Bracingevomix

🚀 BracingEvoMix

BracingEvoMix is a text-to-image model based on Stable Diffusion. It addresses the license risks in traditional merged models and provides high - quality image generation capabilities, especially excelling in generating East Asian faces.

📄 License

License: CreativeML Open RAIL - M
Additional Copyright: sazyou_roukaku (TwitterID @sazyou_roukaku) as of May 31, 2023

This model is licensed under 'CreativeML Open RAIL - M', and there are no changes to the license itself. However, the name of sazyou_roukaku has been added as an additional author. (On June 10, the name was changed due to a Twitter name change. There is no change in the license as it still uses sazyou_roukaku.)

As stated in 'CreativeML Open RAIL - M', regarding the products generated using this model, we have no involvement except for the cases in Usage Restriction A of the license. The use for criminal purposes and specific professional uses such as medical images are prohibited under Usage Restriction A. Please make sure to check the license before use. Also, we assume no liability. Please use the model with the understanding that we are exempted from liability.

On October 1, 2023, BracingEvoMix_v2 was publicly released.

BracingEvoMix_v2

CLIP setting/clip skip: 2 Recommended negative prompt base:

(worst quality:2),(low quality:1,4),(undressing:1.5),(manicure:1.5),(long neck:2),lip,make up,(depth of field, bokeh, blurry, blurry background:1.4)

The part (depth of field, bokeh, blurry, blurry background:1.4) is for when you want a sharp background.

The update of the BracingEvoMix series will stop at v2 for now. The reasons are as follows: Many users who perform additional training are migrating to SDXL, and there are limitations in improving performance with models that meet the selection criteria. Since sxd, which is the base for some lower - body parts, is a relatively early model, there is a high possibility that the improvement in performance will be within the margin of error compared to the effort. Also, the restrictions on available models are too strict to increase the variety of clothing - related types.

However, depending on the demand for SD1.x, I may conduct training myself and release a new series. Nevertheless, I first want to observe the trends in SDXL and the possibility of companies releasing user - friendly models and then respond accordingly.

Restrictions

Note that the sale of the model itself and its use in commercial image - generation services cannot be restricted in principle under the 'CreativeML Open RAIL - M' license unless otherwise stated in Usage Restriction A. The 'OK' indication in the Civitai restriction table is only for considering the burden on the merger, and we do not actively recommend it. We will not assume any liability for any problems that may arise. Please keep this in mind.

Model Evaluation Tables

BracingEvoMix_v2

It is the legitimate successor to the base model 'BracingEvoMix_v1'. The unevenness in saturation and brightness has been improved, and the finger accuracy has also increased. There are also enhancements for the background and horizontal resolution. The output unevenness has decreased, so I think its versatility has increased.

BracingEvoMix_v1

It is the base model. It can represent Asian faces from OpenBra and has very high realism. However, there are unevenness in saturation and brightness. It is also strong in emotional expression, so if you can control it, it has a wide range of expression. But it also has many weaknesses. It combines the advantages and disadvantages of each model used as materials, so it is suitable for experts who can master it.

BracingEvoMix_Another

It is an alternative model. By changing the merging method and deepening the mixing, the obvious changes in saturation and brightness have been stabilized. As a result, the BRA face has become less prominent, and it is more in the direction of a general Asian - face - based merged model. The stability of fingers, especially the beautiful expression of nails and fingertips, is comparable to the latest top - tier models. However, the response to emotional expression has weakened due to the reduction of weaknesses. Note that for general clothing with large breasts, the likelihood of accidental exposure is not much different.

BracingEvoMix_Fast

It is a prototype model based on a more 'chilled_remix' - like arrangement of the alternative model. It is easier to create a more perfectly structured face, and although the effect expression is not as good as 'chilled_remix', it is designed to be a bit more accessible. However, since it does not deviate from the photorealistic area, it does not have as much freedom as 'chilled_remix'.

List of Merged Models

[BracingEvoMix_v2]

• OpenBra
  ©BanKai @PleaseBanKai
• dreamshaper_6BakedVae
  (https://civitai.com/models/4384) ©Lykon
• epicrealism_newEra
• epicrealism_pureEvolutionV5
  (https://civitai.com/models/25694) ©epinikion
• diamondCoalMix_diamondCoalv2
  (https://civitai.com/models/41415) ©EnthusiastAI
• sxd_v10
  (https://civitai.com/models/1169) ©izuek
• Evt_V4_e04_ema
  (https://huggingface.co/haor/Evt_V4-preview) ©haor
• bp_mk5
  (https://huggingface.co/Crosstyan/BPModel) ©Crosstyan

[BracingEvoMix_v1]

• OpenBraβ
• OpenBra
  ©BanKai @PleaseBanKai
• dreamshaper_5Bakedvae
• dreamshaper_6BakedVae
  (https://civitai.com/models/4384) ©Lykon
• epicrealism_newAge
• epicrealism_newEra
  (https://civitai.com/models/25694) ©epinikion
• diamondCoalMix_diamondCoalv2
  (https://civitai.com/models/41415) ©EnthusiastAI
• sxd_v10
  (https://civitai.com/models/1169) ©izuek
• Evt_V4_e04_ema
  (https://huggingface.co/haor/Evt_V4-preview) ©haor

[BracingEvoMix_Another]

• OpenBra
  ©BanKai @PleaseBanKai
• dreamshaper_6BakedVae
  (https://civitai.com/models/4384) ©Lykon
• epicrealism_newEra
  (https://civitai.com/models/25694) ©epinikion
• sxd_v10
  (https://civitai.com/models/1169) ©izuek
• Evt_V4_e04_ema
  (https://huggingface.co/haor/Evt_V4-preview) ©haor

[BracingEvoMix_Fast]

• OpenBra
  ©BanKai @PleaseBanKai
• dreamshaper_6BakedVae
  (https://civitai.com/models/4384) ©Lykon
• epicrealism_newEra
  (https://civitai.com/models/25694) ©epinikion
• Evt_V4_e04_ema
  (https://huggingface.co/haor/Evt_V4-preview) ©haor
• bp_mk5
  (https://huggingface.co/Crosstyan/BPModel) ©Crosstyan

Recommended Settings

CLIP setting/clip skip: 2

There is not much difference in feeling whether or not to include a negative TI of the badhand series. This model recommends using 'EasyNegative' or 'BadBras'. 'EasyNegative v2' is not recommended because the rate of human - body structure breakdown is higher compared to the above.

This model is quite strong in responding to natural - language text prompts. However, for detailed face - shaping other than scenarios, it is my style to specify with word prompts according to personal preferences. You can also use word - only prompts without any problem, so please use the style that is most convenient for you.

I don't feel the effectiveness of quality prompts like 'high quality'. 'Masterpiece' gives a feeling of changing the face - shaping, but it is only slightly effective in improving quality.

However, 'high resolution' is effective for the background and texture. There are various expressions like 'high res' and 'Hires', but I trust 'high resolution' the most.

Also, consider using natural - language words like 'girl', 'woman', or 'young woman' instead of WD - style prompts like '1girl'. Reducing the CGF scale to about 5 can also be a useful technique to weaken the overall contrast.

💡 Usage Tip

If you feel that the skin color is too dark, adding (fair skin:1.2) will make it lighter. Please use it in combination with the above settings.

Negative prompt base:

EasyNegative,(worst quality:2),(low quality:1.4),(undressing:1.5), (disheveled clothes:1.4),(manicure:1.2),(nipple:1.2),(long neck:2),

FAQ

Q1: What is BracingEvoMix?

