[ Stable Diffusion web UI ] Introduction of SDXL1.0 Recommended Checkpoint Models
![[ Stable Diffusion web UI ] Introduction of SDXL1.0 Recommended Checkpoint Models featured Image](/_next/image?url=https%3A%2F%2Fdca.data-hub-center.com%2Fcontent%2Fuploads%2F2024%2F06%2Feye_catch_en_sdxl-recommended-checkpoint-models.jpg&w=3840&q=80)
This article introduces recommended SDXL checkpoint models for use with Stable Diffusion web UI. All models presented in this article are SDXL 1.0 models, and you do not need to use Refiner. The SDXL models can be generated in styles ranging from realistic to illustrative, but they are grouped according to their main style. Let’s take a look at the checkpoint models for illustration, 2.5D, and photorealistic styles with reference images.
Please refer to this article for instructions on how to install the model.
Illustrative checkpoint models
First, let me introduce you to SDXL’s illustration-based checkpoint models. As you can see from the sample images, we have adapted only each model to make it as easy as possible to compare them in four patterns: medieval, maid, cyberpunk, and snow scene.
Animagine XL V3.1
- Creator: 🔗CagliostroLab
- Download: 6.46 GB
- Upload: Mar 21, 2024
- Data Format: Full Model fp16
CagliostroLab’s “Animagine XL V3.1” is a popular model that has been downloaded more than 140,000 times on Cibitai.
Usage Guidelines
Animagine XL allows you to control the quality and style of your illustrations by using its own tags. Following these rules will help you produce high quality illustrations.
Tag Ordering1girl/1boy, character name, from what series, everything else in any order.Special Tags:
Quality Modifiers: You can control the quality of the illustration. Low quality tags can be effectively placed in negative prompts.
Quality Modifier Score Criterion
masterpiece > 95%
best quality > 85% & ≤ 95%
great quality > 75% & ≤ 85%
good quality > 50% & ≤ 75%
normal quality > 25% & ≤ 50%
low quality > 10% & ≤ 25%
worst quality ≤ 10%Rating Modifiers:
NSFW stands for Not Safe For Work and means not suitable for workplace viewing. If you want to generate illustrations for all ages, use the safe tag.
Rating Modifier Rating Criterion
safe General
sensitive Sensitive
nsfw Questionable
explicit, nsfw ExplicitYear Modifier: You can specify the age of the illustration. You can reproduce the illustration style according to the age of the illustration.
Year Tag Year Range
newest 2021 ~ 2024
recent 2018 ~ 2020
mid 2015 ~ 2017
early 2011 ~ 2014
oldest 2005 ~ 2010Aesthetic Tags: The aesthetic tag is similar to the quality tag but allows control over the artistic quality of the illustration. They use the model 🔗shadowlilac/aesthetic-shadow-v2 to evaluate the aesthetic value of content prior to learning.
Aesthetic Tag Score Range
very aesthetic > 0.71
aesthetic > 0.45 & < 0.71
displeasing > 0.27 & < 0.45
very displeasing ≤ 0.27Recommended setting
Negative Promptnsfw, lowres, (bad), text, error, fewer, extra, missing, worst quality, jpeg artifacts, low quality, watermark, unfinished, displeasing, oldest, early, chromatic aberration, signature, extra digits, artistic error, username, scan, [abstract]masterpiece, best quality, very aesthetic, absurdresRecommended VAE: sdxl_vae
Prompt Style: Animagine XL recommends Danbooru style tags.For more information, see HuggingFace's Usage Guidlines.
Anything XL
- Creator: 🔗Yuno779
- Download: 6.46 GB
- Upload: Mar 10, 2024
- Data Format: Full Model fp16
"Anything XL" by Yuno779 is the most popular illustration model at the time of editing, with over 300,000 downloads on Cibitai. By default, the illustration has a slightly animated paint job, but if you set (realistic:0.3), etc., the paint becomes more natural.
Usage Guidelines
You can control the quality and style of your illustrations by using special tags, much the same as in Animagine XL. Following these rules will help you produce high quality illustrations.
Structure of Prompts<|special|>,
<|artist|>,
<|special(optional)|>,
<|characters name|>, <|copyrights|>,
<|quality|>, <|meta|>, <|rating|>，……
<|tags|>, The special(optional) tag is used when you want to write a tag that is weaker than the first special tag. Tags described in the first special tag do not need to be described in the special tag (optional).
Special tags:
Quality Modifiers:Quality Modifier Score Criterion
masterpiece > 95%
best quality > ?
great quality > ?
good quality > ?
normal quality > ?
low quality > ?
worst quality ≤ 10%Rating Modifier Rating Criterion
safe General
sensitive Sensitive
nsfw Questionable
explicit, nsfw ExplicitYear Modifie Year Range
newest 2021 ~ 2024
recent 2018 ~ 2020
mid 2015 ~ 2017
early 2011 ~ 2014
oldest 2005 ~ 2010Recommended setting
Negative Promptnsfw, lowres, (bad), text, error, fewer, extra, missing, worst quality, jpeg artifacts, low quality, watermark, unfinished, displeasing, oldest, early, chromatic aberration, signature, extra digits, artistic error, username, scan, [abstract]When weighting the Negative Prompt, do not weight it too high, such as ugly:2.8. However, for some tags, it is necessary to use a value around 1.5.
masterpiece, best quality, very aesthetic, absurdresblue_pencil-XL
- Creator: 🔗blue_pen5805
- Download: 6.46 GB
- Upload: Apr 5, 2024
- Data Format: Pruned Model fp16
blue_pen5805's "blue_pencil-XL" is a popular SD1.5 model with distinctive eyes. By default, the model is more realistic, but you can add anime style, etc. to create an image that is more like an illustration.
anima_pencil-XL
- Creator: 🔗blue_pen5805
- Download: 6.46 GB
- Upload: May 14, 2024
- Data Format: Pruned Model fp16
blue_pen5805's " anima_pencil-XL" is a model by the same author as the "blue_pencil-XL" introduced earlier. This model is mainly merged with "ANIMAGINE XL 3.0 / 3.1" into "blue_pencil-XL". Compared to "blue_pencil-XL," this model produces illustrations in an animated style. Please note that the license is the same as ANIMAGINE XL 🔗Fair AI Public License 1.0-SD.
CounterfeitXL
- Creator: 🔗rqdwdw
- Download: 6.46 GB
- Upload: Dec 17, 2023
- Data Format: Pruned Model fp16
rqdwdw's "CounterfeitXL" is also an SDXL version of the popular counterfeit in SD 1.5 model. We recommend using "nativeXL." Try different styles using the style tags introduced.
Recommended setting
NegativeEmbedding:negativeXL
Style Tags
- masterpiece: Beautifully and intricately drawn illustrations.
- cinematic style: Stunning illustrations with detailed backgrounds and lighting.
- aesthetic style: Illustrations featuring beautiful color schemes.
- cute style: Cute.
- flat color: Illustrations with strong lines and a minimal color palette.
- anime style: Anime.
- comic style: Illustrations characterized by bold color choices and abstract representations.
AAM XL (Anime Mix)
- Creator: 🔗Lykon
- Download: 6.46 GB
- Upload: Jan 18, 2024
- Data Format: Full Model fp16
Lykon's "AAM XL (Anime Mix)" is created by the same author as "DreamShaper XL," a popular photo-realistic illustration software. It can produce very high quality illustrations. If you do not want to generate adult illustrations, use v1.0 SFW version.
Recommended setting
Sampling method: Euler a
Steps: 20-30 steps
CFG scale: 5-7HimawariMix
- Creator: 🔗min__san
- Download: 6.62 GB
- Upload: Jun 10, 2024
- Data Format: Full Model fp32
"HimawariMix" by min__san is a model that can generate kawaii-style illustrations. It generates beautiful illustrations with details. It has a built-in VAE, so you can generate illustrations without a VAE.
LizMix
- Creator: 🔗hemlok
- Download: 6.46 GB
- Upload: May 29, 2024
- Data Format: Pruned Model fp16
"LizMix" by hemlok produces high quality illustrations with great detail. It is based on Animagine XL V3, so you can use the Animagine XL V3 guidelines to write your own prompts.
Recommended setting
Sampling method: DPM++ 3M SDE Karras または DPM++ 2M SDE Karras
Steps: 20-
Clip skip: 2
CFG scale: 3-8
Denoising strength: 0.6ShiratakiMix XL
- Creator: 🔗Vsukiyaki
- Download: 6.46 GB
- Upload: Mar 24, 2024
- Data Format: Pruned Model fp16
"ShiratakiMix XL" by Vsukiyaki is the SDXL version of the model we introduced in our SD1.5 recommendation. It can generate kawaii illustrations. This model is also trained based on "Animagine XL V3", so you can also write prompts using the "Animagine XL V3" usage guidelines.
2.5D Illustrative checkpoint models
There are not many 2.5D checkpoint models in SDXL, but you can also make them 2.5D by adding realistic, etc. pronts to the illustration-based models shown above.
RealCartoon-XL
- Creator: 🔗7whitefire7
- Download: 12.92/6.46 GB
- Upload: Dec 20, 2023
- Data Format: Full Model fp16/Pruned Model fp16
"RealCartoon-XL" by 7whitefire7 is also an SDXL version of the model introduced in the SD1.5 recommendation. This model also has a built-in VAE, so it can be generated without a VAE. There are two models, a full model and a pruned model, so download whichever you prefer. If you have no particular preference, the pruned model is recommended.
Recommended setting
Sampling method: Eular A または DPM++ 2M Karras
Steps: 45 - 55
Hires steps: 10
Hires Denoising strength: 0.34 - 0.45
Hires upscaler: R-ESRGAN 4x+ または 4k-UltraSharp
Hires upscale: 1.5 または 2.0
Clip skip: 2[Lah] Mysterious | SDXL
![Sample images of [Lah] Mysterious | SDXL-1](/_next/image?url=https%3A%2F%2Fdca.data-hub-center.com%2Fcontent%2Fuploads%2F2024%2F06%2Flahmysterioussdxl_v40_sample-001.jpg&w=1920&q=85)
![Sample images of [Lah] Mysterious | SDXL-2](/_next/image?url=https%3A%2F%2Fdca.data-hub-center.com%2Fcontent%2Fuploads%2F2024%2F06%2Flahmysterioussdxl_v40_sample-002.jpg&w=1920&q=85)
![Sample images of [Lah] Mysterious | SDXL-3](/_next/image?url=https%3A%2F%2Fdca.data-hub-center.com%2Fcontent%2Fuploads%2F2024%2F06%2Flahmysterioussdxl_v40_sample-003.jpg&w=1920&q=85)
![Sample images of [Lah] Mysterious | SDXL-4](/_next/image?url=https%3A%2F%2Fdca.data-hub-center.com%2Fcontent%2Fuploads%2F2024%2F06%2Flahmysterioussdxl_v40_sample-004.jpg&w=1920&q=85)
- Creator: 🔗Lah_Inthe_Futureland
- Download: 6.46 GB
- Upload: Sep 15, 2023
- Data Format: Pruned Model fp16
Lah_Inthe_Futureland's "[Lah] Mysterious | SDXL" can generate illustrations that are somewhat photo-realistic but have a unique atmosphere. It can also generate Asian-style fantasy depictions. If you want to generate photorealistic images, set the CFG Scale to 1-3.
Recommended setting
Sampling method: DPM++ 2M SDE Karras または Euler a
Steps: 30-60
CFG scale: 3-9
ADetailer: ONExample Prompt:
Prompt(masterpiece,best quality:1.2), ultra realistic,1girl, long hair, straight hair, portrait, mysterious forest, firefly, bokeh, mysterious (worst quality, low quality:1.3), (greyscale, monochrome:1.1), cropped, lowres, text, jpeg artifacts, signature, watermark, username, blurry, artist name, trademark, watermark, title, multiple view, extra hand, mask, (animal ear:1.4), Tamarin_XL
- Creator: 🔗antinoice
- Download: 6.46 GB
- Upload: Dec 24, 2023
- Data Format: Full Model fp16
antinoice's "Tamarin_XL" can generate high quality illustrations. Try different styles. This model also has built-in VAE, so it can be generated without VAE. We also recommend that you do not use Negative Prompt and generate it as blank.
Photorealistic checkpoint models
Photorealistic checkpoint models can generate very high quality images because of SDXL's strengths, such as the ability to generate fingers.
SDXL Unstable Diffusers V12
- Creator: 🔗Yamer
- Download: 6.46 GB
- Upload: Mar 17, 2024
- Data Format: Pruned Model fp16
"SDXL Unstable Diffusers V12" by Yamer is a model that can generate artistic images. Although not explicitly stated, the prompt style seems to support both Danbooru style/natural language style.
Recommended setting
Steps: 35-150
Hires upscaler: 4x_foolhardy_Remacri または 4xUltraSharp
Clip skip: 1-2
VAE: SDXL VAERecommended upscaler
Go to the following page and place the .pth file in \stable-diffusion-webui\models\ESRGAN
AlbedoBase XL
- Creator: 🔗albedobond
- Download: 6.46 GB
- Upload: Feb 5, 2024
- Data Format: Pruned Model fp16
"AlbedoBase XL" by albedobond can generate high-quality photorealistic images. This model can also generate illustration-type images, but we classify it here because of its high photorealistic quality. This model also has a built-in VAE, so it can generate images without a VAE.
DreamShaper XL alpha2
- Creator: 🔗Lykon
- Download: 6.46 GB
- Upload: Jul 27, 2023
- Data Format: Full Model fp16
"DreamShaper XL alpha2" by Lykon is a popular model on Civitai with over 110,000 downloads. This model is also a general-purpose model that can be used to create illustrations, and is the SDXL version of the model introduced in our SD1.5 recommendations. The Alpha2 model introduced this time is getting old, so we recommend using the Turbo or Lightning version.
LEOSAM's HelloWorld XL
- Creator: 🔗LEOSAM
- Download: 6.46 GB
- Upload: Jun 13, 2024
- Data Format: Pruned Model fp16
LEOSAM's "LEOSAM's HelloWorld XL" model is capable of a wide range of expressions. For example, it can depict people from Europe, Asia, Africa, etc.
Recommended setting
Sampling method: DPM++ 2M Karras
Steps: ≥ 25
CFG scale: 10
ADetailer: onHigh-Frequency Tagging Word List: Refer to the 🔗High-Frequency Tagging Word List of HelloWorld 6.0. This is V6.0 data, but it can be used in V7.0 as well.
Juggernaut XL
- Creator: 🔗KandooAI
- Download: 6.62 GB
- Upload: Apr 20, 2024
- Data Format: Full Model fp16
KandooAI's "Juggernaut XL" is a very popular model with over 580,000 downloads on Civitai. It can produce a wide range of expressive images in a variety of styles. Read the 🔗Detailed Prompting Guide by Adam for more information on prompting. This model also has a built-in VAE, so it can be generated without a VAE.
Recommended setting
Sampling method: DPM++ 2M Karras
Steps: 30-40
CFG scale: 3-7(less is a bit more realistic)
Hires steps: 15
Hires Denoising strength: 0.3
Hires upscaler: 4xNMKD-Siax_200kRecommended upscaler
Go to the following page and download the file nmkdSiaxCX_200k.pt, rewrite the file extension to .pth and place it in \stable-diffusion-webui\models\ESRGAN
Conclusion
In this article, we have introduced you to SDXL's checkpoint model, which combines the rich expressiveness of SDXL with checkpoint orientation to generate illustrations that exceed what we aimed for. As of this writing, SD3 is now available to the public, further expanding the possibilities of Stable Diffusion. Digital Creative AI will continue to bring you the latest information. Our next issue will focus on the Lora model of SDXL.
