Respiced with updated loras and training. Someone said to try to merge it with bigLove but they're already pretty similiar. The direct merge was too overpowering so it was circularly tossed around in a crazy merge path to get here. People wanted better anatomy this has that. You can use it like bigLovex1 and get similiar results with the 10sxl look, or use it like the other versions with a mix of Pony and natural prompts. It is definitely not bigLove though you can compare seeds to both versions. Use absurdres when doing big images.

Loras seem to work well. The Pony ones trained on 3d/photo work better but strong anime Pony loras still work. This version probably has the strongest Pony in-blocks out of all of them.

X4 - Same prompting and subject focuses, but with a smoother sharper and more real output. Now works with essentially all samplers. Some karras types need lower CFG ~1-3. Some will push 5-7 CFG before burn out. Should still work with all XL and Pony loras and concepts. For more realism always prompt for "realistic, photo" and a negative for "source_anime."

Best with DPM SDE / CFG 3-5 / Steps 10-20

Always at least do some highres for details in a final image. Use a hybrid latent upscale with 4x-Ultrasharp and minimum 8-10 denoise steps at 0.4 denoise strength, also with higher CFG can be better.

This version is more structurally accurate, faces can come through without detailing needed with a few more added steps. Slightly slower but better.

X3 - +offset, less bad seeds, Karras 8-12 step turbo workflow. Use fewest negative prompts as possible for better images.

X2 - Euler Karras/Exopnential

tl;dr: Very stupid diet-Pony model. Entirely female focused. It's called Exponential because its a turbo/dmd/lightning/lcm and it only really works with one sampler and scheduler: Euler A - Exponential. CFG 1.1-3.5, steps 8-20, more steps=more detail and seed lock-in. Can do all sizes and sometimes up to 1920x1080 and 2k. Clip 1 or 2. It's fast, but must usually be highresfixed or after-detailed for something presentable. I just inpaint faces at 0.35-0.4@x1024. 4xUltrasharp highres at 6-10 steps with 0.5-0.6 denoise.

DPM++ SDE Karras 5-step workflow can work but I don't favor the outputs.

Use pony prompt and natural language mix. Pony score tags can do literally nothing, or change the whole image. They usually help and affect limbs/hands and pose. Accepts both Pony and XL Loras. Makes good img2vid outputs for Kling, and the side-nipples,breast holding,sheer clothing, and pasties bypass their filter.

Full info:

I would say 35/65 RealPony-photorealXL split if I had to describe the feel of using it. You can use natural language, Pony prompts, noobAI quality prompts, or just ignore using any pony tags and get a chat-GPT paragraph prompt and press the butone. It'll do ugly, old, fat, amateur, young, but I usually stick with busty and sexy.

Why does EulerA-Exponential work? I don't know. I just noticed you can throw on every turbo lora there is and this sampler/scheduler combo will still produce images where all others will be blocked and fried. LCM and still works of course but we all know how unappealing and fake it outputs, these have detail though.

My username is on it because there is weight from my own AI and 3D works in there. Not finetuned but spiced in through 2 trained loras on to pyramidal merges so it's spread throughout the structure of it.

Basically it's a structural merge from multiple Real Pony models, base pony, and 2kXLfp32 with slight style lora merges based on my AI and 3D works. Can you just merge a RealPony model and 2kXLfp32 and produce this? Go try to. There have been similar models that try something like this is one. It's also not fully Pony; it fails with multiple-character prompts and rare actions without lora support.

This is just my merge that I'll be using and training Loras for. I don't work by finding perfect seeds, so I usually use prompt generations as a canvas for inpainting and stitching and layering to work the image to completion. If any feedback that this doesnt do naked men or anal sex really well, then it is what it is. I like big titty pinups and that's pretty much what we're doing here.

The whole pursuit for the last couple months was that I don't like the 2.5d look of Real pony, and I don't like waiting 40 steps for an image, but I like the idealized bodies and overall sexiness from it's 3D influence. I tried training models for RealPony and it just didn't work. Testing those training runs with this however produce likeness immediately. Pony models don't seem to have enough latent real face accuracy to train realistic facial diversity or likeness so I'll just use this for now.

It wasn't made for porn, but it can do. Prompt for it like with an XL model since it will hijack sex prompts with XL understanding, but there are way better models for that. There are a lot of Pony prompts that will not behave like they do in Pony, straps, sex, gesture, etc can get mixed understanding and cause noise it's probably the key weakness of the merge.

Score tags can do almost nothing or they can entirely change the image. It depends on how booru tagged up the prompt is. BREAK usage can have a huge impact (in A1111) and I'm unsure how to use it best but do use it. The most successful base prompt and negative I've found is in most of the preview images. You can remove source_3D, source_anime from the negative to get a 3D or fake look that can be aesthetic. That negative looks odd, but it's well thought-out. Those are strong core Pony tags I got from the base literature and they introduce a lot of manga noise to images, so having them in the negative helps almost all seeds that are trying to look realistic. Bad quality, bad hands, etc have little to no effect and I think they haven't in any model since NAI 1.5. Best quality, highres do however work in the positive but don't really improve most seeds.

Can tend to produce younger characters so be very careful with negatives and prompts. Turns out the 1girl anime girls are very child-like with XL outs and XL knows how to draw children so it just is what it is, protect yourself with child, loli, kid, etc in negative and that usually solves it. Keep in mind a girl is usually implying someone very young, so use woman if that's what you intend. You won't get milfs or old ladies unless you prompt for them. And this model can definitely do hot moms if that's your thing.

Very advanced users:

IF you want to merge this with another model or do dumb things like I did; you can make your own 'exponential' version of a model to merge with this. I used A1111 main branch and the supermerger extension to make this merge string, or if there is any other merger that works like that that than can target induvial lora block weights. You need the Loras first obviously.

It will look something like:

sd_xl_turbo_lora_v1:0.4,sdxl_lightning_8step_lora:0.12:OUTALL,sdxl_lightning_8step_lora:1:INALL,dmd2_sdxl_4step_lora_fp16:0.12:OUTS,LCM_LoRA_Weights_SDXL:0.12:OUTD,

Yeah that's what is turboing this model. How much turbo? Yes. There is variability that you'll have to test and adjust per model, sometimes you will want more 8-step lora on all the outs, or more overall turbo_lora up to 0.55, or you can remove LCM entirely, but imo LCM is a key addition for the face coming in at around 10 steps and smoothing out the image. When testing if EulerA-Exponential, CFG 2-3 and 12 steps produces an image with proper structure with prompts that fit the model. it's done. Then you can merge it with this in a merging-block-weight form like GRAD_A or GRAD_V and even push it to a full 0-1 spectrum across the weights. Keep M0 at 0 or 1 to prevent it from getting cooked.