An anime mix inspired by the simplicity and style flexibility of the original NAI model. Ether One is focused on LORA compatibility as well as providing a flat, slightly rough aesthetic. Let's take it back to square one.
✨ Please Share Your Cool Creations Below! ✨
Version Three re-tunes weights to allow improved style variety, while also retaining that slight 'rough' aesthetic version Two had. Swapped out a few LORA's to better compliment the current style approach.
If you have any feedback I'd be happy to hear it.
Use whatever sampler, steps, CFG, clip skip you prefer.
Sample images were generated with
Sampler: DPM++ 2M Karras
Steps: 20 - 28
CFG: 7 - 11
Clip Skip: 1
ENSD: 31337
Increasing CFG can lead to more vibrant results.
Negative Prompt Example:
"NSFW, (worst quality, low quality:1.3)"
I recommend this simple negative prompt if you wish to achieve a similar style to the preview images. You can further alter positive and negative prompts based on the type of style you'd like to achieve.
Style focused negative prompt examples:
* Photorealism feel:
"sketch, drawing, anime, cartoon, illustration, NSFW, (worst quality, low quality:1.3)"
* Flat Anime/Cartoon:
"realistic, photorealistic, photograph, 3D, NSFW, (worst quality, low quality:1.3)"
Waifu Diffusion VAE kl-f8-anime2
The VAE is baked. I strongly recommend this VAE as the model was tuned with its color and saturation in mind.
Sample images were further upscaled using AUTO1111 Hi-res fix. Please utilize it if you wish to achieve similar results.
Latent (nearest-exact)
Upscale by: 2
Denoising strength: 0.55
Are your results changing too much after Hi-res fix? Please try 4x-foolhardy-Remarci.
(Also recommended for simple style illustrations, anime, cartoon, etc.)
4x-foolhardy-Remarci
Upscale by: 2
Denoising strength: 0.45 - 0.55
To achieve the aesthetic featured in the gallery I recommend using the suggested positive and negative prompts of AOM3.
masterpiece, best quality,
NSFW, (worst quality, low quality:1.3)
The model was mixed and tested using these prefixes and will give you the most consistent style results. Of course you're free to prompt however you'd like as booru-style models tend to be quite flexible in prompting.
If you look at the prompts included with the sample images they demonstrate that even with a very loose and vague style of prompting you can achieve fun, high quality images.
Experiment and give the AI some freedom to interpret your ideas!
General guidelines:
Prompt in the style of anime/booru-style models, "masterpiece, best quality, ..... "
More anime style images
Prompt in the style of photography/non-anime models, "professional photograph of ..... "
More realistic images, higher style variety
I have no training or real knowledge of machine learning. I merged the model using the information provided on WarriorMama777's AOM model page as well as the info from these Rentrys. I very arbitrarily mixed the UBM weights of each model, modifying the values with trial and error until I was visually satisfied with it.
I utilized the SuperMerger extension for AUTO1111 by hako-mikan.
Thank you to WarriorMama777 for providing AOM and other various mixes as well as detailing your workflow which inspired me to try mixing my own models.
Thank you to hako-mikan for creating the SuperMerger extension allowing a quicker workflow for merging.
Thank you to those that have contributed to these MBW Rentrys.
Thank you to gsdf, rqdwdw for Counterfeit.
Thank you to nuigurumi for BasilMix.
Thank you to Xerxemi and Xynon for SD-Silicon.
Thank you to Jemnite for Lunar Radiance.
Thank you to Yang for Watercolor Oil Painting.
Thank you to OliviaOliveira for Eflorescense.
Thank you to those at the LMFResearch Society for their archive of SD LORA's and checkpoints.
Thank you to AUTO1111 for creating the web UI.
Thank you to StabilityAI for starting everything with Stable Diffusion.
Thank you to CivitAI and Huggingface for providing places where open source AI can flourish.
Thank you to the entire SD community for continuing to openly share and create.
* Version Three:
Version Three re-tunes weights to allow improved style variety, while also retaining that slight 'rough' aesthetic version Two had. Swapped out a few LORA's to better compliment the current style approach.
If you have any feedback I'd be happy to hear it.
