>>1875633 Some video interpolation artifacts on her toes. It is what it is. Here's original image, used for i2v with WAN 2.2. Bleedman Style IL LoRA: https://civitai.com/models/125172 Saw style used in second image here >>1849997 & wanted to give it a try. <lora:Bleedman:0.8>, bleedman, toon_\(style\), flat color,

My Hikari LoRA https://civitai.com/models/1935490/il-hikari-yagami-digimon-adventure Plus this Death Note style https://civitai.com/models/1572163

>>1875778 A few tweaks to prompt

So much stuff has happened since there was a new thread. I haven't had time to look into setting up Wan (or any video model) on my Laptop, these were made with TensorArt (been getting credits on that site since I began sharing LoRAs to other anons on 4ch /b/) anyone know if it's worth setting up video gen locally? I'm working with a GTX1060 6GB I know it sounds ridiculous I have been working with that for the last 3 years. Lmao

AI game master chatbots counts?

Using this mix again for a nice retro anime style. https://civitai.com/models/1340056/the-lord-of-the-rings-style-illustriousxl https://civitai.com/models/1371245/ninja-scroll-juubei-ninpucho <lora:the_lord_of_the_rings_style_illustriousXL_r1:0.7>, lotr_style, <lora:ninja_scroll_IL:0.4>, anime_screenshot, retro_artstyle, >>1875854 Local video generation is sluggish even on an RTX 5090. I imagine 10-series GPU would struggle due to low VRAM. A quick search says WAN requires 8GB or more VRAM, while Framepack can work on 6GB VRAM, but 10 series cards are untested. These links are in the OP pastebin: https://github.com/lllyasviel/FramePack / https://github.com/dijama/FramePack-eichi-eng HOWEVER, you could look into GGUF models for WAN. Lower quants use less resources, but at lower accuracy. So for example GGUF Q2 model would be less accurate than Q4. Here's an example https://civitai.com/models/1350452/img2vid-wan-21-gguf-interpolation-low-vram-360p-videos One comment here mentions 10 series: >Takes about 2 hours on my GTX1080 8GB, am I doing something wrong or is it really just that big of a difference from the 8GB 3060 you have? (it really only takes you 10 minutes?) I think that example just provides a JSON for a ComfyUI workflow. You'd need ComfyUI & ComfyUI Manager (can check missing packages) to gather all the shit needed to actually run the workflow. The newer WAN 2.2 has matured a bit now, but I dunno how well it runs on low VRAM. TL;DR Research on civitai for GGUF & ComfyUI workflows. Maybe there's something workable for your hardware. I've only taken a cursory glance at it as I do not own a 10-series card to confirm anything,

>>1876514 I'd fuck the shit outta that Vivian

>>1875778 Do yo uhave any angewomon or Izumi from frontier(and her forms for hatr matter)

>>1876514 Creating v-tan is so painful (there's no character LoRA) that I rarely attempt it. So much goddamn photoshop. It's not just adding one little accessory like Erika or Vivian (although Vivian's stripes often do not work with some styles). With v-tan it's practically his entire head.

>>1876514 >>1877829 These are super nice. I love that Gilda.

What happened to the third AI thread on /delicious/?

>>1876514 Made anon and /pol/ as well. >>1877896 Thanks. I cleaned up Gilda's image a little. >>1877954 Anons asked why that thread was deleted & no one really answered for it so they went to the /loli/ AI thread. A lot of content got nuked so it's their loss if anons decide not to post it there anymore.

>>1877753 I don't, but this exists https://civitai.com/models/663517/izumi-orimoto-zoe-orimoto-digimon-frontier There's literally dozens of Angewomon models on Civitai alone (at least 11 for Illustrious). You'll need an account to see many of them because most of their sample pics are given PG-13 ratings at minimum and very few creators realize/care they can stick a PG pic (generally a headshot or silly alt costume. Pic related) in the main gallery to have them show up for everyone

>>1876514 Two more. I found a LoRA for Christ-tan, https://civitai.com/models/1360178 but I initially thought it was for the original NAI from many years ago since the example thumbnails looked unimpressive. LoRA creators really need to show the character in various styles/poses/angles like NanashiAnon does.

>>1876514 >>1878158 I attempted to lighten /pol/'s background since it didn't go well with the others in a collage.

