Trained on images of artists whose artwork I find aesthetically pleasing. Example images have very minimal editing/cleanup.
If you like my work (models/videos/etc.), feel free to contribute here:
Discord channel: https://discord.gg/ZKX3Zym8FQ
Tips/Donations: https://streamlabs.com/kasukanra/tip
Reproducibility
If you want to reproduce the images below, I've prepared a more detailed version of the prompt settings and seed values in these pastebin(s):
v2-0: soonTM
v1-8: https://pastebin.com/EAf3GLpv
v1-6: https://pastebin.com/c7zphaN1
v1-5: https://pastebin.com/Nqz6vNFN
v1-3: https://pastebin.com/xs8mmybZ
v1-1: https://pastebin.com/2XrkSMfV
If you want to use this with the graphic illustration LoRA I created here: https://civarchive.com/models/11477/graphic-illustration-style, make sure to add the prompt "in the style of <lora:graphic_illust_12:0.4>". Tune the strength of the LoRA to what you think is appropriate, but don't use 1. It will blow out the image.
Stable Diffusion Workflow Videos
Comfy Org Meetup in Tokyo, Japan
FLUX.1 [dev]: Training LoRA with SimpleTuner and AI-Toolkit
SD3 2B Medium Training
Paints-UNDO | Create a timelapse for any 2D image
Stable Diffusion 3 Early Preview
High-resolution advanced inpainting with ComfyUI (SDXL, IP-Adapter, rgthree)
Improving Stable Diffusion images with FreeU (SDXL, LCM, Turbo) and ComfyUI API
Autogen/AI Agents & Local LLMs autonomously create realistic Stable Diffusion model
Worldbuilding an IP with 3D and AI ft. MVDream | Part 1
ControlNet 1.1 for Character Design | Part 2
Creating a PNGtuber with Character LoRA
Training LyCORIS (LoCon/LoHa) for Style | LoRA Part 2
ControlNet for Character Design | Part 1
Training LoRA for Style | LoRA Part 1
Creating Stylized Vector Logo Characters
How I train DreamBooth for Style:
Model Demonstration:
Example settings: Steps: 50, Sampler: Euler, CFG scale: 7, Size: 512x512
Example prompt: DBfantasyart style, masterpiece, 1girl, beautiful female portrait of sorceress
Description
Some of the v1.8 generations are used in combination with another style LoRA I created called graphic illustration style: https://civitai.com/models/11477/graphic-illustration-style.
The example images are not perfect, but I left in the imperfections instead of correcting them so you can see what kind of quality you get straight out of txt2img. If there's a metadata icon included on the bottom right of the image, it means that there were no external modifications.
FAQ
Comments (15)
For the latest version v1.8 I downloaded also the graphic illustration style which you uploaded to your profile and downloaded the ckpt, but there's no safetensors?
Should I load the ckpt as a model inside webgui? Or should I use a previous version of the safetensors?
I've been getting some very weird results which I think might be an issue related with the above.
Thank you!
I uploaded a safetensors model for v1.8 recently. The v1.8 checkpoint should work without issue as long as you use the token "DBfantasyart". I also added a pastebin with prompts as a reference to generate results. Maybe the downloaded LoRA is named differently?
@kasukanra I managed to get some decent results for characters (not great and im unsure why as I used the same prompt I use to test every other model) , but as far as backgrounds go, they seem to be very chaotic even, should I go for a lot of prompts or keep things fairly simple?
Also great videos as well! :)
@RykeWollf13 Thanks for the kind words!
Anyway, can you check if the downloaded LoRA name and the LoRA name in the prompt are the same name? Also, did you turn down the LoRA weight to around 0.4? Using the graphic_illust LoRA at strength of 1 will blow out the image really hard.
Also, can you try the prompt for the background image in v1.8 (the city in the clouds)? Does that reproduction work for you?
I also updated the pastebin with the prompt to create the base image for the character. I then took it into img2img with ControlNet along with inpainting to create the final image.
@kasukanra I can see I'm already making a tonne of mistakes right from the beginning.
The title of the Lora being graphic_illust_12 and the weight being applied there in the prompt correct?
example: (graphic_illust_12:0.4)
one of the main issues I'm getting is that I'm getting a really epic preview, but then it switches to something completely messed up, is it due to the CLIP skip: 2?
Not sure if there's a way i can share some images with you of what I'm getting before and after, but it's a bit frustrating when I see such an awesome model but me failing so miserably at it haha.
@RykeWollf13 For the fantasy art style model and the graphic illustration LoRA, all my generations are done with CLIP skip of 1.
Also, I'll be adding more and more example images with their prompts in the pastebin or the metadata on the bottom right of the image for reference.
@kasukanra turning down clip skip back to 1 really helped (now i need to revisit all the other models I tested and was getting weird results in hmm) but still on the anatomy part I'm getting some weird artifacts even when adding a lot of negative prompts to try and correct it.
Main issue right now is getting the face to pop up correctly.
Overall, the style is super solid and easily one of my favourites.
@RykeWollf13 I'm glad to hear that changing the CLIP settings helped!
As for the face, it is almost always an issue. I rarely get something nice on the first try, and it requires inpainting. As for the anatomy, you can always inpaint in some limbs (if you have anatomical knowledge), then run it through ControlNet img2img and other stuff.
I keep wanting to go back and re-train this model on 768 resolution images to hopefully fix some of the problems you mentioned, but I think the diffusers repository is currently broken (for me and some other people) https://github.com/ShivamShrirao/diffusers/issues/208#issuecomment-1439091407.
@kasukanra thanks for all the help, I'm gradually getting there even with the faces on the base model itself which is pretty incredible, I just need to keep playing around with the prompts a little bit here and there but it's getting there.
As for inpainting I need to learn how to properly do that, if you have a guide you recommend I appreciate it. I'm a classically (fine-art) trained freelance artist so anatomy isn't an issue, just looking into some models to either get some inspiration or generally help me in my work flow. :)
@RykeWollf13 Nice, another artist among us. I'm a digital concept sculptor.
Also, I mostly really only use diffusion models to get ideas as well. For inpainting, I'm not really sure if there's a good guide out there. I have a video where I kind of go over it as an aside, but it's not a full on deep dive into it. I show how I use inpainting slightly starting at 8:13 in this video: https://youtu.be/w-wMXn-EsGw?t=493.
Maybe bringing a txt2img output to final concept or illustration will be the topic of my next video.
@kasukanra do we still need to use (graphic_illust_12:0.4) for the latest version 2.0?
@RykeWollf13 It's up to you. I usually use the LoRA to tighten up the image if it's too painterly.
@kasukanra oh which lora? I thought the graphic ilst was a hypernetwork ahah
@RykeWollf13 lol graphic_illust is a LoRA and the one I was talking about.
@kasukanra oh silly me must've gotten it confused haha. Cheers! ;)
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.