AniVerse XL - v1.0

You can find my model in MAGE:

https://www.mage.space/play/d3699ef7c4a6e852e7e51e6755dffb13

->
Surpriiiiiise!!!

As they say: better late than never!

First of all:
The images in the main gallery are divided as follows:

• I created the first 8 images

• The rest are made by nuaion, who is much better than me!
  

Now, let's jump straight into the possible questions and answers to clear up any doubts:

• Q) Have you finally bought the new PC?

A) No, I'm still saving money to buy it, as I mentioned before, I hope to get it by the end of 2024 or early 2025. This is the PC I want to build: https://www.amazon.it/hz/wishlist/ls/371J30CMA0EC5/ref=nav_wishlist_lists_2 As always, if you'd like to help me with this project, you can donate on my Ko-Fi page: https://ko-fi.com/samael1976

• Q) So how did you manage to create a trained model of SDXL?

A) It was tough, but after countless attempts (not exactly a million, but really many), I managed to get the SDXL training to work on a 2060 with 12GB of VRAM.

For more details, here’s the link to my article: https://civitai.com/articles/5672/how-to-train-a-sdxl-style-with-only-12gb-of-vram-rtx-2060-onetrainer

• Q) Will you stop training for Stable Diffusion 1.5 now?

A) Not at all. As far as I'm concerned, Stable Diffusion 1.5, despite its limitations, remains a little gem. Sure, projects involving Stable Diffusion 1.5 will be delayed, but for now, I have no intention of stopping the creation of models for Stable Diffusion 1.5.

• Q) This model is not like AniVerse for SD 1.5

A) You’re absolutely right! I thought a lot about whether to name it AniVerse XL or something else. In the end, I chose AniVerse XL because this first model is just an "alpha test."

Think of it this way: to create the true AniVerse XL will take time, and each new version will get closer to resembling AniVerse. In any case, there are various tags in the training. If you want an image more similar to the AniVerse style, include this in the positive prompt: 4n1v3rs3.

• Q) Why will it take so long?

A) Unfortunately, injecting a specific style into SDXL is very difficult, if not impossible. However, with each new version, I will inject more AniVerse images, which should (theoretically) eventually make it resemble AniVerse 1.5. It’s not guaranteed, but that’s my plan. Plus, with a 2060, each training session takes 25 to 30 days of 24/7 PC work solely for training.

• Q) Is it suitable for creating NSFW images?

A) No, sorry, this version of AniVerse XL was not designed for creating NSFW images. It’s not impossible, but you’ll have difficulties.

• Q) Will you make a Pony version as well?

A) Yes, it's already in training. (I hope that the training working good)

• Q) What about the future?

A) In the future, I’d like to explore other Open Source models, first and foremost Pixart Sigma. It’s an exciting time after the release of Stable Diffusion 3.

• Q) Will you create a model for Stable Diffusion 3 as well?

A) It’s too early to say. It’s just been released, and there are still many issues, including regarding the usage license. So, I prefer to wait and see how things develop.

• Q) Do you have any settings to suggest for your model?

A) As mentioned, I’m not great with SDXL, so I rely on you and hope you can help me find the best settings.
- But I found these setting:

• VAE: a special VAE is already included, you don't need it - (Thanx to nuaion)

• Clip skip: 2

• Upscaler: 4x-Ultrasharp or 4X NMKD Superscale

• Width: 720 (o 768) - (probably more, but I've not tested)

• Height: 1280 - (probably more, but I've not tested)

• CFG Scale: 3~7 -> Steps: 20~30 (also, I've not tested so much)

• For 2D (or very close to 2D) - see this comparison post

  • CFG: 3~7

  • Sampling favorite method: Restart

  • Embedding to use: zPDXL2 + SimplePositiveXLv2 (in positive prompt) + unaestheticXL_bp5, zPDXL2-neg (in negative prompt) maybe the best

  • Sampling alternative favorite method: Euler a

  • Embedding to use: zPDXL2 + unaestheticXL_bp5, zPDXL2-neg

    EXAMPLE OF GENERAL PROMPT

  • POSITIVE PROMPT: zPDXL2, 4n1v3rs3, (Style of the image), your subject, the description, (background), more details
    EXAMPLE: zPDXL2, 4n1v3rs3, (Oil Painting), woman, wearing red dress, (sunrise background), masterpiece, high details... (ect, ect)

  • NEGATIVE PROMPT: unaestheticXL_bp5, zPDXL2-neg, moles, freckles, ugly, artifacts

