metadata

license: creativeml-openrail-m
language:
  - en

ether-blu-mix-ebm

TOU:

Please don't sell this model merge.

Please don't place it on paid generation services.

You are free to sell the images you create with it.

You are free to merge this into other models.

🔹☁️ Hello, this is Ether Blu Mix ☁️🔹

A simple mix aimed at providing a semi-realistic anime style while retaining the knowledge trained on Danbooru-styled tags. Though its base model, AOM3, is fit for anime styled illustrations it's capable of other aesthetics with the correct style of prompting.

The mix has strong biases towards females and feminine faces, likely due to their prominence within the models as well as the anime aesthetic leaning towards feminine\androgynous faces.

Version 2 Update:

With version two, problems such as unprompted NSFW generations as well as poor quality hands and fingers have been reduced. The tradeoff seems to be a slight loss in realism.

If you prefer the realistic look of v1, please add "realistic" to the start of your prompts for a similar look in v2.

Placing "NSFW" into the negative prompt is still recommended if you're concerned about occasional nudity or more suggestive generations as the mix is still biased towards suggestive and NSFW imagery.

Since "NSFW" can greatly alter the output, I recommend experimenting with adding or removing it.

From my short amount of experimentation: including NSFW tends to produce more stylized generations, while excluding it tends to introduce more realistic elements. (?)

☄️ Generation Recommendations

Use whatever sampler, steps, cfg you prefer.

#*V2 Update:

With version two the model has lost some realism in exchange for a slightly more stylized aesthetic. If you prefer the realistic look of v1 please prepend "realistic" to achieve a similar style.

Sample images were generated with
- Sampler: DPM++ 2M Karras
- Steps: 28
- CFG: 7
- Clip Skip: 1

The mix is still biased towards suggestive and NSFW imagery due to the models included. See "* Version 2 Update:" above for a more detailed explanation.

Negative Prompt Example:
- "NSFW, (worst quality, low quality:1.3), watermark, signature"

*V2 Update: NSFW is now optional

The sample images were generated using Waifu Diffusion VAE kl-f8-anime2. This VAE is not mandatory. Please use whichever VAE you prefer.

Most 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

Prompting Advice:

Prompt however you like. The model was merged with both Danbooru-style tag prompting as well as natural language in mind. You're free to mix and match prompting styles, the model should understand it.

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.

I personally like to use short prompts mixing prose and danbooru-styled tags. Experiment and give the AI some freedom to interpret your ideas!

General guidelines:

Prompt in the style of anime\danbooru-style models, "masterpiece, best quality, ..... "
- More anime style images
Prompt in the style of photography\non-anime models, "professional photograph of ..... "
- More realistic images

📘 Model Merging Info - V2

I have no training or real knowledge of machine learning. I merged the model using the information provided on WarriorMama777's AOM model page. I very arbitrarily used the UBM weights given to merge and create AOM3 and its derivatives and modified the values with trial and error until I was visually happy with it.

I utilized the SuperMerger extension for AUTO1111 by hako-mikan.

▼ Merging

MODEL A: AOM3.safetensors

MODEL B: pastelmix-better-vae-fp16.safetensors

  Weight Sum merge - Use MBW

  Alpha: 1
  0,0.7,0.7,0.7,0.7,0.7,0,0,0,0,0.7,0.7,0.7,0.7,0.7,0.7,0.7,0.7,0.7,0.7,0.7,0.7,0.7,0.7,0,0

      Model: AOM3-PastelSumS2-07

▼

MODEL A: AOM3-PastelSumS2-07

MODEL B: deliberate_v2.ckpt

  Weight Sum merge - Use MBW

  Alpha: 1
  0,0.3,0.3,0.3,0.3,0.3,0,0,0,0,0.3,0.3,0.3,0.3,0.3,0.3,0.3,0.3,0.3,0.3,0.3,0.3,0.3,0.3,0,0

      Model: AOM3-PastelSumS2-07-DelibV2SumS2-03

▼

MODEL A: AOM3-PastelSumS2-07-DelibV2SumS2-03

MODEL B: chilloutmix_NiPrunedFp16Fix.safetensors

  Weight Sum merge - Use MBW

  Alpha: 1
  0,0.2,0.2,0.2,0.2,0.2,0,0,0,0,0.2,0.1,0.2,0.2,0.2,0.2,0.2,0.1,0.1,0.1,0.2,0.2,0.1,0.2,0.2,0.2

      Final Model: EBM-v2

Models Used:

AOM3.safetensors
  D124FC18F0232D7F0A2A70358CDB1288AF9E1EE8596200F50F0936BE59514F6D

chilloutmix_NiPrunedFp16Fix.safetensors
  59FFE2243A25C9FE137D590EB3C5C3D3273F1B4C86252DA11BBDC9568773DA0C

deliberate_v2.ckpt
  9ABA26ABDFCD46073E0A1D42027A3A3BCC969F562D58A03637BF0A0DED6586C9

pastelmix-better-vae-fp16.safetensors
  D01A68AE76F97506363F387F5F28BB564AD9E20924844FD5945E600B72D39E79

💙 Credits:

Thank you to WarriorMama777 for providing AOM and other various mixes as well as detailing your workflow and inspiring 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 the creator of ChilloutMix.

Thank you to XpucT for creating Deliberate.

Thank you to andite for Pastel-Mix.

Thank you to AUTO1111 for creating the web UI everyone uses.

Thank you to StabilityAI for starting everything with Stable Diffusion.

Thank you to the entire SD community for continuing to openly share and create.