duo, gay, 2boys' to my prompt. Specific poses also help. Often when I put something like 'anal sex' the POV only shows most of one character's body, but adding 'hug' puts the whole body in the frame.

@GayPurpleSkunk Is correct, there's never a real separation between how a model will infer individual tags; so if you prompt pikachu and charmander in the same prompt for example, those tags may bleed into one another during generation - which is further compounded by any additional tags you add to the prompt.

There are ways to optimize the chances however.

- Using a lora for Character A, and a character the model was trained on for Character B tends to have some reliability to it.

- Prompting characters who would commonly appear in a model's training images together works consistently well. ash ketchum and pikachu for example.

- Leveraging BREAK can help to mitigate bleed, but it's only so effective, I personally follow the following format for this:

Character A tags, including individual action tags

BREAK

Character B tags, including individual action tags

BREAK

common tags; shared action tags; setting tags; quality tags; etc

Pardon?

@TheGarandMann Is he wrong though

@Maxiimus Nah

@LatteLeopard thank you!

@LatteLeopard nice! Thx Latte

@LatteLeopard guess I gotta go back to using Nova Furry XL until next week. Damn. Looking forward to when you get the new model up

@LatteLeopard Stellar! Looking forward to it. Thanks for being so active.

@LatteLeopard Give yourself as much time as you need! <3 These are great models

@LatteLeopard good to know i just hope it stays somewhat cosisent on what i can gen, since most lora updates tend to throw it out of whack, lol.