<h2 id="tame-is-back!">TAME is back!</h2>For those familiar with TAME Pony, this model retains the same spirit, with rich vivid colors, highly detailed backgrounds, detailed realistic anatomy, and excellent prompt following. In other words, it looks nice and can do cool stuff.Quickstart:seeds2/bong_tangent/8 steps - very high quality, slowerer_sde/simple/8 steps - faster, decent qualityIf the model is producing nudity or sex that you didn't prompt for, add <a target="_blank" rel="ugc" href="https://civarchive.com/models/2775340/krea2-textfusion-refusal-reduction-lora">Textfusion Refusal Reduction</a> lora with NEGATIVE weights.Note: I've uploaded the BF16 but it is dequantized from the int8, so the quality should be comparable to int8. I only recommend it if you want to use TAME for merging or you have trouble running the int8. I cannot produce a full quality BF16 at this point because one of the merged models is only available in FP8.As before, I've written quite a lot here. I'm not going to feed you one of those nonsensical AI-written descriptions that tell you nothing about what the model is good at.TAME Krea is a complex merge of models and loras built using advanced block merging. I started with a block merge of DarkRealism v4 and AnalogMadness v2, which resulted in a flexible, highly stable model with a nice look. I then tuned it to my preferences and enhanced its capabilities and look by merging in a bunch of loras (again, using advanced block merging), mixed with a bit of magic and my own unpublished secret-sauce lora, which I trained on about 250 hand-captioned photos (mainly to enhance female anatomy and skin textures).Unlike some other merge fine-tunes, TAME Krea is not overly tuned toward realism. It maintains the flexibility of base Krea2Turbo quite well, so it can do almost any style, while adding vivid colors, more details, better anatomy, and its own unique TAME style. It is very flexible and capable, and will give nice results for almost any prompt you throw at it (but more on prompting below).All example images were generated with the int8_convrot version at 1440x1024, 8 steps, using a simple workflow (attached to the images), no upscaling. The only special node in the workflow is the <a target="_blank" rel="ugc" href="https://comfy.icu/node/Krea2PromptWeight">Krea2 Prompt Weight node from comfyui-kjnodes</a>. I highly recommend using this node in your own workflows with this model (or any model, honestly). The reasons are explained below under 'known issues'.Settings:For testing and quick generation, I prefer er_sde/simple at 8 steps. This is just a personal preference because it is fast and usually comes out well. But if you want to maximize detail and quality, try seeds2/bong_tangent at 6-8 steps. The results are next level, even at lower resolutions. res_2s is also great, but I am not a fan of the overly detailed skin texture it produces. The model also works well with euler_ancestral, euler, heun_pp2, and many others. Beta is a good all-around scheduler, works well with most samplers, but simple, sgm_uniform, normal, ddim_uniform, beta57, and bong_tangent are all good choices. Each sampler/scheduler combo produces a different look, more so for some prompts than others, so it's often worth playing around a bit when you have a prompt you like, to see what comes out best.Prompting:Be clear and precise and logical. If you prompt for a woman holding a cup of coffee in one hand, a purse in the other, and keys in the other, and playing with her hair at the same time, you can expect the model to screw up part of your prompt somewhere along the line because it knows women shouldn't have three or four arms and doesn't know what to do with the rest of the items (if you want a woman with three or four arms you have to ask for that specifically and then your problem is solved). I've seen many complex prompts that give contradictory instructions because the writer had lost the plot and got tangled up in their own prompt. Also, the model can handle long prompts, but don't make them long just for the hell of it. Tell the model what it needs to know, and nothing more. Don't repeat the same phrases in different ways unless you really need to emphasize something. And in that case you are better off to use the Krea2 Prompt Weight node (more on that in 'known issues').The default style of the model is more tuned towards aesthetics than ultra realism, so if you are looking to maximize realism, prompt for styles: amateur flash photo, documentary-style photo, unpolished street photo, grainy webcam photo, fashion glamour photo, disposable camera photo, retro 1970's photo, whatever you want. Not everything you come up with will be a legit style, but try around, or ask AI for the right keywords. Use lighting effects: natural lighting, golden hour, chiaroscuro, low key, high key, strong shadows, high