I trained this model using a dataset composed almost entirely of real-world photography. My goal was to achieve realistic textures and lighting, moving away from the "plastic" look of purely synthetic datasets. While this implies the model isn't technically "perfect" yet, it does a crazy good job at what it was designed for.

As usual, the model doesn't require a trigger word.

v1

The Training Experience

This is the MRB version using my whole dataset to get a taste of what it's like to train on """this""" base model. Honestly, creating a LoRA on this base model for the first time kind of sucks (to train). Don't expect the plug-and-play quality of the LoRAs I made with Klein (personally my favorite model, only now outclassed by this one—at least in the T2I part).

I've noticed that most LoRAs out there using this base model share a common issue: using the LoRA at a full strength of 1.0 burns the image. I'm quite sure this is because we all used the first version of the Ostris AI-Toolkit (I personally don't know if there will be another version, but I hope so). If any fellow LoRA creators are reading this and have found a better tool or have any advice for this base model, please share it!

🔥 !!IMPORTANT!! 🔥

💪 LoRA Strength

Since it gives burned outputs at full strength, I tested extensively to find the sweet spot. This is what made me go crazy: sometimes the LoRA gives great results at a strength of 0.8, and other times it works best as low as 0.5 (this is the general range, but sincerely, play around with it as you like).

For demonstration purposes, I added a number in the top corner of the showcase images indicating the strength I used (usually in the top right, with some exceptions). (Note: One of the images says 7.5, but it is actually 0.75. I only noticed the typo after I finished generating everything!)

🛠️ Prompting & Generation

To get the best results, you NEED to use the KJ Prompt Builder. It helps you build that polarizing JSON prompt, which is currently the only reliable way to get a good result since typical text prompts don't work well with this base model. Personally, I think this is a great change because it gives the user so much more control over the generated image (I can't tell you how much I loved creating those images with this tool). It obviously has its limitations, but it definitely does its job.

I personally didn't apply the LoRA to the unconditional model (as you can see if you drag and drop my presentation images). I haven't tested the difference yet, so feel free to experiment and let me know!

Since I don't have massive local compute power, I use a paid cloud GPU service (renting a 5090) for testing and training. Testing every single LoRA epoch with every possible strength combination is a real pain and quite expensive. I'm not begging, but... :) if you really like my work and want to support it, consider donating on my Ko-Fi! (Or I'll decimate the population of oranges in this world 🍊).

⚠️ Known Issues & Tips

Clothing: Unfortunately (maybe I was just unlucky in my tests), this LoRA struggles to provide a fully covered top and will almost always generate cleavage. I'll work to fix this problem in the next version, but for now, I hope it doesn't bother you too much.

Bounding Boxes: I suggest not creating a specific bounding box just for the "bazoongas" (which is what I initially did). Instead, factor them in while making the base bounding box for the woman. But as always, feel free to experiment!

🚀 What's Next?

This is the only model where I'm 100% sure I'll make another version of MRB. It gives crazy good results, and I really love the quality of the base model. The first thing I'm going to do for V2 is write the JSON captioning manually, since for this version I relied on the auto-captions generated by the Toolkit.

After saying all that, have fun with it... and use the damn Prompt Builder! <3