• For 2.5D - see this comparison post

  • CFG: 3~7

  • Sampling favorite method: DPM++ 2M or Euler_max

  • Embedding to use: zPDXL2 + SimplePositiveXLv2 (in positive prompt) + unaestheticXL_bp5, zPDXL2-neg, moles, freckles, ugly, artifacts"

    EXAMPLE OF MY GENERAL PROMPT

  • POSITIVE PROMPT: zPDXL2, 4n1v3rs3, (Style of the image), your subject, the description, (background), more details

    EXAMPLE: zPDXL2, 4n1v3rs3, (Anime 2.5D style), Rei Ayanami, wearing white EVA bodysuit, (sunrise background), masterpiece, high details... (ect, ect)

  • NEGATIVE PROMPT: unaestheticXL_bp5, zPDXL2-neg, moles, freckles, ugly, artifacts

• For 3D - Realism - see this comparison posts: 1 - 2 - 3 - 4 - 5

  • CFG: 2.5~7

  • Clip Skip: 1~2

  • Sampler & Scheduler(1): DPM++ 2M & Karras

  • Sampler & Scheduler(2): DPM++ SDE & Karras (the best for realism, but very slow generation)

  • Sampler & Scheduler(3): Euler_max & Karras

  • Sampler & Scheduler(4): UniPC & Exponential

  • Embeddings to use: zPDXL2 + SimplePositiveXLv2 (in positive prompt) + unaestheticXL_bp5, zPDXL2-neg

    EXAMPLE PROMPT FOR REALISM:

  • POSITIVE: 4n1v3rs3, zPDXL2, (Style of the image), your subject, the description, (background), more details
    EXAMPLE: 4n1v3rs3, zPDXL2, (Analog photo by Rutkowski), (Hindu young man), 25 years old, (elaborate ancient Hindu temple ruins background), (midnight hour, high quality, film grain), focus on orange and green

  • NEGATIVE: unaestheticXL_bp5, zPDXL2-neg, moles, freckles, ugly, artifacts

  • MORE INFO (full prompt info): Clicking in this post, you can see all the setting I used, and choose your CFG Scale

• Using Turbo or Lightning - See this post

  • CFG: 2.5~3.5

  • Clipskip: 2~3

  • Lora to use: SDXL Lightning LoRAs

  • Sampling method: dpm 2 turbo

  • Embeddings to use: zPDXL2 + SimplePositiveXLv2 (in positive prompt) + unaestheticXL_bp5, zPDXL2-neg

AVOID THIS COMBINATION: (Euler a + Uniform) - (Euler_max + SGM Uniform) - (DPM++ 3M SDE + Exponential) - (DPM++ 2M SDE + Exponential)

DOWNLOAD ALL THE EMBEDDING AND LORA THAT I USED:

• unaestheticXL_bp5 (put in your negative prompt)

• SimplePositiveXLv2 (put in your positive prompt)

• Add Details: Detail Tweaker (in positive prompt)

• Pony PDXL Negative Embeddings: High Quality V2

• Pony PDXL Negative Embeddings: Photo Real

• LoRA lightning: SDXL Lightning LoRAs

I hope I’ve cleared up any doubts.

Now for the thank yous:

1) A million thanks to everyone in the civitai Italia Telegram group, whose support kept me from giving up on SDXL.

2) To nuaion, who supported me by providing the base model for the training and thousands of images for the dataset.

3) To GattaPlayer, without whom I would never have managed to achieve anything useful. I don’t know how many hours we spent testing and configuring the training.

4) To MarkWar, because this journey probably wouldn’t have happened without him.

5) To Furkan, who showed me how to train with 12GB of VRAM. His setup didn’t work for me, but he remains a wonderful person, always ready to help those in need.

6) To all of you who support me with your kind words, comments, and by sending your images. I read all your messages and I alway try to look at all the images you create!

Now it’s your turn! Try out the model, send in your creations, and let me know in the comments what you think of this model!

I love you all and I hope that you enjoy this model!

<----------------------------------------------->
------------- ITALIAN VERSION ----------------

<----------------------------------------------->

Sorpresaaaaaa!!!

Come si suol dire: meglio tardi che mai!

Spero di aver fatto cosa gradita con questo modello :)

Prima di tutto:
Le immagini della galleria principale sono suddivise così

• Io ho fatto le prime 8 immagini

• nuaion, che è molto più bravo di me, mi ha gentilmente creato le restanti!

Passo subito alle possibili domande e risposte, così da levarvi qualche dubbio

• Q) Hai comprato finalmente il nuovo PC?

