stable-diffusion text-to-image

ご利用の際は下記のライセンス内容を十分にご確認ください。

If you can read English, please refer here.

DeDeDeはアニメ調の人物を出力しやすいように調整されたStable Diffusionモデルです。
ベースモデルのDreamLike Diffusion 1.0へTrinart Characters v2 Derrida をStable Diffusion 1.4を用い差分マージしました。
そこからDreamLike Photoreal 1.0でIN0~5を調節、さらに30000枚のSD2.1、Novel AI、WD1.3/1.4、CoolJapan Diffusion 2.1、Dreamlike Photoreal 2.0で出力された画像でチューニングされています。

利用の際は以下のPrompt/Negative Promptをおすすめします。
P: best quality, masterpiece
NP: 3d, flat shading, flat color, retro style, 1980s, 1990s, 2000s, 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017 2018 2019, lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry, artist name, inaccurate limb

例

<img src="https://huggingface.co/nakayama/DeDeDe/resolve/main/img/image06.png" style="max-width:400px;" width="50%"/>

(((best quality, masterpiece, 8k))), detailed anime style of 1girl sitting in room and reading book wearing school uniform and wavy detailed pink hair pink and detailed yellow eye yellow, smiling
Negative prompt: [[3d]], (((flat shading, flat color))), retro style, 1980s, 1990s, 2000s, 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017 2018 2019, lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry, artist name, inaccurate limb
Steps: 25, Sampler: Euler a, CFG scale: 8, Seed: 2277801742, Size: 512x768, Model hash: 6d1729a039, Denoising strength: 0.75, Clip skip: 2, ENSD: 31337, Hires upscale: 1.5, Hires upscaler: Latent

<img src="https://huggingface.co/nakayama/DeDeDe/resolve/main/img/image01.png" style="max-width:400px;" width="50%"/>

(((best quality, masterpiece, 8k))), detailed anime style of anime 1girl bust shot sitting and dipping in river and wetty wearing white transparent onepiece dress with detailed wavy pink hair pink and hetailed yellow eye yellow, water splash in gorgeous scene secret garden
Negative prompt: [[[3d]]], (((flat shading, flat color))), retro style, 1980s, 1990s, 2000s, 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017 2018 2019, lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry, artist name, inaccurate limb
Steps: 20, Sampler: DPM++ SDE Karras, CFG scale: 8, Seed: 1117106368, Size: 512x768, Model hash: 6d1729a039, Denoising strength: 0.7, Clip skip: 2, ENSD: 31337, Hires resize: 768x1152, Hires steps: 5, Hires upscaler: Latent

<img src="https://huggingface.co/nakayama/DeDeDe/resolve/main/img/image02.png" style="max-width:400px;" width="50%"/>

(((best quality, masterpiece))), detailed anime style of bunny girl bishoujo from front wearing intricate frill jirai kei bikini with detailed wavy pink hair pink and detailed yellow eye yellow and lying on the bed in (((glitter harajuku kawaii messy room with flower and candy)))
Negative prompt: 3d, (((flat shading, flat color))), retro style, 1980s, 1990s, 2000s, 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017 2018 2019, lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry, artist name, inaccurate limb, eating
Steps: 50, Sampler: DPM++ SDE Karras, CFG scale: 5.5, Seed: 911018641, Size: 768x512, Model hash: 6d1729a039, Denoising strength: 0.2, Clip skip: 2, ENSD: 31337, Hires resize: 1152x768, Hires steps: 10, Hires upscaler: ESRGAN_4x

<img src="https://huggingface.co/nakayama/DeDeDe/resolve/main/img/image03.png" style="max-width:400px;" width="50%"/>

(((best quality, masterpiece))), detailed anime style of 1girl bishoujo full body standing and looking from viewer and wearing classical frill dress with cape with wavy detailed pink hair pink and detailed yellow hair yellow in scenic view british fantastic landscape with flowing and golden hour, dynamic pose
Negative prompt: [[[[3d]]]], (((flat shading, flat color))), retro style, 1980s, 1990s, 2000s, 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017 2018 2019, lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry, artist name, inaccurate limb
Steps: 15, Sampler: DPM++ SDE Karras, CFG scale: 5.5, Seed: 2149903685, Size: 768x512, Model hash: 6d1729a039, Denoising strength: 0.75, Clip skip: 2, ENSD: 31337, Hires resize: 1152x768, Hires steps: 10, Hires upscaler: Latent

<img src="https://huggingface.co/nakayama/DeDeDe/resolve/main/img/image04.png" style="max-width:400px;" width="50%"/>