contrast, strong blacks, washed out, overly bright, whatever. If the results aren't coming out realistic enough, or you want to focus on the details of human skin, ask for detailed skin textures with visible skin pores and fine vellus hair.DON'T ask for extreme quality 8k ultrarealism unrealengine superb best quality maximum quality ultra quality best ever superduperamazing image. At best it will do little or nothing, at worst it will degrade your output and/or reduce prompt following. Something like 'extremely detailed' might help you (although this model already tends toward extreme detail, so you might find 'low detail' or 'simple background' more useful), but just use it once, not twenty times in different ways.Known Issues and Workarounds:I've spent many, many hours testing and tweaking, but there are still a few nagging issues that I haven't manage to fully solve yet (but there are easy workarounds).<ol><li>The model is highly tuned toward NSFW. Most of the time it will put clothes on your subjects if you ask it to, but occasionally it stubbornly refuses. This is often because of something specific in your prompt that is triggering the model. For example, I had one prompt that just kept making a hole in the woman's leggings to show off her vulva, no matter what. In the end, I found that removing the term 'close-up' from the prompt fixed it. Why? I don't know, maybe because the model is overly tuned on close-up anatomical details. Despite many attempts, I couldn't found a way to prevent this behavior without compromising the model's NSFW abilities to some extent, so in the end I gave up. That said, there are two fixes you can apply in your workflow: one is to use the Krea2 Prompt Weight node I keep talking about, with NEGATIVELY weighted phrases (at -1). The other is to use the TextFusion Refusal Reduction lora, <a target="_blank" rel="ugc" href="https://civarchive.com/models/2775340/krea2-textfusion-refusal-reduction-lora">available here</a>, with NEGATIVE weights. Start small (0.2-0.3) and increase if the model is being very stubborn. At some point it should work. The output will look a bit different, but should still be nice.</li><li>Sometimes sweat looks viscous and a bit like cum. Female arousal fluids also tend to look like cum. Also cum occasionally appears when it shouldn't. These seem to be common problems on other models too. I haven't found a general fix for this without changing the way the cum itself looks, or massively reducing it. However, for your own workflow, you again have two options. You can fix it either by using the Krea2 Prompt Weight node, or by using a cum lora with NEGATIVE weights (start with 0.3-0.5), e.g. <a target="_blank" rel="ugc" href="https://civarchive.com/models/2738551/too-much-cum-krea2">this one</a>. I've tried several and they all worked. Keep in mind that it will change the output slightly. And at high weights it reduces NSFW, so if you make the weight too strong on an NSFW image, it will likely start putting clothes on your subject.</li><li>The model occasionally adds watermarks to images. This is unpredictable and only happens once in a while, but it is very annoying, and is not caused by my own dataset, so probably a lora. I will try to track it down and eliminate it, but in the meantime, you can use the Krea2 Prompt Weight node (are you starting to understand why I recommend it?) with exactly this text: 'watermark, logo in corner, text in corner (watermark, logo in corner, text in corner:-1)'. Alternatively, you can try changing seeds or alter your prompt slightly to try to remove the context that might pull up the watermark. But I haven't had a lot of luck with that, so better use the Prompt Weight node.</li></ol>I'm sure there are other issues users will discover along the way. Feel free to let me know in the comments, in case I make another version. Since I'm releasing this for free, and it takes a lot of time and effort on my part, I cannot tell when or if that might be. Don't be a jackass and demand things (most of you are good people, but I've seen a few bad apples on here), just let me know politely if you find issues, and I may or may not find a way to fix them.Acknowledgements:Thank you to the creators of DarkRealism and AnalogMadness, which are both fantastic models and form the starting point of this model, and to all the wonderful creators of loras and models that make their work freely available on this platform.

TAME Krea

*This BF16 is dequantized from the int8 version. It is NOT a native BF16. Therefore I do not recommend it unless you need it for some reason (e.g. you cannot run the quantized versions or you want to use it for merging). I cannot make a full quality BF16 because one of the models I merged from is only available in FP8.

Files with SHA256: 5219741d2c021962bfe9e685dc31de4bf657715ab5a8e42f10d739d347ee73cd