SD

ChillyMix (early access)

#Asian

ChillyMix (early access) - AI Model cover image
Uploaded at
May 28, 2023, 8:20 PM
Uses
1.33k
Reviews
Dissatisfied
Permissions
  • Allow image generation & sharing
  • Allow users to download this model
  • Allow commercial use of generated images

Description

This is another attempt to make a free/open version of the ChilloutMix.

There are chilled_re_generic v1/v2 and chilled_remix, but they are not good enough to generate the desired results.

Merged models are the following: (all models are merged with the SuperMerger extension)

V2 version only: use the chillyMixBase - include pyro's_pov models (results NSFW fix)

analogDiffusion_10 (Fp16)

F222 (Fp32) see also F222 at https://rentry.co/sdmodels

r34_e4 (Fp16) - from https://rentry.co/sdmodels

neverendingDreamNED_noVae (Fp32) by @Lykon

muse_v1 (Fp32) by @9muses_arts

Merge Recipe The following three steps are modified recipe based on https://rentry.org/hdgrecipes#instructions_1 (use the Analog Diffusion model instead of the Novel AI model)

(This recipe is for the V1 version. V2 recipe uses the chillyMixBase )

analogDiffusion_10 x 0.95 + v1-5-pruned-emaonly x 0.05 -> MERGE1 (fp16)

MERGE1 + (f222-Fp16 - v1-5-pruned-emaonly) x 1 : Add difference -> MERGE2 (fp16)

MERGE2 x 0.75 + r34_e4-Fp16 x 0.25 -> MIX_BASE (fp16)

These three models are used by PoV Skin Texture - r34 Lucid Black (it is used by the ChilloutMix as described it's description.)

In this recipe, you can get the MIX_BASE but it is not good to use it alone. It sometimes produces overfitted results. but simply merge with a good model like the NED to fix this issue. and it could maintain some chilloutmix's own property.)

add the NED(NeverEndingDream - NO VAE version) to the MIX_BASE:

(for V2 recipe case, MIX_BASE == chillyMixBase)

MIX_GOOD = MIX_BASE x 0.25 + neverendingDreamNED_noVae x 0.75 (fp16)

Finally, mix the Muse V1-fixed (fixed by the model toolkit extension) for a realistic model. (Muse v1 seems very similar to the ChilloutMix.)

MIX_GOOD + museV1_v1-fixed are merged with the SuperMerger with the following settings:

MIX_GOOD x (1-alpha) + museV1_v1-prunedFp16-fixed x alpha (0.0,1.0,0.925925925925926,0.740740740740741,0.7,0.7,0.7,0.7,0.7,0.7,0.7,0.740740740740741,0.925925925925926,1.0,0.925925925925926,0.740740740740741,0.7,0.7,0.7,0.7,0.7,0.7,0.7,0.740740740740741,0.925925925925926,1.0)

Selected alpha MBW(Merge Block Weights) are modified R_SMOOTHSTEP/4

(MBWs are selected for more realistic and ChilloutMix like. I'm not sure that MBWs are optimal but it seems best for my use cases.)

Recommends All posted images use "DDetailer script extension" or "After Detailer" for beautiful faces.

DPM++ 2M Karras sampling method.

"EasyNegative" "badhandv4" (or "bad-hand-5") "ng_deepnegative_v1_75t" embeddings recommended in the negative prompt.

Set Clip skip: 2, ENSD: 31337

Useful AUTOMATIC1111's webui extensions SuperMerger extension

Lora Block Weight extension

webui model toolkit to examine Loras and models.

civitai extension to add several useful features for civitai.

ChangeLog 2023/05/29 : V2 version preview - use the chillyMixBase: it includes pyro's pov checkpoint and results in NSFW enhancement and more realistic skin.

License All used models here have "Have different permissions when sharing merges" license permission, so I do not try to add any restriction on it.

CreativeML-Open Rail-M

This model permits users to:

✔Use the model without crediting the creator

✔Sell images they generate

✔Run on services that generate images for money

✔Share merges using this model

✔Sell this model or merges using this model

✔Have different permissions when sharing merges

Comments