(((best quality, masterpiece, 8k))), detailed anime style of 1boy cowboy shot wearing samurai outfit with flat chest and fighting pose, fist, motion blur
Negative prompt: 3d, flat shading, flat color, retro style, 1980s, 1990s, 2000s, 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017 2018 2019, lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry, artist name, inaccurate limb
Steps: 20, Sampler: Euler a, CFG scale: 8, Seed: 2136917906, Size: 512x768, Model hash: 6d1729a039, Denoising strength: 0.6, Clip skip: 2, ENSD: 31337, Hires resize: 768x1152, Hires steps: 20, Hires upscaler: Latent

<img src="https://huggingface.co/nakayama/DeDeDe/resolve/main/img/image05.png" style="max-width:400px;" width="50%"/>

(((best quality, masterpiece, 8k))), detailed photorealistic style of 1boy cowboy shot wearing mad max outfit with flat chest and fighting pose, fist, motion blur, abandoned city background
Negative prompt: (((flat shading, flat color))), retro style, 1980s, 1990s, 2000s, 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017 2018 2019, lowres, bad anatomy, bad hands, text, error, missing fingers, extra digit, fewer digits, cropped, worst quality, low quality, normal quality, jpeg artifacts, signature, watermark, username, blurry, artist name, inaccurate limb
Steps: 20, Sampler: Euler a, CFG scale: 8, Seed: 3787706619, Size: 768x512, Model hash: 6d1729a039, Denoising strength: 0.7, Clip skip: 2, ENSD: 31337, Hires resize: 1152x768, Hires steps: 5, Hires upscaler: Latent

マージ・学習手順について

以下、モデル横に記載されている記号列は、Automatic1111 Webui コミットハッシュ c98cb0f8ecc904666f47684e238dd022039ca16f 時点での、モデル選択時に記載されているckptのハッシュ値です。

  1. Dreamlike Diffusion 1.0にTrinart Derridaを差分マージする | Interpolation Method | Primary Model | Secondary Model | Tertiary Model | Merge Name | | --- | --- | --- | --- | --- | | Add Difference @ 1.0 | DreamLike Diffusion 1.0(0aecbcfa2c) | TrinArt Characters v2 Derrida(42d3f359b0) | Stable Diffusion 1.4(fe4efff1e1) | DDD_pre1(d1ac03017b) |

  2. DDD_pre1にDreamlike Photoreal 1.0でIN00~IN05を階層マージで編集する | Model: A | Model: B | Weight | Base alpha | Merge Name | | --- | --- | --- | --- | --- | | DDD_pre1(d1ac03017b) | Dreamlike Photoreal 1.0(f403e4e2a5) | 0.45,0.45,0.4,0.35,0.3,0.25,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0 | 0 | DDD_pre2(601ec74593) |

  3. DDD_pre2に対し、自前で用意した他Diffusion Modelの出力からなる素材画像にて学習させる
    用意の際に利用したサービス/モデルは、SD2.1、Novel AI、WD1.3/1.4、CoolJapan Diffusion 2.1、Dreamlike Photoreal 2.0。
    総数は30000程、flipしたものと合わせてlearning rateは5e-6、60000Step学習させた。
    これにより生成したモデルをDDD_pre3(4709475652)とする。

  4. DDD_pre3にDDD_pre2を加重平均でマージする | Interpolation Method | Primary Model | Secondary Model | Merge Name | | --- | --- | --- | --- | | Weighted Sum @ 0.5 | DDD_pre3(4709475652) | DDD_pre2(601ec74593) | DeDeDe(6d1729a039) |

フレーバー

DeDeDe_ip2p_0.7_0.8.ckpt/DeDeDe_ip2p_0.7_1.0.ckpt

Instruct pix2pixモデルからタスクベクトルを抽出して加算したモデルです。
それぞれDeDeDe 0.8/Instruct Pix2Pix 0.7、DeDeDe 1.0/Instruct Pix2Pix 0.7の大きさとなります。
以下はInstruct Pix2Pixから継承したライセンスとなります。
Copyright (c) 2023 Ren Nakayama
Released under the MIT license
https://huggingface.co/nakayama/DeDeDe/blob/main/MIT-License

DeDeDe_controlnet_.pth/DeDeDe_webui_controlnet_.safetensors

ControlNetで用いられているモデルにDeDeDeをマージしたものです。
具体的な方法はこちらをご確認ください。

DeDeDeP

https://huggingface.co/nakayama/DeDeDeP
DeDeDeをベースにDreamlike Photoreal 1.0と階層マージを用いて編集したものです。
DeDeDeと比較してより写実的に出力が寄るようになっています。

備考

SNSなどに出力した作品をアップロードする際に、タグなどの機能があれば #DeDeDeArt などをつけていただければ嬉しいです。
私が見に行くので。

ライセンスについて

当モデルはDreamlike Diffusion 1.0 / Dreamlike Photoreal 1.0の影響下にあるため、上記モデルにおける修正されたCreativeML OpenRAIL-M licenseが適用されます。
以下はDeepLで翻訳された修正分の日本語訳となりますが、解釈において優先される言語は英語となります。