TAME Pony: The Authenticity MachinE

Version 2.5 is here!

Like version 2, the changes in this version have been achieved only through merging.

Since I had a lot of feedback that people missed the vivid colour and Pony knowledge (characters, men) from version 1, I went back and reworked the recipe from the start to try to get as much of that as possible back into the model while retaining the improvements from version 2.

I also improved the diversity somewhat. It's still limited, but at least now you can actually prompt for Asian girls and get them!

I did have to compromise slightly on anatomy and realism, but overall I think it worked pretty well. Let me know what you think in the comments.

Version 2.5 Usage:

I like CFG 5 for this one, but anything from 2-7 generally works well. It has the same wide CFG range like version 2, so feel free to crank it up higher as long as you raise the steps to compensate.

I haven't tested the samplers as thoroughly as I did with version 2, but the DPM samplers are still working well with SGM Uniform or Simple. Still not working out of the box with Karras, but it does work if you set sigma min to 0.1 (note that Version 2 required 0.3). You can set this in options/sampler parameters in A1111, or with a node in ComfyUI.

In my opinion, DPM++ SDE and DPM++ 3M SDE with SGM Uniform scheduler typically give nicer results than Euler a on this model, but try around.

As always, please post images and feedback so I can see what everyone is up to!

P.S. This will likely be the last version for a while, as I don't think I can squeeze much more out of merges. Version 3 will come, but not until I manage to do more custom training.

Version 2:

No new training this time, but hundreds of hours of merging, testing, and tweaking to squeeze more quality and realism out of the model. Two more SDXL models (bigASP and NightVision) and two more Pony models (CinEro and One-Trick) were introduced to the mix. Version 2 still has the realism, responsiveness and capabilities of the TAME you know and love, but with improved anatomy, clarity, image quality, sampler compatibility, lighting, and artistic capabilities.

This is NOT just a porn checkpoint. Yes, it can do realistic XXX really well, but there is much more it can do, so look at the example images and try around.

Version 2 Usage:

Same guidelines as version 1, but with a few extra tips:

CFG 3 will give good, realistic, high quality results, but the output will be less vivid than version 1. If you prefer that bright and colourful feel, increase CFG to between 5-7.

If you want to get artistic, higher CFG even up to 20+ can give interesting results! But increase the number of steps if you start seeing colorful artifacts or other issues.

All the DPM samplers are now working well, IF you choose SGM Uniform or Simple as your scheduler. If you are using an old version of Auto1111 or something where you cannot independently choose the scheduler, they may not work.

If that is the case, or if you want to use one of the non-working schedulers, you can go to settings, sampler parameters, and set sigma min to 0.3 (may not be the optimal value, but works pretty well for me). This should fix Karras and most other schedulers except KL Optimal. BUT remember to set it back to 0 at some point because it will negatively affect the results if you use schedulers that were already working!

Please post interesting results so myself and others can see what you have been up to with the model, and get some new ideas of what it is capable of!

Version 1:

This model is all about maximum realism and sexiness. It aims to achieve a new level of realism for Pony models. While there are a lot of amazing Pony realistic models out there, most of them suffer from "Ponyness": their Pony heritage is immediately clear when you look at faces or anatomy. TAME certainly has its flaws, some of which I hope to remedy in future versions, but from what I can tell the amount of "Ponyness" is very low.

Many creators are jumping on the Flux bandwagon now, which is understandable. It's a great model. But for those of us stuck with older GPUs I don't think it is the best option (if it is an option at all). I've also noticed a degree of "Fluxness" is present in most/all of the fine tunes.

TAME began with a series of checkpoint merges, using block weight merge to combine a set of realistic PonyXL and SDXL models in a way that maintained the prompt adherence and flexibility of Pony. I then trained the resulting model on my own dataset to further improve the realism.

The Authenticity MachinE will not win any awards for creativity but it is damned good at making realistic pictures of women in any state of dress or undress.

Quick start:

Sampler: Euler a

Steps: 20

CFG Scale: 3

Resolution: 912x1280, 1024x1400, 1280x1536

Hires fix: ESRGAN_4x

Upscale by: 1.5

Hires steps: 10

Denoise: 0.3

Usage guide:

Score tags: You do not need score tags with TAME. Putting them in may not hurt but it likely won't help either, so why waste the prompt space?

Quality words: You do not need quality words (8k, masterpiece, best quality, etc) with TAME. They are a waste of prompt space.

Negative prompts: You do not need negative prompts with TAME unless there are specific things you want to exclude.

Positive prompts: Prompt length doesn't matter too much, but keep it simple. Words and phrases with commas in between. The model understands Pony style prompts, but does not do well with natural language prompts. TAME usually responds well to gentle prompting (take a look at the example image prompts), so don't use a lot of emphasis e.g., (large breasts:1.8) unless the model is being stubborn. Start by just telling it what you want, then play with emphasis, rearranging words, and more advanced techniques if you aren't getting the right results.

