AniVerse

You can find my model in MAGE: <a target="_blank" rel="ugc" href="https://www.mage.space/play/d8d5657480d78a0e0b3b8fafa7d55bde">https://www.mage.space/play/d8d5657480d78a0e0b3b8fafa7d55bde</a> -&gt; PRUNED VERSIONThis AniVerse version was literally the hardest model that I've made, but I also think it's also the best version out of all the ones released.I have a lot of things to write, but I'll start by giving you the details for your creations, immediately after a short parenthesis.Why the wording HD and 2.0? I imagine you're wondering.HD: Because this version of AniVerse is trained (almost) entirely at 1024X1024px.2.0: Because there is a slight shift towards an increase in realism compared to previous versions////////////////////////////////////////////////////////////////////////////////////////////////////This version gives you more scope for use.You can use it as usual, to create 2.5D images in a semi-realistic, cartoonish or more realistic style.It all depends on the positive prompt and negative prompt you will use.You can find examples of prompts that I used.In summary I can tell you that:1) <a target="_blank" rel="ugc" href="https://huggingface.co/Xynon/models/blob/main/experimentals/TI/bad-image-v2-39000.pt">Bad-Images-39000</a> = More cartoonish images2) <a target="_blank" rel="ugc" href="https://civitai.com/models/7808?modelVersionId=9536">Easy Negative</a> and <a target="_blank" rel="ugc" href="https://civitai.com/models/71961?modelVersionId=94057">FastNegativeV2</a> = Usual Aniverse 2.5D style images3) <a target="_blank" rel="ugc" href="https://civitai.com/models/4629/deep-negative-v1x">ng_deepnegative_v1_75t</a> and <a target="_blank" rel="ugc" href="https://civitai.com/models/100191/ti-easynegativev2-textual-inversion-embedding">EasyNegativeV2</a> = Images like the classic Aniverse, with a pinch of greater realism (not much)////////////////////////////////////////////////////////////////////////////////////////////////////But what does training at 1024px mean?You can push your creations, with some peace of mind, to 1024x576 (or 1024x768) resolution, without using HiRes.fix. I went even further, up to 1280x720 with a margin of error of approximately ~5% (if you follow my advice):<ul><li>VAE: VAE is included (but usually I still use the<a target="_blank" rel="ugc" href="https://huggingface.co/stabilityai/sd-vae-ft-mse-original/resolve/main/vae-ft-mse-840000-ema-pruned.safetensors"> 840000 ema pruned</a>)</li><li>Clip skip: 2</li><li>Upscaler: <a target="_blank" rel="ugc" href="https://civitai.com/models/116225?modelVersionId=125843">4x-Ultrasharp</a> or <a target="_blank" rel="ugc" href="https://civitai.com/models/141491?modelVersionId=156841">4X NMKD Superscale</a></li></ul><ul><li>Sampling method: DPM++ 2M SDE Karras</li><li>Width: 576 (o 768)</li><li>Height: 1024</li><li>CFG Scale: 3 -&gt; Steps: 15 CFG Scale: 4 -&gt; Steps: 20 CFG Scale: 5 -&gt; Steps: 25 CFG Scale: 6 -&gt; Steps: 30...and so on...</li></ul>MY FAVORITE PROMPT:<ul><li>(masterpiece, best quality, highres:1.2), (photorealistic:1.2), (intricate and beautiful:1.2), (detailed light:1.2), (colorful, dynamic angle), RAW photo, upper body shot, fashion photography, YOUR PROMPT, (highres textures), dynamic pose, bokeh, soft light passing through hair, (abstract background:1.3), (sharp), exposure blend, bokeh, (hdr:1.4), high contrast, (cinematic), (muted colors, dim colors, soothing tones:1.3), morbidNEGATIVE PROMPT:</li><li>(worst quality, low quality), negative_hand-neg, bad-hands-5, naked, nude, braless, cross, sepia, black&amp;white, B&amp;W, painting, drawing, illustration</li></ul>YOU CAN ALSO USE THESE NEGATIVE EMBEDDINGS:<ul><li>1) <a target="_blank" rel="ugc" href="https://civitai.com/models/7808?modelVersionId=9536">Easy Negative</a>, <a target="_blank" rel="ugc" href="https://civitai.com/models/56519/negativehand-negative-embedding">negative_hand-neg</a>, <a target="_blank" rel="ugc" href="https://civitai.com/models/116230/bad-hands-5">bad-hands-5</a></li><li>2) <a target="_blank" rel="ugc" href="https://huggingface.co/Xynon/models/blob/main/experimentals/TI/bad-image-v2-39000.pt">Bad-Images-39000</a>, <a target="_blank" rel="ugc" href="https://civitai.com/models/56519/negativehand-negative-embedding">negative_hand-neg</a>, <a target="_blank" rel="ugc" href="https://civitai.com/models/116230/bad-hands-5">bad-hands-5</a></li><li>3) <a target="_blank" rel="ugc" href="https://civitai.com/models/4629/deep-negative-v1x">ng_deepnegative_v1_75t</a>, <a target="_blank" rel="ugc" href="https://civitai.com/models/116230/bad-hands-5">bad-hands-5</a></li><li>4) <a target="_blank" rel="ugc" href="https://civitai.com/models/71961?modelVersionId=94057">FastNegativeV2</a>, <a target="_blank" rel="ugc" href="https://civitai.com/models/56519/negativehand-negative-embedding">negative_hand-neg</a>, <a target="_blank" rel="ugc" href="https://civitai.com/models/116230/bad-hands-5">bad-hands-5</a></li><li>5) For MEN images: girl, woman, female, tits, <a target="_blank" rel="ugc" href="https://huggingface.co/Xynon/models/blob/main/experimentals/TI/bad-image-v2-39000.pt">BadImage_v2-39000</a>, <a target="_blank" rel="ugc" href="https://civitai.com/models/56519/negativehand-negative-embedding">negative_hand-neg</a>, <a target="_blank" rel="ugc" href="https://civitai.com/models/116230/bad-hands-5">bad-hands 5</a></li></ul>I created over 1000 images to be sure and when I talk about an error rate of around 5% I naturally mean the classic elongated or double bodies due to too high resolution.Sometimes three arms or three legs may appear, but it is another type of error, not related to resolution, so I excluded it from the error calculation.If you want to avoid this kind of error, just go down to a resolution of 1024X576px, to return to a sort of safe zone.The hands!!! I finally managed to improve the hands!!!I said improve, but they are not yet perfect and I am very happy with the details that have been achieved in the creations.There are other things that I would like to improve in the next version (or rather I'm already working to eliminate any cuts/artifacts that sometimes happen on the face)I would also like to improve the angle of the shadows and other small things (like the teeth).But I can say I am absolutely satisfied with this version.Now a little chitchat to tell you how sweaty this version was.The PC with which I'm training, every so often (I'm still trying to define the cause) decides to crash, so I have to turn it off and start training again (and sometimes I'm away from home and lose entire days).The training lasted approximately 490 hours, which is equivalent to approximately 21 continuous days of training.To find the fix for the hands, after having the final model, I worked on it for another 2 weeks.I repeat, it was hard, long and difficult, more than a few times I thought about throwing everything away, but in the end, thanks also to a bit of luck, I managed to create this version of AniVerse which is truly satisfying me.I hope it gives you the same satisfaction it is giving me.We'll talk again soon... with a new project (which has already been in progress for a while, but at this time of writing, I haven't the faintest idea if it will lead to the desired results).Samuel //////////////////////////////////////////////////////////////////////////////////////////////////// Questa versione di AniVerse è stata letteralmente il modello più difficile da fare, ma penso anche che sia anche la versione migliore tra tutti quelli pubblicati.Ho un sacco di cose da scrivere, ma inizio con il darvi subito i dettagli per le vostre creazioni, subito dopo una breve parentesi.Perchè la dicitura HD e 2.0? Immagino che ve lo starete chiedendo.HD: Perchè questa versione di AniVerse è allenata (quasi) totalmente a 1024X1024px.2.0: Perchè c'è un leggero stacco verso un aumento del realismo rispetto alle versioni precedenti.