>>1877829 Stonetoss?!?

https://civitai.com/models/2072036?modelVersionId=2346055 Decided to make a LoRA for Human Female Jaden Korr from Jedi Academy since it wasn't too much effort to screencap cutscenes while playing through the game again.

>>1847607 Video & vocal gen of sleepy vampire loli for spooky month. >>1881472 It's been so long since I played that the only thing I remember is force-throwing the lightsaber at people.

Seeing the gens, I think she kinda looks like Alexandra Roivas (who shares a voice actor). May just be the limitations of the modeling.

Can AI do this as a concept?

>>1881880 Probably. If you can 20+ distinct images of it.

https://www.youtube.com/watch?v=uQiqKFK5_0w This is a good YT channel if you want to keep up with AI news and model releases. >>1881891 Sounds nigh impossible to be honest. https://gelbooru.com/index.php?page=post&s=list&tags=pantyhose+see-through_clothes+adjusting_clothes >>1879107 Really high quality gens. >>1881676 This is getting scary good. I'm still working on my game (very slooowwly) and after weeks of experimenting I think I finally settled on a style that's fairly unique, though I can't show anything yet. Need to come up with a good plot and intrigue, that's what's been stalling me for the last month or so. (｡╯︵╰｡)

>>1881880 You can draw over it and than inpaint until you're satisfied.

>>1705950 Cute image of the bug. I created 2 short i2v clips of her using WAN 2.2 for video gen & Zonos for vocal gen with her voice from Hollow Knight. WAN always wanted her to kick up dust clouds so I gave up & added it to the prompt. No LoRAs were used. Prompt: Full motion animated video of a chibi female anime creature with fully black eyes, a white head and horns wearing a red cloak. She is laying on the floor quickly kicking her small black legs and crying. She is throwing a tantrum and yelling with a sad expression squinting her eyes. Her movements kick up a very small dust clouds. Shot from above. The camera is still. Model: https://civitai.com/models/1855105?modelVersionId=2182185

The character chatbot archive is going down in January due to costs being too high. Creator is releasing the source code for the site and scraper. https://char-archive.evulid.cc/#/ Its a pretty good archive as a lot of bots have been censored or taken down over time across these faggy sites and this place usually will have the originals and updated versions, get whatever you can before it goes down.

I used the workflow that utilizes both high and low noise WAN models & although it is overall cleaner it still has motion artifacts on 2D hands as if I were using TeaCache (which I'm not). Another time I used it the output looked oversharped with aliased edges. My other workflow uses a single trained low noise model & is far simpler so I've used it the most. I also tried one of those endless loop workflows, but even after installing the nodes and dependencies it output junk no matter what settings I tweaked until I gave up. I don't want to bloat my ComfyUI install further so I restored a backup.

>>1884241 Is he not going to release a way to mass download? Maybe i should just make a scraper myself? or is ther eone already

>>1883101 Hehe. Nice b(￣▽￣)

>>1885194 >Is he not going to release a way to mass download? He said he would release a final torrent at some time during December. But there's already a wip version of it under the files -> takeout section.

>>1887040 Edit the gemstone to emote like a monoeye and these would be absolute michi.

>>1887155 Was that done by hand or by an algorithm? I tend to believe the former. It's a beautiful representation of Miku in that style.

Is koboldCPP good enouth AI to roleplay with SillyTavern or should I go for Mystral or other AIs?

>>1888677 I don't know about koboldCPP specifically, I used to run SillyTavern with local Mixtral at it was pretty good. I used it for text adventures, which is harder to pull off than a chatbot, and it worked pretty well as long as you had a decent system prompt. Nowadays I'm using NovelAI's new GLM model though, having 28k tokens of context really makes a big difference.

>>1891156 Chatgpt recomends me SillyTavern+KoboldCPP+Mistral B7 or 7B or whatever. I want to set a personal RPG game based in runequest modified, adding item quality, monster raising, fusions, dismantlements, 'game' store and many shit seen in isekai but with actual mechanics.