Don't fill up your prompt with nonsense words. Look at this example I copied from another model's gallery (RealVis XL V5.0):

photograph the little catgirl, cat ears, wearing fur dark coat, 50mm . cinematic 4k epic detailed 4k epic detailed photograph shot on kodak detailed cinematic hbo dark moody, 35mm photo, grainy, vignette, vintage, Kodachrome, Lomography, stained, highly detailed, found footage, (masterpiece), (best quality), (ultra-detailed), the little catgirl, cat ears, wearing fur dark coat, illustration, disheveled hair, detailed eyes, perfect composition, moist skin, intricate details, earrings, cinematic still the little catgirl, cat ears, wearing fur dark coat . emotional, harmonious, vignette, 4k epic detailed, shot on kodak, 35mm photo, sharp focus, high budget, cinemascope, moody, epic, gorgeous, film grain, grainy, the little catgirl, cat ears, wearing fur dark coat, detailed, elegant, highly colorful, warm light, sharp focus, beautiful, intricate, expressive, rich deep colors, cinematic, cute, enhanced quality, creative, positive vibrant romantic atmosphere, depicted, perfect background, professional, thought, iconic, best, thoughtful, pretty, attractive, charming, confident, passionate

I can't speak for RealVis because I didn't create it, but I strongly suspect that much of that prompt is doing nothing. What I can tell you is if you use that prompt with TAME you will not get great results. Here is a quick rewrite:

catgirl,earrings,cat ears,wearing dark fur coat,warm lighting

See how simple that is? From 142 words down to 10, that's more than a 90% reduction. Now, observe the difference in the actual renders using TAME:

Long prompt:

Not terrible, but probably not what the user was looking for.

Short prompt:

See the difference? 90% less words and a much better image that is likely also closer to what the user was looking for (romantic atmosphere, elegant, etc). In fact, it looks so good that I decided to use it in the model showcase!

Adetailer: I do not recommend using Adetailer with TAME. In my experience it makes things worse. Most of the time you are better off with just Hires fix. If you are having trouble with e.g. a really small face then maybe give it a go, but don't expect miracles. Often changing the settings (sampler, prompt, resolution, steps, denoise, etc) will get you a better solution. For example, full body poses often render best at a more narrow resolution (square resolution often makes the faces smaller) and e.g. work better with Euler a than with DPM++ SDE. Too few steps can make things blurry while too many can cause artifacts and strange eyes. Higher denoise can fix more issues but can also mess things up.

Resolution: This model was trained on a high quality dataset with a max training resolution of 1536x1536. That means you can often generate at up to 1536x1536 without deforming your subjects. The model also works well at pretty much any width-height combination, even extreme ones.

For example the image below was generated at 2048x408:

And this one was generated at 504x2048:

These are just test images done on the fly, without Hires fix, but as you can see neither of them has any obvious deformities or other major issues. The second image would look significantly better with Hires fix and the extra finger could likely be eliminated by trying a different seed.

Basically you have about 2.4 million pixels to work with, and as long as you don't exceed that by much (width x height) the model works. For example, you can crank the width all the way up to 2048 and set the height to 400, 800, or 1200... but if you go much above 1200 you will need to reduce the width or you will start to see weird things happen.

Even a tiny change in the resolution typically has a large effect on the image (pose, etc). So be creative, try different resolutions!

Upscaling: I highly recommend using Hires fix, but make sure you choose a good upscaler. My favourite is 4x_NMKD-Siax_200k, but 4x_foolhardy_Remacri and ESRGAN_4x are also good choices. I generally set it to 1.5x resolution with approximately half the number of steps used for the first pass and a denoise strength of 0.2-0.4 (default for me is 0.3). You can of course upscale further if you desire.

Samplers: Euler a is highly recommended as it will give you good results with a range of settings. DPM++ SDE can work well with some images (avoid it for full body poses) but it really has to be dialed in or it will look terrible. Other working samplers include DPM++ 2S a, DPM++ 3M SDE, Euler, DPM2, DDPM, and LCM. Most others do not work properly with this model.

Schedulers: I haven't played around much with different schedulers, so you are on your own with that. I use Forge (an offshoot of Auto 1111) which has limited options for sampler/scheduler combos, but I haven't noticed huge differences in any case.

CFG Scale: Typically 1.5-5 is best (I keep it at 3 most of the time), but you can try going higher if you wish. Worst case you get a bad looking image or two, right?

Steps:

Euler a - 15-30 steps + 8-15 hires steps (for best quality I typically use 25 + 12)

DPM++ SDE - 6 steps + 4 hires steps (for pure speed you can even drop...