////////////////////////////////////////////////////////////////////////////////////////////////////Questa versione vi da più margine di utilizzo.Potete utilizzarla come al solito, per creare immagini 2.5D in stile semirealistico, cartoonistico oppure più realistico.Tutto dipende dal prompt positivo e dal prompt negativo che userete.Potete trovare gli esempi di prompt che io ho utilizzato.In sintesi vi posso dire che:1) <a target="_blank" rel="ugc" href="https://huggingface.co/Xynon/models/blob/main/experimentals/TI/bad-image-v2-39000.pt">Bad-Images-39000</a> = Immagini più cartonistiche2) <a target="_blank" rel="ugc" href="https://civitai.com/models/7808?modelVersionId=9536">Easy Negative</a> e <a target="_blank" rel="ugc" href="https://civitai.com/models/71961?modelVersionId=94057">FastNegativeV2</a> = Immagini solito stile di Aniverse 2.5D3) <a target="_blank" rel="ugc" href="https://civitai.com/models/4629/deep-negative-v1x">ng_deepnegative_v1_75t</a> e <a target="_blank" rel="ugc" href="https://civitai.com/models/100191/ti-easynegativev2-textual-inversion-embedding">EasyNegativeV2</a> = Immagini come il classico Aniverse, con un pizzico di realismo maggiore (poco)////////////////////////////////////////////////////////////////////////////////////////////////////Ma il training a 1024px cosa comporta?Beh, che potete spingere le vostre creazioni, senza HiRes.fix, con una certa tranquillità, fino a 1024X576px o 1024X720px. Io mi sono spinto anche oltre, fino a 1280X720px (e viceversa) con un error rate di circa il ~5% (se seguirete i miei consigli):<ul><li>VAE: VAE is included (but usually I still use the<a target="_blank" rel="ugc" href="https://huggingface.co/stabilityai/sd-vae-ft-mse-original/resolve/main/vae-ft-mse-840000-ema-pruned.safetensors"> 840000 ema pruned</a>)</li><li>Clip skip: 2</li><li>Upscaler: <a target="_blank" rel="ugc" href="https://civitai.com/models/116225?modelVersionId=125843">4x-Ultrasharp</a> or <a target="_blank" rel="ugc" href="https://civitai.com/models/141491?modelVersionId=156841">4X NMKD Superscale</a></li></ul><ul><li>Sampling method: DPM++ 2M SDE Karras</li><li>Width: 576 (o 768)</li><li>Height: 1024</li><li>CFG Scale: 3 -&gt; Steps: 15 CFG Scale: 4 -&gt; Steps: 20 CFG Scale: 5 -&gt; Steps: 25 CFG Scale: 6 -&gt; Steps: 30...and so on...</li></ul>MY FAVORITE PROMPT:<ul><li>(masterpiece, best quality, highres:1.2), (photorealistic:1.2), (intricate and beautiful:1.2), (detailed light:1.2), (colorful, dynamic angle), RAW photo, upper body shot, fashion photography, YOUR PROMPT, (highres textures), dynamic pose, bokeh, soft light passing through hair, (abstract background:1.3), (sharp), exposure blend, bokeh, (hdr:1.4), high contrast, (cinematic), (muted colors, dim colors, soothing tones:1.3), morbidNEGATIVE PROMPT:</li><li>(worst quality, low quality), negative_hand-neg, bad-hands-5, naked, nude, braless, cross, sepia, black&amp;white, B&amp;W, painting, drawing, illustration</li></ul>YOU CAN ALSO USE THESE NEGATIVE EMBEDDINGS:<ul><li>1) <a target="_blank" rel="ugc" href="https://civitai.com/models/7808?modelVersionId=9536">Easy Negative</a>, <a target="_blank" rel="ugc" href="https://civitai.com/models/56519/negativehand-negative-embedding">negative_hand-neg</a>, <a