chillyMixBase版本chillyMixBase-V0-Fp16 (ID: 80454)

This is the chillyMixBaseV0 and chillyMixBaseV1-NED (as an example)

chillyMixBaseV0 used by chillySauceMix and chillyMix V2.

If you are not interested in checkpoint merge or development, you don't have to download it.

• chillyMixBaseV0 checkpoint alone can produce too many overfitting results.

  • it seems that the F222 is broken: the M00 block of the F222 is broken!

• for this reason, chillyMixBaseV0 produces some erroneous results such as:

  • resulting images could be too much exaggerated

  • torn or mixed or broken

  • these broken images might be harmful to children or pregnant women ?

• For this reason, it is not used alone, but rather mixed with other models. chillyMixBaseV1-NED is an example (see below)

chillyMixBaseV0 Merge Recipe

please see also chillySauceMix(by @mixboy) description

All models have merged with the SuperMerger extension

• pyro's_pov models - (pyro_pov_a + pyro_pov_b)

• analogDiffusion_10 (Fp16)

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

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

See also merging recipes of legacy merged models https://rentry.org/hdgrecipes

Merging Steps are modified recipe of https://rentry.org/hdgrecipes#instructions_4

• pyro_pov_a + pyro_pov_b = pyAB

• analogDiffusion_10 x 0.85 + pyAB x 0.15 = MERGE1

• MERGE1 x 0.95 + v1-5-pruned-emaonly x 0.05 = MERGE1Fix

• MERGE1Fix+ (F222_v1 - v1-5-pruned-emaonly) x 1 = MERGE2

• MERGE2 x 0.75 + r34_e4 x 0.25 = MERGE_BASE

Now we can get MERGE_BASE named ChillyMixBaseV0.

Cosine Similarity

What about cosine similarity? some cosine similarities have been checked recently and found out that the results are interesting.

Output_blocks cosine similarities

base: chilloutmix_NiPrunedFp32Fix

| OUT03 | OUT04 | OUT05 |OUT06   |OUT07   |OUT08   |OUT09   | OUT10  | OUT11 |

|91.6694|91.3029|93.1590|93.6105 |94.5325 |96.1938 |99.4821 |99.6246 |99.9055|

total output blocks similarity: 95.4978%

OUT09, OUT10, OUT11 is ~99%!!

Input_blocks cosine similarities

How about input_blocks?

|   IN01   |   IN02   |   IN04   |   IN05   |   IN07   |   IN08   |   MID0   |

| 93.2390% | 95.7006% | 90.4897% | 92.1093% | 83.8271% | 85.8409% | 73.0998% |

total input blocks similarity: 87.7581% (MID0 seems too low relatively)

based on this information, output_blocks are fair enough but input_blocks seem bad. simply replace the input blocks with the input_blocks of good checkpoint, it will work great.

Why use it?

If you're interested in developing checkpoint files, you might be wondering how a lot of checkpoint files were created, or what the original source of the merged checkpoint files was.

However, many checkpoints are being merged and distributed, either not properly identifying what the original source is, or ignoring the licenses of the original source. This can cause potential licensing problems, and there are checkpoints that risk being retired due to licensing issues in the near future.

To resolve these issues, it is very important to use freely available checkpoints with a free/open license.

Interestingly, a bad checkpoint file + a bad checkpoint file + a bad checkpoint file and so on can get good results. so you can merge this bad checkpoint as one of your basis models.

When trying to merge this model?

• my model needs more realistic skin (F222 and pyro_pov contribution)

• my model needs minimal nsfw capability. (F222 and pyro_pov contribution)

• my model needs anatomic improvement (F222 contribution)

Examples

chillySauceMix is an example.

• neverEndingDream (NED) is a 2.5D style but more anime-like model.

• mixing ChillyMixBaseV0 could add a more realistic skin and make less anime style.

• ChillyMixBaseV0 x (1 - alpha) + neverendingDreamNED_noVae x alpha

  • chillySauceMix is alpha == 0.6

  • for chillyMix V2 case, alpha == 0.75 used before merging Muse_v1.

• ChillyMixBase_V1-NED = ChillyMixBaseV0 + NED's input_blocks

  : replace all input blocks with NED's input blocks. (this is the recipe of V1-NED)

  • using SuperMerger extension :

    • ChillyMixBaseV0 (1-alpha) + NED x alpha

    • alpha = (1,1,1,1,1,1,1,1,1,1,1,1,1,1,0,0,0,0,0,0,0,0,0,0,0,0)

      • BASE == 1 (include NED's text encoder) for practical reasons.

      • you can exclude text encoder by setting "BASE == 0"

NeverEnding Dream (NED)'s cosine similarity

NED's cosine similarity has been checked between the chilloutmix

base: chilloutmix_NiPrunedFp32Fix

|  OUT03   |  OUT04   |  OUT05   |  OUT06   |  OUT07   |  OUT08   |  OUT09   |  OUT10   |  OUT11   |

| 91.5396% | 91.6970% | 91.1852% | 92.7715% | 94.1585% | 96.0288% | 99.3220% | 99.1343% | 99.9986% |

Total output blocks similarity - 95.0928%

It is not a surprise that the similarity seems high. but OUT03~OUT08 mean is 92.89676% and it is not so high.

base: chilloutmix_NiPrunedFp32Fix.safetensors

|   IN01   |   IN02   |   IN04   |   IN05   |   IN07   |   IN08   |   MI00   |

| 99.8909% | 99.2414% | 98.8610% | 99.2940% | 98.3178% | 98.3939% | 96.9688% |

total input blocks similarity - 98.7097%

why does this show high similarity?

• IN00, IN03, IN06, IN09, IN10, and IN11 blocks are omitted in this similarity check. so the similarity of the input blocks is over estimated.

• and, ... probably NED might be based on some older(not restrictive license) version of chilloutmix or chilloutmix's source models.

the follows are some similarities of the input blocks between chilloutmix.

• BRAv5 - 96.8725%

• dreamshaper v5 - 89.4525%

• dreamshaper v6 - 91.0372%

• muse_v1 - 98.1695%

• henmixreal_v40 - 96.6106%

• diamondcoalmix - 91.0779%

• epicrealism_newAge - 87.0339%

• epicrealism_newEra - 86.6671%

• sxd_10Pruned - 67.8242%

• BracingEvoMix_v1 - 93.4567%

• Evt_V4_e04_ema - 88.2055%

How to merge this model?

• as you already know, use Checkpoint Merger (included by default) or SuperMerger extension (recommend)

• you can try to merge it by 0.1 weight ratio like as:

  • your_model x 0.9 + ChillyMixBaseV0. x 0.1

  • reduce/increase the ratio and find the best result. (SuperMerger can use X/Y/Z plot)

• or like as ChillyMixBase_V1-NED you can merge specific block levels.

License

Like as ChillyMix, its license has no commercial restriction. it is open and free under CreativeML-Open Rail-M license:

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

Donation

Feel free to donate to me if you want to support my work! at ko-fi https://ko-fi.com/mixboy

描述:

2023/05/25: released

训练词语: