V3.0:

2 months of work, and I think 3.0 is finally ready for showtime:

A completely new redesign built using different ideas, using a primarily human only model and applying 'patches' as it were for doing what is essentially a furry conversion yielding much improved detail.

Tips below still apply, and some tags might spawn humans where my tip would be to try a different word with the same meaning, say a booru/e621 tag of said subject.

Or try the furryfication slider lora. But that is best used with solo subjects.

Of note: the bigger version should help with fur and background detail and such, if you can it is recommended to use it instead of the smaller float16 converted one.

Many excel sheets, Visio documents, 6 gigabytes of test images, possible 300 or more GB of test models made over the span of the 2 to 3 months.

And yes it is still tagged as a Pony model even if I am pretty sure it has become so diluted at this point it might not respond nowhere near the strength it used to. But hey, try it anyway.

If you end up getting too many humans, check the length of your negative prompt or other non-booru tags that might give problems if the underlying base model were to respond too much to it. Recommending to keep thing short and simple.

Or try v2.2 and see if it works better for you if you are not chasing the 'best quality' as I am :P

Models used:

• Base: Brehmsstrahlungmix 1.0

• Furry additives: Furry Enhancer 6.1 + OneForAll Pony Reality + Lucid Citrus Pony Kiwi + ???? Master ???? (a rule34 dataset, got removed from CivitAI) + Quintillus + RealDream 14

• Negatives: Natural Vision 2.7

Results so far:

• Better lighting

• Background detail and items, say at an office a whole computer setup appears

• Better usability of sentences and tag based prompts, see our standard dragon lady who is no longer brown and has the correct expression.

• Plus whatever else. Quirks are still being worked out to figure out what to finetune exactly.

• Shoutout @freek22 for Furry Enhancer 6.1

• Don't forget to join the Furry Diffusion Discord for discussions and posts of everyone's work: --> https://discord.gg/furrydiffusion (I'm mostly active there)

V2.X: SDXL, Pony compatible

Having problems with rampant humans everywhere? Try adding source_furry and/or 'furry art'.

Tip:

As some have asked for help for a point to start your prompt from, something like:

source_furry, source_9, score_8_up, score_7_up, furry art,
Rest of your prompt here

Optional photorealism, realistic, and for negative;

low quality, bad composition, blurry, bad ???????, bad proportions,ugly, lacking, mediocre, mutation, conjoined, fused, clone, mirrored, disjointed, incomplete, watermark, signature, artist name, camera, extra ????, skin, human feet, ((human, humans)), score_4,

Should be a good starting point. I have noticed over the last long while, depending on the subject you're trying to do having too many quality tags can be detrimental so a KISS approach, (keep it simple, **) would be good.

Older notes:

New v1.9 after 8 months!:

1. Fluffyrock tags

2. Many hundreds of tests run in total over 5+ models

3. Should be an excellent allrounder.

Now back to the old stuff:

This model mix was found by 'accident' a month or two ago and has decent requests on Furry Diffusion for it to be available for download so here it is. It can do both safe and mature works.

The mix is supposed to handle traditional art and classic 'toony' furry styles, but as the base models used for mixes are seemingly moving away from that it is a good idea to try all versions if you cannot get something specific to work right.

As such: 1.2 is a good starting point. 1.4 could/will result in different poses and scenes.

But with that said, the fault is obvious and I am working almost every day for more, better tuning, finding the faults or strength of the model.

For instance:

• Extremely high background detail, even without LoRa's. Especially when using hires fix can get very fancy things built into the background of whatever scene you, the director is trying to create.

• Newer models might have issues with eyes, as the artist data gets a bit 'messy'. Thus, it is recommended to explicitly mention what color eyes you want, and pupils if applicable. Adding or removing artist tags can also help a ton in this case.

• Extremely niche Kinks/fetish works might need older model versions like 1.2, this is a known thing so I'm working on it.

Otherwise, new models are at the mercy of others since this is a tuned mix of other tuned mixes for maximum detail and universability, or that is the idea.