📗 Model Merging Info - v3
I have no training or real knowledge of machine learning. I merged the model using the information provided on WarriorMama777's AOM model page as well as the info from these Rentrys. I very arbitrarily mixed the UBM weights of each model, modifying the values with trial and error until I was visually satisfied with it.
I utilized the SuperMerger extension for AUTO1111 by hako-mikan.
▼ Merging
MODEL A: EtherRealMix-3x1-FINAL-fp16.safetensors
MODEL B: Counterfeit-V3.0_fp16.safetensors
MODEL C: watercolorOilPainting_v10-pruned-clipfix.safetensors
Triple Sum merge - ✔️ Use MBW
Alpha: 1
0,0.5,0.5,0.5,0.5,0.5,0.4,0.3,0.3,0.3,0.3,0.3,0.3,0.3,0.5,0.5,0.5,0.5,0.5,0.5,0.5,0.5,0.5,0.5,0.5,0.5
Beta: 1
0,0.3,0.3,0.3,0.3,0.3,0.3,0.1,0.1,0.1,0.1,0.1,0.1,0.1,0.6,0.6,0.5,0.4,0.3,0.3,0.8,0.8,0.8,1,1,1
Model: EtherOne-p1
▼
MODEL A: EtherOne-p1
MODEL B: Basil_mix_fixed.safetensors
MODEL C: AOM3-Silicon29sum50
Triple Sum merge - ✔️ Use MBW
Alpha: 1
0,0.8,0.8,0.8,0.6,0.6,0.6,0.3,0.3,0.3,0.4,0.4,0.4,0.4,0.2,0.2,0.1,0.1,0.1,0.1,0,0,0,0,0,0
Beta: 1
0,1,1,1,0.8,0.6,0.4,0.1,0.1,0.1,0.3,0.3,0.3,0.3,0.1,0.1,0.1,0.2,0.2,0.2,0,0,0,0,0,0
Model: EtherOne-p2
▼
MODEL A: EtherOne-p2
MODEL B: lunarRadianceMix_luasNight.safetensors
Weight Sum merge - ✔️ Use MBW
Alpha: 1
0,0.2,0.2,0.2,0.2,0.1,0,0,0,0.1,0.1,0.1,0.1,0.1,0,0,0,0,0,0,0.2,0.2,0.2,0.2,0.2,0.2
Final Model: EtherOne-v2
▼ LORA Mix
Iridescence:
0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0.1,0.2,0.2,0.2,0.2,0.2,0.1,0.1,0.1
sukebenishiki_e08:
0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0.5,0.6,0.5,0.4,0.3,0.3,0.2,0.2,0.2
artist-rasra:
0,0.3,0.3,0.3,0.3,0.2,0.1,0,0,0,0,0,0,0,0,0,0,0.4,0.4,0.5,0.3,0.3,0.3,0.2,0.2,0.2
rabbit-(wlsdnjs950)-1037it-novae:
0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0.2,0.3,0.3,0.2,0,0,0,0,0
Models Used:
*EtherRealMix-3x1-FINAL-fp16.safetensors
D173E91970FE9E4A7003D87F059FDD5F9BFCBDA3C2EA99B1FB10D44A3C348D9D
*Counterfeit-V3.0_fp16.safetensors
CBFBA64E662370F59D4AA2AA69BF16749FCE93846CCCE20506AEE5DF01169859
*watercolorOilPainting_v10.safetensors
0CBEA7D5041FC10F273D148A89D56F297D5919059F864040F17F3ABF709B410B
*Basil_mix_fixed.safetensors
0FF127093F5BE455057742C40CEF578407B6933F240EE8DC5ED0F3061196FB38
*AOM3.safetensors
D124FC18F0232D7F0A2A70358CDB1288AF9E1EE8596200F50F0936BE59514F6D
*Silicon29-dark.safetensors
de329f31ad222a0d765897aa38796811276691288b104099523649bdf1ef9708
*lunarRadianceMix_luasNight.safetensors
1E50E77B17170BE3AA87B73814798A51617774DFF5FE1F884773686F7B98B258
1. The rights to reposted models belong to original creators.
2. Original creators should contact SeaArt.AI staff through official channels to claim their models. We are committed to protecting every creator's rights.