A) No, sto ancora mettendo via i soldi per comprarlo, come avevo già detto, spero di riuscirci verso fine 2024, inizio 2025. Questo è il Pc che vorrei montarmi: https://www.amazon.it/hz/wishlist/ls/371J30CMA0EC5/ref=nav_wishlist_lists_2 Come sempre ricordo che se vi fa piacere aiutarmi in questo progetto, esiste la mia pagina Ko-Fi per le donazioni: https://ko-fi.com/samael1976

• Q) Ma allora come hai fatto a fare un modello trained di SDXL?

A) E' stato difficile, ma dopo un milione di tentativi (non dico esattamente un milione, ma veramente tanti), sono riuscito a far funzionare il training di SDXL su una 2060 con 12gb di VRam.

Per ulteriori dettagli vi lascio il link al mio articolo: https://civitai.com/articles/5672/how-to-train-a-sdxl-style-with-only-12gb-of-vram-rtx-2060-onetrainer

• Q) Ora smetterai di fare training per Stable Diffusion 1.5?

A) Non mi passa neppure per l'anticamera del cervello. Per quanto mi riguarda Stable Diffusion 1.5, seppur con tutte le sue limitazioni, rimane una piccola chicca. Certo, i progetti riguardanti stable Diffusion 1.5 subiranno uno slittamento, ma di certo, al momento, non ho intenzione di smettere di fare modelli per stable diffusion 1.5

• Q) Questo modello non è come AniVerse per SD 1.5

A) Hai perfettamente ragione! Ho pensato veramente tanto se dargli il nome AniVerse XL o mettergli un altro nome. Alla fine ho optato per AniVerse XL, perchè questo primo modello è solo un "alpha test".

Mettetela così, per creare il vero AniVerse XL, ci metterò del tempo, e ogni nuova versione si avvicinerà sempre di più ad assomigliare ad AniVerse

In ogni caso, dentro al training ci sono svariati tag, se volete un immagine più simile allo stile di AniVerse, nel prompt positivo inserite: 4n1v3rs3

• Q) Perchè ci metterai tutto questo tempo?

A) Purtroppo iniettare uno stile specifico a SDXL è molto difficile, se non impossibile. Però ad ogni versione nuova inietterò sempre di più immagini di AniVerse, questo dovrebbe (in linea teorica) prima o poi, farlo assomigliare ad AniVerse 1.5. Oh ragazzi, non è detto, ma questo è nella mia testa. Oltre al fatto che avendo una 2060, ogni training mi porterà via dai 25 ai 30 giorni di 24h/24h di PC che lavora solo di training.

• Q) E' adatto a fare immagini NSFW?

A) No, mi dispiace, questa versione di AniVerse XL non è stata pensata per creare immagini NSFW. Non è detto che tu non ci riesca, ma avrai delle difficoltà.

• Q) Farai una versione anche Pony?

A) Sì, è gia in training. Spero che il modello finale sia funzionante.

• Q) E in futuro?

A) In futuro mi piacerebbe esplorare anche altri modelli Open Source, primo fra tutti Pixart Sigma. E' un momento di fervento dopo l'uscita di Stable diffusion 3.

• Q) Farai anche un modello per Stable Diffusion 3?

A) E' troppo troppo presto per dirlo. E' uscito da pochissimo, ci sono ancora tante criticità al riguardo, anche riguardo la licenza d'uso. Per cui, preferisco mettermi in un angolino ed aspettare per vedere come si evolverà la situazione.

• Q) Hai qualche setting da suggerire per il tuo modello?

A) Questo è quello che ho trovato:

• VAE: a special VAE is already included, you don't need it - (Thanx to nuaion)

• Clip skip: 2

• Upscaler: 4x-Ultrasharp or 4X NMKD Superscale

• Width: 720 (o 768) - (probably more, but I've not tested)

• Height: 1280 - (probably more, but I've not tested)

• CFG Scale: 3~7 -> Steps: 20~30 (also, I've not tested so much)

• For 2D (or very close to 2D) - see this comparison post

  • CFG: 3~7

  • Sampling favorite method: Restart

  • Embedding to use: zPDXL2 + SimplePositiveXLv2 (in positive prompt) + unaestheticXL_bp5, zPDXL2-neg (in negative prompt) maybe the best

  • Sampling alternative favorite method: Euler a

  • Embedding to use: zPDXL2 + unaestheticXL_bp5, zPDXL2-neg

    EXAMPLE OF GENERAL PROMPT

  • POSITIVE PROMPT: zPDXL2, 4n1v3rs3, (Style of the image), your subject, the description, (background), more details
    EXAMPLE: zPDXL2, 4n1v3rs3, (Oil Painting), woman, wearing red dress, (sunrise background), masterpiece, high details... (ect, ect)