target="_blank" rel="ugc" href="https://civitai.com/models/116230/bad-hands-5">bad-hands-5</a></li><li>2) <a target="_blank" rel="ugc" href="https://huggingface.co/Xynon/models/blob/main/experimentals/TI/bad-image-v2-39000.pt">Bad-Images-39000</a>, <a target="_blank" rel="ugc" href="https://civitai.com/models/56519/negativehand-negative-embedding">negative_hand-neg</a>, <a target="_blank" rel="ugc" href="https://civitai.com/models/116230/bad-hands-5">bad-hands-5</a></li><li>3) <a target="_blank" rel="ugc" href="https://civitai.com/models/4629/deep-negative-v1x">ng_deepnegative_v1_75t</a>, <a target="_blank" rel="ugc" href="https://civitai.com/models/116230/bad-hands-5">bad-hands-5</a></li><li>4) <a target="_blank" rel="ugc" href="https://civitai.com/models/71961?modelVersionId=94057">FastNegativeV2</a>, <a target="_blank" rel="ugc" href="https://civitai.com/models/56519/negativehand-negative-embedding">negative_hand-neg</a>, <a target="_blank" rel="ugc" href="https://civitai.com/models/116230/bad-hands-5">bad-hands-5</a></li><li>5) For MEN images: girl, woman, female, tits, <a target="_blank" rel="ugc" href="https://huggingface.co/Xynon/models/blob/main/experimentals/TI/bad-image-v2-39000.pt">BadImage_v2-39000</a>, <a target="_blank" rel="ugc" href="https://civitai.com/models/56519/negativehand-negative-embedding">negative_hand-neg</a>, <a target="_blank" rel="ugc" href="https://civitai.com/models/116230/bad-hands-5">bad-hands 5</a></li></ul>Ho creato più di 1000 immagini per essere sicuro e quando parlo di un errore rate di circa il 5% naturalmente intendo i classici corpi allungati o con doppio corpo dovuti ad una risoluzione troppo elevata.A volte possono comparire tre braccia o tre gambe, ma è un altro tipo di errore, non legato alla risoluzione, perciò l'ho escluso dal calcolo dell'errore.Se volete evitare questo genere di errore, basta scendere ad una risoluzione 1024X576px, per rientrare in sorta di zona sicura.Le mani!!! Finalmente sono riuscito a migliorare le mani!!!Ho detto migliorare, ma non sono ancora perfette e sono contentissimo dei dettagli che hanno raggiunto le creazioni.Ci sono altre cose che vorrei migliorare nella prossima versione (o meglio sto già lavorando per eliminare eventuali tagli/artefetti che a volte capitano in faccia)Vorrei anche migliorare l'angolazione delle ombre e altre piccole cose (come i denti).Però posso dirmi assolutamente soddisfatto di questa versione.Ora un po' di chitchat per dirvi quanto è stata sudata questa versione.Il PC con il quale sto trainando, ogni tot tempo (sto cercando ancora di definire la causa) decide di impiantarsi, perciò mi tocca spegnere e ripartire con il training (e a volte sono fuori casa e perdo giornate intere).Il traning è durato circa 490 ore, che equivalgono a circa 21 giorni continuativi di training.Per trovare il fix delle mani, dopo aver avuto il modello finale, c'ho lavorato per ulteriori 2 settimane.Ripeto, è stata dura, lunga e difficile, più di qualche volta ho pensato di buttare via tutto, ma alla fine, grazie anche ad un po' di fortuna, sono riuscito a creare questa versione di AniVerse che mi sta veramente soddisfacendo.Spero che vi dia le stesse soddisfazioni che sta dando a me.Ci risentiamo presto... con un nuovo progetto (che è già in corso d'opera da un po', ma in questo momento in cui scrivo, non ho la più pallida idea se porterà ai frutti desiderati).Samuele

aniverse_v20HDPruned.safetensors

V2.0 (HD) - Pruned

V5.0 - Pruned

V5.0

V4.0 - Pruned

V4.0

V3.0 - Pruned

V3.0

V2.0 (HD)

V1.6 - Pruned

V1.6

V1.5 - Pruned

V1.5

THX Ed. (1.4) - Pruned

THX Ed. (1.4)

V1.3 - Pruned

V1.3

V1.2 - Pruned

V1.2

V1.1 - Pruned

V1.1

V1.0 - Pruned

V1.0

Files with SHA256: 7e776b531fae4b3608f2059114fef59285a040dd25d8898a43b6ab64540efe7c