Type | |
Stats | 3,935 |
Reviews | |
Published | Jul 7, 2023 |
Base Model | |
Usage Tips | Clip Skip: 2 |
Hash | AutoV2 0DCB1BC5AB |
Ardmix is a model focusing on male anime characters. Ardmix is in its early phase. There are a lot of things that need to be improved. Any feedback is more than appreciated.
News / Bulletin board
[26.08]
Forgot to upload a Safetensors file of the current version. I'll upload it soon.
[12.10]
Pruned (no-ema) + safetensor uploaded
---
- test v0.2-experimental, if good: release
Contact
Ardmix Discord server: https://discord.gg/Cnay2KfyDk (WIP)
Discord: ard_ai (Feel free to contact me about anything. I'm very active on Discord.)
Quick start guide
Sampler: Euler a, (DPM++ M2 Karras)
Steps: ~50, (try to stay above 20)
CFG scale: 7
Some Loras are tested and they work well with the model
Try to keep your prompts simple
NSFW: Not yet thoroughly tested
Use Booru-styled tags
I highly recommend the usage of 'a1111-sd-webui-tagcomplete', An a1111 sdwui extension for autocompleting Booru tags.
The model works best with 512x512, 512x768 and 512x1024.
512x1024 is especially good for full-body portraits.
Twitter advice: Twitter will crop 1:2 ratio images in a very unfortunate way, so 512x768 is a very good choice if you want to upload to Twitter
Loras that nicely complement this model:
add_detail: Adds much more detail to the renders. But be aware, this LoRA quickly starts to mess with hands (using Ardmix)
How to write prompts?
Check the preview images. A more detailed work-through will come soon.
Use '1boy' to target male characters. The required weight might vary from prompt to prompt. A weight between 1.0 to 1.3 is recommended.
An additional step might be to add '1girl' to your negative prompt. This isn't required but might help if the model acts very stubbornly.
File guideline - Which version to choose?
Each version released comes with a few different files. The question that may remain for some probably is: Which one to choose?
For simplicity, always choose the fp16, no-ema-versions. no-ema-models are pruned checkpoints, meaning some data (weights) are pruned away.
This sounds like the quality is worse... Actually no, the weights that are deleted are close to zero, meaning they would have had a very minor impact on the final image. Non-pruned models are good for further training but are inefficiently large for image generation.
For the paranoid people: EXAMPLE COMING SOON
If you want to know how to simply prune a model, check out this well-detailed video:
Prune Models, Save Space 🤩 Best Settings for Converting & Pruning | AUTOMATIC1111
Roadmap
Increase the amount of male training data (Trigger words like 1boy, 2boys).
Bake the style of the preview images into the model itself
Mixed models
Cetus-Mix https://civitai.com/models/6755?modelVersionId=78676
Dark Sushi Mix 大颗寿司Mix https://civitai.com/models/24779?modelVersionId=29645
Ardmix wouldn't be possible without the hard work of the creators of the models listed above. Thank you! (Both models are marvels on their own, it's definitely worth giving them a thorough try)