  • NEGATIVE PROMPT: unaestheticXL_bp5, zPDXL2-neg, moles, freckles, ugly, artifacts

• For 2.5D - see this comparison post

  • CFG: 3~7

  • Sampling favorite method: DPM++ 2M or Euler_max

  • Embedding to use: zPDXL2 + SimplePositiveXLv2 (in positive prompt) + unaestheticXL_bp5, zPDXL2-neg, moles, freckles, ugly, artifacts"

    EXAMPLE OF MY GENERAL PROMPT

  • POSITIVE PROMPT: zPDXL2, 4n1v3rs3, (Style of the image), your subject, the description, (background), more details

    EXAMPLE: zPDXL2, 4n1v3rs3, (Anime 2.5D style), Rei Ayanami, wearing white EVA bodysuit, (sunrise background), masterpiece, high details... (ect, ect)

  • NEGATIVE PROMPT: unaestheticXL_bp5, zPDXL2-neg, moles, freckles, ugly, artifacts

• For 3D - Realism - see this comparison posts: 1 - 2 - 3 - 4 - 5

  • CFG: 2.5~7

  • Clip Skip: 1~2

  • Sampler & Scheduler(1): DPM++ 2M & Karras

  • Sampler & Scheduler(2): DPM++ SDE & Karras (the best for realism, but very slow generation)

  • Sampler & Scheduler(3): Euler_max & Karras

  • Sampler & Scheduler(4): UniPC & Exponential

  • Embeddings to use: zPDXL2 + SimplePositiveXLv2 (in positive prompt) + unaestheticXL_bp5, zPDXL2-neg

    EXAMPLE PROMPT FOR REALISM:

  • POSITIVE: 4n1v3rs3, zPDXL2, (Style of the image), your subject, the description, (background), more details
    EXAMPLE: 4n1v3rs3, zPDXL2, (Analog photo by Rutkowski), (Hindu young man), 25 years old, (elaborate ancient Hindu temple ruins background), (midnight hour, high quality, film grain), focus on orange and green

  • NEGATIVE: unaestheticXL_bp5, zPDXL2-neg, moles, freckles, ugly, artifacts

  • MORE INFO (full prompt info): Clicking in this post, you can see all the setting I used, and choose your CFG Scale

• Using Turbo or Lightning - See this post

  • CFG: 2.5~3.5

  • Clipskip: 2~3

  • Lora to use: SDXL Lightning LoRAs

  • Sampling method: dpm 2 turbo

  • Embeddings to use: zPDXL2 + SimplePositiveXLv2 (in positive prompt) + unaestheticXL_bp5, zPDXL2-neg

AVOID THIS COMBINATION: (Euler a + Uniform) - (Euler_max + SGM Uniform) - (DPM++ 3M SDE + Exponential) - (DPM++ 2M SDE + Exponential)

DOWNLOAD ALL THE EMBEDDING AND LORA THAT I USED:

• unaestheticXL_bp5 (put in your negative prompt)

• SimplePositiveXLv2 (put in your positive prompt)

• Add Details: Detail Tweaker (in positive prompt)

• Pony PDXL Negative Embeddings: High Quality V2

• Pony PDXL Negative Embeddings: Photo Real

• LoRA lightning: SDXL Lightning LoRAs
  ------------------------------------------------------------------------------------------------

Spero di aver spianato ogni dubbio.

Ora passo ai ringraziamenti:

1) Un milione di grazie tutte le persone del gruppo di civitai italia su telegram, che con il loro appoggio, sono riusciti a non farmi demordere dal fare SDXL.

2) A nuaion che mi ha supportato aiutandomi sia con il modello base per il training, sia fornendomi migliaia di immagini per il dataset.

3) A GattaPlayer senza di lei non sarei mai riuscito a portare a casa qualcosa di utile. Non so quante ore abbiamo speso a provare e testare configurazioni per il training.

4) A MarkWar, perchè questo viaggio senza lui probabilmente non ci sarebbe mai stato.

5) A Furkan, che mi ha mostrato la via di come poter fare training con 12gb di Vram, anche se il suo setting non mi funzionava, ma per me lui rimane una bellissima persona, sempre pronto ad aiutare chi ha bisogno.

6) A tutti voi che mi supportate con le vostre belle parole, con i vostri commenti e inviando le vostre immagini. Sappiate che li leggo tutti e che guardo tutte le immagini che create!

Ora lascio la parola a voi! Provate il modello, inviate le vostre creazioni e fatemi sapere nei commenti cosa ne pensate di questo modello!

Vi adoro tutti!