BustySigrid - Flux

Sigrid is a beautiful woman with long blonde hair, striking blue eyes, a thin waist, long legs, and a generous bust. At rest, her mouth is usually open just a little bit, but she has a really nice smile that really lights up her eyes.

Sigrid is great for all your blonde superhero needs. But she also looks good in a cozy sweater with jeans. Or a bikini. Or anything, really.

This was one of the more expensive LoRAs I've made. It took me three tries on civitai to get models that were pretty good, and even now I'm not 100% happy with it. So I may revisit Sigrid later once my buzz total has recovered. LoRA training info is below for the curious.

Unfortunately, this LoRA isn't perfect. It's inconsistent at times in a few ways, but odd images are generally correctable with a bit extra in the prompt.

• She sometimes has brown hair. Do you know the old movie "Gentlemen Prefer Blondes" with Marilyn Monroe? Well, Flux doesn't. It's really damn hard to train Flux to put blonde hair on a person every time. Adding "with long blonde hair" to your prompt will pretty much fix it without altering the rest of the image.

• Her eyes are sometimes less blue than other times. I don't know why. As a catchall, starting your prompt with "a photo of a bustysigrid woman with long blonde hair and blue eyes" will take care of both problems easily enough.

• She very occasionally gets rendered as a really overweight woman. I'm at a loss to explain why. The picture usually looks very realistic and true-to-life, but nothing at all like Sigrid in any of my training photos. Your guess is as good as mine. Forget that seed and move on. You don't even need to change your prompt, just use another seed.

My sample images were mostly rendered with CFG 3 and 1024x1024 resolution since that's close to a standard baseline for Flux. If you want the best of photo realism, though, I've found the following settings work really well for Sigrid as well as many other models:

• DPM++ 2M / beta (called "dpmpp2m" in comfyui)

• Steps: 25

• Flux Guidance: 2 or 2.5

• any resolution from 768 upwards

• I've started using SDXL-like style suffixes on my prompts to boost images in one direction or another. For photo realism, I use this at the end of my prompt: "A realistic and detailed photograph taken with a 35mm film camera. The photo has extraordinary detail with a shallow depth of field, bokeh effects, enhanced contrast, and high dynamic range. The image is epic, cinematic, and gorgeous."

• For another fun effect, try this alternate prompting style: "Multiple views in comic book panels. A bustysigrid woman is cooking dinner in her kitchen. A graphic illustration image in a comic book art style. Graphic novel art with edges clearly outlined in black strokes and filled with vibrant colors. Flat shading and high detail."

So this is Sigrid. I hope you enjoy making pictures with her. If you do use this model and post to Civitai, I'd love it if you either post directly to this model page or else add the model reference as a resource in your posting. That way I'll get to see the pictures other people are making with her, which is always fun.

LoRA training stuff for anyone who's interested...

I'm by no means a training expert as I'm still mostly just getting by. Unfortunately, I haven't yet been able to make Flux training work on my machine at home, so I'm limited in experimentation by how much buzz I have here. I only get a couple of shots at each LoRA before risk going broke. That said, here is what I learned this time around.

In short, more repeats are better than more images, if you have to choose one. At first, I was using a long list of training images. First it was 73 then it was 122 when that first effort didn't work out well. The problem is that with that many images, and I knew I wanted about 20 since every other Flux LoRA I've trained, epoch number 20 has turned out to be the one I published... anyway with that many images, having more than a couple of repeats for each epoch made the buzz cost VERY high. So I did a batch of 2 on the first run with 4 repeats, than on the second run a batch of 1 with 2 repeats (lots more images in that batch).

Attempt 1: 5840 steps

Attempt 2: 4880 steps

Both of them failed.

So I went back to the drawing board and went with the setup that had worked well with other models of mine in the past. I cut out a bunch of the images and boosted the repeats.

Attempt 3: 31 images, 20 epochs, 10 repeats, batch of 2, making 3100 steps.

Around 3000 steps is usually a good number, so I hear, so this training set put me near that number. And once again epoch 20 turned out to be the winner, although a couple others were close.

Now keep in mind this is completely anecdotal. I didn't have the luxury of tweaking one of those values at a time to see what popped out. Also this is a character model, so the requirements are probably quite a bit different for style models and pose models and other such things.

Happy modeling and good luck!