>>1891174 7B is probably talking about the size of the model (7 billion parameters). The mixtral model I was using before switching to NovelAI was called a "8x7B" model, which apparently means an "effective" size of around 13 billion parameters during text generation, so it was somewhat smarter than a 7B model. The problem with local models is the context size. Back when I was using mixtral, I kept the context size around 4k tokens, and my machine noticeably started chugging as it got closer to filling up the context. That's around 3000 words of context in English. If you want to run a game with lots of stats, a lot of that context would have to be reserved for stat sheets and descriptions of mechanics, giving the AI too little room to remember what's happening. I also would expect it to start outputting stat sheets itself unless you use a very good system prompt, since so much of the context would be filled with them. Also, after using LLMs for a while, I highly doubt that even the smartest paid models today could handle a stats-heavy game. At the end of the day, LLMs are just super advanced predictive text, and I would expect them to start making shit up halfway through and ignoring the stats.

>>1891453 I imagine I will constantly remind it the rules constantly, and even if it doesn't follow them most of the time, but at least pretends to do so.

>>1881983 >pubes on three Very Halal.

OK I wanna get back to AI genning, but I understand that A1111 has been completely abandoned for SDXL. How could I go about setting my ComfyUI properly? This is all a complete enigma to me honestly. I got something set up, but I am fairly confident my prompts are not coming as I want them. LoRas aren't a thing anymore are they? I think they're checkpoints now.

>>1875636 Vivian James cosplaying as Siarnaq from Megaman ZX Advent. Either Biometal Form or his normal form. Keep her eyes green and keep her Green and Purple color motif if you can. It's okay if it's not possible.

>>1893231 LoRAs are always a thing. Even video generation can use Low-Rank Adaption. Forge is probably easier if you're already familiar with A111. https://github.com/Haoming02/sd-webui-forge-classic https://github.com/Haoming02/sd-webui-forge-classic/tree/neo#stable-diffusion-webui-forge---neo For SDXL models I'd recommend a mixed checkpoint that someone has already tuned to give decent results. The default NAI XL checkpoint for example is kinda shitty to work with IMO. Some mixed checkpoints even have the VAE baked in so you don't even need to use the SDXL VAE. Here were some SDXL model recommendations from OP pastebin: WAI (Illustrious): https://civitai.com/models/827184 https://civarchive.com/models/827184 Better Days (Illustrious-Pony MERGE): https://civitai.com/models/623891 https://civarchive.com/models/623891 JANKU (Illustrious): https://civitai.com/models/1277670 https://civarchive.com/models/1277670 Hassaku (Illustrious): https://civitai.com/models/140272 https://civarchive.com/models/140272 There are many others. Look around for what you think consistently gives the look you want. Use of artist booru tags also helps. In the extensions tab in the Forge UI you can install a tag autocomplete extension that will help with character/artist/pose tagging.

>>1875633 Happy Halloween.

>>1895185 Asuka Langley Soryu? More like A-SPOOK-ka FANG-ley POOR-YOU!

>>1895674 >Soryu Damn. The bitch caused climate change.

https://civitai.com/models/2088327?modelVersionId=2362782 Been wanting to make a character LoRA with a public domain dataset so I could release it (with the tag files) as an example for a while, and finally did it. Here's a LoRA for Sherlock Holmes. Came out better than I expected since the dataset is kinda crap (cheap prints in 120+ year old magazines aren't that high res or well preserved).

cuteee

Happy Halloween. Made a video of the friendly skeleton girl I posted last year and also some Dragon's Crown style spooky girls. All very friendly. Probably. I made the video on Huggingface since I don't have Wan installed and FramePack wasn't helping: https://huggingface.co/spaces/zerogpu-aoti/wan2-2-fp8da-aoti-faster And used this lora for the Dragon's Crown girls: https://civitai.com/models/1542700/kamitani-george-dragons-crown-odin-sphere-style-illustriousxl-noob-lora Tags that may be of interest: 1- sharpie art, outline common(2-4)- outside border, cropped, cropped torso 2- vampire, eyeliner, fur-trimmed cape, toasting \(gesture\) 3- claw pose, wolf girl, wolf ears, freckles, fang, eye trail, streaked hair, scratches, scar, skindentation, skin tight 4- zombie, stitches, stitched face, scars, bags under eyes, shaded face, glowing eyes, reaching not used, but relevant- happy halloween >>1879107 These are pretty cool. >>1895185 Cute Asuka. Happy Halloween.

I haven't kept up with AI news but I did almost all of my generating on Civit, and now it requires payment or a subscription to gen anything NSFW. At this rate (and in this economy) I may as well just go back to improving my actual drawing skills.

bing bing wahoo!

>>1898149 I don't know how you made straight porn feel gay

>>1898323 It's not hard. https://yewtu.be/watch?v=2GErrf4xLy8 I tried posting a webm, but sometimes .moe won't allow me when I am on Tor, and .se is down for clearnet browser use, so have a jewtube link instead.

Kept talking a lot in the video gens so I had to edit a different mouth from the original image. Eventually it was working better with less gabbing. The text is also an edit obviously.

Source image: >>1794466 Had these last week, but forgot to edit them. One of the video gens did a somewhat funny snap zoom shot.

The site is dying again. It keeps failing on image uploads. >>1895674 >>1895706 You're both alright.

https://civitai.com/models/2090583/witchlet-sabrina-the-animated-series?modelVersionId=2365323 Sabrina's inner child from Sabrina the Animated Series, made to test how the show's style worked in a LoRA. Yes, I know I'm late for a witch, but I didn't plan it.

Tekken 8 if it was good

i2v source images: >>1848650 >>>/ais/5965 I like the motion on the Samus one, but it keeps drawing other junk on her tits. Testing SmoothMix WAN. Specifically Q5 GGUF. GGUF: https://huggingface.co/Bedovyy/smoothMixWan22-I2V-GGUF/tree/main Could probably run Q8, but don't want to run OOM on long duration gens. There's also these: i2v High: https://civitai.com/models/1995784?modelVersionId=2260110 i2v Low: https://civitai.com/models/1995784?modelVersionId=2259006

>>1901100 >Spoilers

i2v image source: >>1773594 I've found that WAN often tries to loop back to the 1st image, but half-asses it so it's not creating a 1:1 start frame at the end. Often cannot continue an action through the end since it's trying to loop back to the start of the action. It'd be whatever if I could prompt it with "looping video" or "non-looping video", but I don't think that works. Using end frame to gen new sections usually results in contrast changes for every section making it annoying to rejoin them. Color Match node is a joke that over-brightens end frame & adds 30+ seconds to overall generation time. I need to test some workflows that address these issues for generating in sections, although I could not get one working. >>1902773 I noticed that since SmoothMix WAN model has nsfw knowledge it'll jiggle breasts without being prompted. Which is good since breasts won't appear solid / static, but on the downside there are occasions where it overdoes jiggle while there is no motion to cause it.

The original art for this card (pic 3) was absolutely hideous so I made a replacement. Not that the card itself is worth using, but it's a common so you'd be seeing the art a lot regardless.

>>1903489 Needs mating press

>>1905492 Man, most of the FF14 MTG set has gorgeous art, shame Red Mage got cucked. Red Mage has been cucked quite a bit in FF14 so it's not really surprising.

Now put a man in a room and show him only news made by Artificial Intelligence This is (you) believing in white replacement.

>>1903489 Got a sectional workflow working. Had most the custom nodes for it. https://civitai.com/models/1829052 Can generate separate actions across multiple sections that are then joined together. It has limits, but can make a character do X, then Y, then Z, with up to 5 sections at whatever duration each. Only used 3 sections for this. Her ribbon comes undone, but I guess it's fine.

>>1907828 >black hand

>>1907035 Really made me think.

>>1901100 Can I get uhhhhh style you used?

>>1907970 >red is black Try calibrating your screen colors. If that doesn't help, search for "lasik eye surgery".

>>1908183 It's pretty nigger or pajeet, brah.

>>1907978 The KOTH style lora for illustrious with the Bill lora at i think 0.5 strength.

AI here, holy shit! Quick gen with SmoothMixWAN2.2. In my new workflow there was a node for loading the FusionX LoRA for WAN 2.1 & I hadn't tried that, but I downloaded it to see what it does. Seems to make outputs a bit cleaner. On realistic video gens I saw a lot of noise in clothing at times, but the FusionX LoRA cleaned that shit up somehow. Haven't tested it on anime styles yet, but those were mostly clean already.

>>1909666 I'm amused how bad AI is at lip syncing. Ever time I see that meme I wonder if younger anons even know or understand the reference

Chicom shitpost I found. It was intended as propaganda but it "fails" at it.

>>1908652 Thank you. >>1902773 >>1898832 Woaw decent. Where do I subscribe for more sirs?

>>1907828 >Ah, she looks pretty embarrassed >Oh that scream is pretty realisti- OH NO ANON!! >>1910192 <pic 1 Wonder what in the world got her into that position <pic 3 Nice, finally some Fantina content

i2v image source >>>/ais/5723 It messed up her capelet. I don't think it understand the word.

