santa hat
deerdeer nosedeer glow
Sign In

AstolfoMix

78
707
11
Verified:
SafeTensor
Type
Checkpoint Merge
Stats
227
Reviews
Published
Oct 22, 2023
Base Model
SD 1.5
Usage Tips
Clip Skip: 2
Hash
AutoV2
41429FDEE1
default creator card background decoration
6DammK9's Avatar
6DammK9

AstolfoMix

An exotic merge of 20 models. See this article for description. Ignore the content below will update soon.

Abstract

I present AstolfoMix, a merge model focusing on absurdres, and well as drawing most content out of the box, disregarding the art style, as an attempt to the bias-variance tradeoff. Currently it is in anime style, and Astolfo is so cute!

Introduction

See the HuggingFace model page. I am too little time to repeat the contents there. Also all the intermediate / variants are available there. Also check out this model page for more models.

Related Works

I am not sure, seems that "uniform merge" is an original idea. However I did some information gathering about merges. Check out auto-MBW-rt for the future "boosting" approach, instead of my weaken "bagging" approach. Sorry for manual MBW fans, I don't have any perference, or the "absurdres" is never covered in such exploration.

Methodology

Disclaimer: it is trolling, unless you already know what I'm writing. I won't explain anything here. Also it may be duplicated in Experiments, because I need to be tested independently.

CS-related theory

Try your best to read my article. Seek for any PHD students / Profs if that is BS to them (Maybe I will make a paper formally). Also I still need time to conduct test write Python codes which is almost impossible in a 996 life.

Receipe

Full receipe here. Remember, use this toolkit to replace with original SD's CLIP. Whole procedure should be carried in WebUI only, without any lines of coding.

Experiements

Artworks

Included in my article, also my artworks, and my usual routine (no manual image editing). For the last part of "correlation", I'll complete it "soon". No score metric is involved, because it has been discussed, and I prefer basic metrics on differentiation before attempting to correlate aesthetic with logical justification. For my personal feeling, it "draw" anime stuffs with realistic scale, which is unique across all the models available in internet.

Prompts

This is a base model. Use what you are familiar with, without thought of "trigger words". PNG info in CivitAI is incomplete. See my artworks instead. Ignore prompt weight if you're using ComfyUI, or you don't know how it works. "Who What Where" is enough.

Samplers

Use what you like. I use default sampler because I am simple, Also it is not bad at all.

CFG / STEPS / ADDONS

Thanks for the very first review, any "default range" (CFG 7.5) works (obviously).

From experience, CFG 4.5 4.0 will be the minimum optimal value. Plot if you doubt. Feel free to use adetailer for local latent upsampling, or built-in HiRes fix for global (2.0x for 1024x1024, or 2.5x for 768x768, lower if image breaks). Default Dynamic CFG (mimic 1! Stick with the paper!), with default FreeU (Stick with the paper!). It is just a base model.

Embeddings / LoRAs

Thanks for the very first review, any generatl embedding (trainned from NAI / SD or any models "close" to them) will be fine.

Check out this post and this post for reference. (They may be appeared as cross posts). In the worst case it loss the ability to draw with Hi-Res upsampling. Moreover (it may be triggering), it supports VBP's original 273 style embeddings. Mirror will be activated by request.

Diversity of contents / Art styles

"Yes". But keep in mind, it is just a "good" base model. "Base as most contents". It cannot replace embeddings / LoRAs, or any other tools. You will need prompting skills to do minor stuffs (e.g. non human, non binary stuffs etc.)

Discussion

I think this is not my discussion. Just try to use my model like a base model: mix it further, LoRA / Finetune, everything is welcome. I may "beat" most models in such hires, and reaching SDXL's level, but I'm sure it comes with price (e.g. image layers as "v-pred", and the "bias" which must be larger then SDXL with 1024 ARB training with LAION dataset out of the box), and that is way out of my capability.

Conclusion

Just try my model! AstolfoMix only represent my personal "model selection", everybody can use the "uniform merge" to make you own base model, before great SDXL model comes. Also do not waste your time on the risky MBW / finetuning if you don't have abundant resources.

Appendix

Model used

It should be same as the full receipe.

  1. VBP

  2. CBP

  3. MzPikas TMND Enhanced

  4. DreamShaperV8

  5. CoffeeWithLiquor

  6. BreakDomain

  7. AIWMix

  8. Ether Blu Mix

  9. majicMIX realistic

  10. Silicon29

  11. BP

  12. CGA9

  13. LimeREmix

  14. CyberRealistic Classic

  15. ORCHIDHEART

  16. BB95 Furry Mix

  17. Indigo Furry mix

  18. AOAOKO [PVC Style Model]

  19. GuoFeng3

  20. YiffyMix