New! Check out my SDXL version of this model:
https://civitai.com/models/130869/fantastic-characters-sdxl

Introducing Fantastic Characters! There's witches, and robots, and creatures, oh my! (and much, much, more)

Need to make a new character? Maybe it needs to be something from every day life, like a construction worker. Or maybe you need something more fancy like a futuristic cyberpunk ninja. How about a nurse with robotic implants? Aliens, humanoid creatures, robots, mobsters, cowboys... the list goes on and on. Mix and match. Impress your friends and relatives! Make a new Fantastic Character today!

Use the fantchar keyword to trigger this model. The trigger word is not strictly necessary, but it does seem to help with the finer details.

This model is designed for showing characters in portraits, solo images, posing for the camera, etc. The model is tuned primarily towards realism and 3d, but it can generate some more cartoony or artsy stuff sometimes. You'll get the best results from characters with humanoid anatomy (two arms, two legs, torso, head, etc.), though it's good at making lots of different types of creatures and characters in that general shape (see the example images.) I prefer to use this model with Realistic Vision V4.0 as the base for a more realistic look, and A-Zovya RPG Artist Tools (V3+VAE) as the base for a more fantasy look. I've also seen it spit out nice results with others like DreamShaper 7, Deliberate V2, and RevAnimated.

Here's the info on the versions:

• V1: trained with the least amount of images. Tends to be a well rounded model, but weights more towards 3D types of characters, and the output can often look like it belongs in a video game (which might be good or bad, depending on your use case).

• V2: trained with about 2x the images as V1. Often makes more realistic images than V1, and includes more details. The downside is that it is somewhat less versatile and can struggle with wildly non-human characters.

• V3: trained with about 3x the images as V1. Can make a wide variety of characters, but can sometimes look a little overtrained. If your space vikings look sunburned, try reducing the weight to 0.9 or 0.85.

• V4: trained with about 6x the images as V1. Can make the widest variety of characters. The overcooked look from V3 appears to be fixed. Increased the dimension to 128 for this version since I ended up doubling the number of source images.

I have several ideas on how to improve this for a V5 version, but that will require a ton more training time. I might even train V5 as a full checkpoint. I'll only go down that road if people really like this model though. So if you like it, add a review, or post some images, or even a comment would help. Let me know you like the model and I'll be much more likely to give V5 a go or make a checkpoint. I love to get feedback on the models I create!

Final notes and tips: I used Clip Skip 1. Anime base models might be less fantastic, but I haven't tried it so let me know. There's very little "Action" going on in the source images, so you might want to avoid using verbs in your prompt. ("Viking with axe" is better than "Viking swinging axe", for example). Be careful with your negative prompts, especially if you want something more on the 'creature' side than the 'human' side. There are several common negative prompts that will conflict or work against this Lora, so add them in as needed and remember that less is sometimes more. Weights between 0.8 and 1.0 seem to work best. Most of the source images are in the 768x512 or 512x768 buckets, but there's lots of buckets and most sizes seem to work fine for me. I didn't use any explicitly NSFW images in the dataset, so if you're getting lots of topless women, that probably isn't coming from this model. There are some attractive cosplayers in the dataset though, so it might be biased more towards making hot bartenders than plump barmaids. I did try to include a diverse mix though, so it might be fine. Your mileage may vary.

Good luck, and post some images!