santa hat
deerdeer nosedeer glow
Sign In

Ether Blu Mix

6.0k
32k
17k
131
0
Verified:
SafeTensor
Type
Checkpoint Merge
Stats
14,219
15,174
Uploaded
Jul 10, 2023
Base Model
SD 1.5
Usage Tips
Clip Skip: 1
Hash
AutoV2
6965F33B20
gamerdan69's Avatar
gamerdan69
Joined Dec 8, 2023
0
0
0
0
0

🔹☁️ Hello, this is Ether Blu Mix ☁️🔹

Ether Blu Mix is a stylized anime model specialized in creating character illustrations with vibrant colors, expressive backgrounds and a painterly feel. Based on AOM3 and Counterfeit 3, I consider this my interpretation of a pastel-mix style of model.

Please Share Your Cool Creations Below!

Meet "Ether Real Mix" - EBM's big sister.

* Version 5 Update:

Ether Blu Mix 5 looks to further enhance the vibrant, painterly look with my first attempt at LORA mixing. There's also a lot more Counterfeit 3 allowing for more compositional variety.

EBM 4 did a lot of things right; as a base it created detailed, vibrant images reminiscent of a classic anime style thanks to YOZORA. Though, personally, it seemed too biased at depicting characters at a somewhat close distance, typically creating waist-high portraits. Adding more weight to Counterfeit 3 should hopefully add more compositional variety, while slightly reducing YOZORA and Color Box's weights to maintain EBM 4's style. Basil Mix was added as a supplementary realistic model in tandem with AyoniMix. It's a personal favorite that plays nice with anime (booru tag) style models.

EBM 5 is my first attempt at LORA mixing. I was initially hesitant at merging LORA's to checkpoints as they might introduce unwanted bias, since LORA's tend to be trained on small datasets. After plenty of testing, as well as learning from Luna's documentation of Kenshi and Aurora, I applied a light mix of LORA's to help enhance the vibrant, painterly style, while also improving background expression.

Your prompts might react differently as the base of the model is no longer just AOM3, so please experiment and have fun!

Placing "NSFW" into the negative prompt is recommended if you're concerned about suggestive generations or instances of unprompted nudity. Due to AOM3 this mix is might be biased towards suggestive and NSFW imagery.

* Huggingface Archive

☄️ Generation Recommendations

Use whatever sampler, steps, CFG 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)"

Waifu Diffusion VAE kl-f8-anime2

Strongly recommended as the model was tuned with this VAE's 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.

  • 4x-foolhardy-Remarci

  • Upscale by: 2

  • Denoising strength: 0.45 - 0.55

🖊️ Prompting Advice

As Ether Blu 5 leans heavier into the anime aesthetic 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

📘 Model Merging Info - v5

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: Counterfeit-V3.0_fp16.safetensors

Add Sum merge - Don't Use MBW

  • Alpha: 0.5

    • Model: AOM3-Counter3Sum-05

  • MODEL A: AOM3-Counter3Sum-05

  • MODEL B: nightSkyYOZORAStyle_yozoraV1PurnedFp16

Add Sum merge - Don't Use MBW

  • Alpha: 0.4

    • Model: AOM3-Counter3Sum-05-YozoraSum-04

  • MODEL A: AOM3-Counter3Sum-05-YozoraSum-04

  • MODEL B: colorBoxModel_colorBOX.safetensors

  • MODEL C: AOM3-Counter3Sum-05

Add Difference merge - Don't Use MBW

  • Alpha: 0.2

    • Model: AOM3-Counter3Sum-05-YozoraSum-04-ColorBoxDiff-02

  • MODEL A: AOM3-Counter3Sum-05-YozoraSum-04-ColorBoxDiff-02

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

  • MODEL C: Basil_mix_fixed.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

  • Beta: 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

    • Model: AOM3-Counter3Sum-05-YozoraSum-04-ColorBoxDiff-02-PastelSumS1-02-BasilSumS1-02

  • MODEL A: AOM3-Counter3Sum-05-YozoraSum-04-ColorBoxDiff-02-PastelSumS1-02-BasilSumS1-02

  • MODEL B: ayonimix_V5.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: EtherBluMix-5

LORA Mix

  • xlimo768: 0.25

  • dpep4-768-000004: 0.3

  • sttabi_v1.4-04: 0.05

  • Jordan_3: 0.2

  • lightingv2: 0.4

Models Used:

*AOM3.safetensors
D124FC18F0232D7F0A2A70358CDB1288AF9E1EE8596200F50F0936BE59514F6D

*Counterfeit-V3.0_fp16.safetensors
CBFBA64E662370F59D4AA2AA69BF16749FCE93846CCCE20506AEE5DF01169859

*pastelmix-better-vae-fp16.safetensors
D01A68AE76F97506363F387F5F28BB564AD9E20924844FD5945E600B72D39E79

*colorBoxModel_colorBOX.safetensors
93A20525F51798309AF22AC6E8285E61CF563A88F1EE00AA9F544CA939E63181

*nightSkyYOZORAStyle_yozoraV1PurnedFp16
E7BF829CFFD9EA034DA3B59FA13421F52AA4E4ED74FCBBEEFBF9CFB2F97C329E

*ayonimix_V5.safetensors
B46B04432B070FD88D69161B62C43EDCD4340DFC27FFB2E5FE7EDA64950C7E34

*Basil_mix_fixed.safetensors
0FF127093F5BE455057742C40CEF578407B6933F240EE8DC5ED0F3061196FB38

💙 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 Luna for sharing your process on creating Kenshi and Aurora. It deepened my understanding of model making as well as inspired me to mix LORA's.

Thank you to hako-mikan for creating the SuperMerger extension allowing a quicker workflow for merging.

Thank you to gsdf, rqdwdw for creating Counterfeit.

Thank you to YozoRaAru for Night Sky YOZORA and Color Box.

Thank you to andite for Pastel-Mix.

Thank you to Ayoni for AyoniMix.

Thank you to nuigurumi for BasilMix.

Thank you to CTD for Dpep4, Xlimo and lightingv2.

Thank you to dolphinz for Sttabi.

Thank you to SatyamSSJ10 for Jordan 3.

Thank you to AUTO1111 for creating the web UI.

Thank you to StabilityAI for starting everything with Stable Diffusion.

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