Lora file in case you want to mix and match with other models : https://civarchive.com/models/18633/concept-character-designer-ccdwr-lora
Trained with 100 images focusing on "micro-crops" of character equipment and accessories using https://openart.ai/discovery platform!
Not capable of outputing sci-fi / futuristic images
High coherence on carryied equipment and accessories
Awesome for stamped / embroidery multilayered outfits
Good for generating fantasy scenarios as well
Description
One of the models with the best achieved trainings and fixed through many merges to remove overfiting.
Just in case you need the Lora alternative: https://civitai.com/models/18633/concept-character-designer-ccdwr-lora
FAQ
Comments (8)
Unfortunately this model is not working too great for me. I love the example images, but in practice the pics I generate - even when using the exact prompts from the example images - have blurred and distorted faces. Perhaps this is my 4gb VRAM and 540 x 540 image size but I tried it a bunch of different ways and couldn't get any good clear faces. In addition, applying the artstyle to other images via Img2Img generally results in a very chaotic and mishmashed look, it does not "play well" with others.
After about 90 gens I hadn't really seen anything that deviated from the example images as well, despite the 5gb this seems to hew very closely to a small set of data. I kept seeing the same basic layouts, poses, and backgrounds.
I want to commend the author for this model, if the target was the example images, it hits it very well (other than the blurry face issue, which might be something to do with me and not them). And it's not yet another waifu dungheap. But if you want it to generate something that isn't similar to the examples, it really doesn't do that.
not sure it'll make a huge difference, but could you try 512x512 generations? Also, if you have 4 GB VRAM, I think there's likely an optimization setting --low-vram or similar (even xformers) has small distortions and introduces irreproducibility.
I've just posted some tests and configs with the lowest parameter as possible to output "something" .... more precisely "small portraits" that looks like icons!
Could you make a Lora for this? Much better to have a 300 mb file rather than a 5 GB one :-)
Sure, however, I don't think that you would achieve the same quality, since Lora hacking tends to output something that looks like a filter.... Probably by the end of the day I can give you a Lora file
Extracted a Lora, and it works well with different checkpoints giving a different 'flavor' to this concept.
Thank you for the share!
Not uploading it because you're the poster and I'll wait for your own upload.
This is a great utility for extraction
https://github.com/sashaok123/LoRA-EXTRACTOR
@AK_3D thanks!
I've used kohya's GUI, and i couldn't increase the rank higher than 256 to output something "with higher fidelity" : https://civitai.com/models/18633/concept-character-designer-ccdwr-lora
@ivangarciafilho Yes, it's RAM dependent. I was able to extract a 320 version. Happy to share it with you.
Thank you for sharing some good models!