A1: Traditional merged models may contain NAIleak models or DreamlikeLicense, which can cause difficulties in making decisions for full - scale business use. BracingEvoMix is a model merged based on OpenBraβ and OpenBra, which were born as a result of direct discussions with Mr. BanKai, the developer of BRA, and donations from volunteers. All other models are also training models, which significantly reduces the probability of including risky models compared to combinations in merged models.

It is considered the lowest - risk model among merged models that can generate East Asian faces. However, even if they are called training models, we cannot know the details of their contents, so the risk is not zero. Although the selection is also based on the cosine similarity rate, please note that we cannot guarantee that there is no contamination.

Q2: Is it impossible to have a completely uncontaminated model?

A2: If only additional training is performed, data bias will occur, and there will inevitably be unevenness in capabilities, including the background. It is almost impossible for an individual to create a complete training model, so merging is necessary. However, since we are not experts, we cannot analyze the contents of these training models.

It is possible to remove data that matches the data of a specific model through differential extraction. However, taking the difference itself may meet the conditions for a derivative model. Also, it is necessary to download the NAIleak model itself, which is also difficult.

Q3: What are the selection criteria for each training model?

A3:

• sxd_v10: It is an NSFW model. While including an NSFW model can cause problems such as an increased likelihood of undressing, it is used to reinforce the slightly lower training level of the lower body, which is one of the weaknesses of OpenBra, and to supplement various pose - training data. Since increasing the proportion too much will make exposure more likely, only the minimum amount is merged. Sxd is selected because, although it is a relatively early model, it has a large amount of training and no restrictions.
• dreamshaper_5Bakedvae/dreamshaper_6BakedVae: Dreamshaper is a well - known training model used in commercial image - generation services. It is adopted for enhancing diversity because it has a large amount of additional training data and high data diversity.
• epicrealism_newAge/epicrealism_newEra: It is considered to have the strongest specifications among current training models. It is adopted for background reinforcement and its high capabilities.
• diamondcoalmix_diamondcoalv2: It is a training model with data on East Asian faces. The author has created many LoRAs, and based on the cosine value, it is judged to be problem - free. It is adopted for reinforcing East Asian data.
• Evt_V4_e04_ema: It is a model created by training an illustration - training model that publicly claims not to contain NAIleak data called ACertainty. Merging an illustration model can affect face data regardless of race and age, so it is adopted to reduce the rate of portrait - right infringement and to create a face shape preferred by the Japanese.

⚠️ Important Note

Although ACertainty may have distilled data from NOVEL AI, it does not conflict with patent law, so it is considered to be problem - free. ACertainty:

• https://huggingface.co/JosephusCheung/ACertainty
• https://huggingface.co/JosephusCheung/ASimilarityCalculatior

• bp_mk5: It is a training model based on ACertainty. BracingEvoMix_Fast uses it for rebalancing.

Q4: The quality of fingers in the original version is lower compared to models like chilled_remix.

A4: Since the models used are limited to training models that meet certain criteria, the data stability is lower compared to the latest models. However, the fingers are still output sharper than in the initial models, and I feel that the average output is comparable to models around mid - March.

I'm sorry, but please also consider using controlnet for control.

Note (June 14): Another and Fast have improved the finger accuracy, and I think the output is comparable to the average of the latest photo models as of June.

Q5: Are there any problems or contradictions in the current restrictions?

A5: 'diamondCoalMix_diamondCoalv2', 'dreamshaper_5Bakedvae', and 'dreamshaper_6BakedVae' can lift the restriction of 'Have different permissions when sharing merges'. Since the other models have no restrictions, all models are publicly released without restrictions this time.

Note that if there are any changes in the license or restrictions of the merged models, the release is based on the license and restrictions as of May 31, so it complies with 'creativeml - openrail - m'. I have stored screenshots of the relevant models in MergeModel_LicenseSS_v1.

If a serious problem occurs with the merged models, I may stop the public release of the model and call for a halt to its use. However, I will not set any additional restrictions on my side.