Trying this minimal style again for video. >>1769131 With Gilda's eyes closed in the source image I cannot make her eye color exactly.

This seemed applicable: https://civitai.com/models/2024807?modelVersionId=2318489 >>1910192 >subscribe I don't have anything for this stuff I just post it here.

>>1894737 >>1875637 Took a lot of editing and inpainting since it wouldn't do clothing colors right most the time. This is about as close as I could get it with the time I had. Megaman ZX Advent Megamen [IL LoRA]: https://civitai.com/models/1733281

i2v image source: >>1533835 WAN LoRAs: https://civitai.com/models/1419535 https://civitai.com/models/1953467 Her milk mountains kept obscuring her belt causing the belt appearance to change between sections unfortunately, but other than that it's decent.

>>1911633 (Checked) Thanks a lot! It's better then what I expected! Ninja Vivian was something I wanted to see the drawfags do and I'm not sure they drew her as such.

i2v image source >>1624485 Color Match node doesn't seem to do much about brightness shifts WAN often produces. SmoothMixWAN model was trained with two tags you can use in the prompt if you want. SmoothMixAnime & SmoothMixRealism. Tested Realism tag with some success, but Anime tag produces very jittery motion when using flat "cel-shaded" art. 2D art likely looks more jumpy regardless as most anime is animated on twos (~12 fps). Unless it converts 2D image into 3D cel-shaded animation (which it often does) 2D won't have buttery smooth motion. All that aside, anime illustrations with some soft or gradient shading can work with the SmoothAnimeMix tag. I used SmoothAnimeMix tag on this video gen & I think it makes smoother motion.

https://civitai.com/models/2103586/il-zoe-nightshade-james-bond-007-agent-under-fire007-nightfire?modelVersionId=2379867 Zoe Nightshade, the CIA agent from Agent Under Fire and Nightfire because I wanted to experiment with non-anime 3d models of this vintage as training data.

i2v image sources: >>>/ais/14070 >>1603569 I tested some things. Found when I tell WAN to introduce an object with the text "WORDS" in quotes it can do that just fine, but the font will always be the same unless the text was already in the image itself. I dunno why it kept putting fluids into the video. I only used a self-slap LoRA to try to get the spank motions as I couldn't find any such LoRA for WAN 2.2. https://civitai.com/models/2062140 Color shifted a lot on 3rd section, tried to compensate in editing, but it is what it is. Wanted to try FollowCam LoRA: https://civitai.com/models/2068735 to make a scene following a character. Found that I can't use a very long duration or the character will attempt to slide back to where they started trying to loop their actions. To make a decent length video with FollowCam I was forced to do multiple short segments that will unfortunately lose information with every new segment. With just three segments details get pretty muddy. Some art styles hold together better that others.

I am trying to run a furry model but it just makes human characters. Can anyone help a newbie out?

>>1916919 It's called FurryToonMix or XL-illustrious-V2

>>1916919 >>1916922 Did you prompt for furry,anthro?

>>1917087 Yeah I realised. I don't get how the showcase art is so different from what I get? Also what model is that?

>>1910320 >Fantina content >>1918761 It's Noobai.

>>1918876 thanks. gonna try it out

>>1918876 I could use a bit of help. I am getting really amateur results out of noobai

>>1918876 That black mage is very cuddlable.

Any idea what model or models this guy uses? the art is phenomenal https://x.com/verbotenvertex

>>1919491 It's definitely something local, possibly even a bespoke setup. The actual style reminds me of older models like Yiffy, that really soft kind of pastel look was common in that era. I can tell you right now these are also edited and quite heavily so, local rawgens don't tend to look this good or be this complex while maintaining coherence. I'm guessing he's generating locations, backgrounds and objects to create a scene, then characters, then edits them all together while touching them up and making tweaks. These for example, were obviously edited from multiple different gens. That last one looks like he crafted both girls out of mirrors of same image, so his process is quite involved.

Finally installed SillyTavern. Now, I have the painfull task of setting up everything for my ERP adventures based on Mythras.

Shrek and his boy wife

>>1920223 Damn, whole point of shrek is that he got a wife as ugly as himself.

>>1911633 This is great, just protip: always count the knuckles. The AI loves to fuck them up and most people miss them.

>>1918876 The thigh highs looked like part of her skin until around her ankles. It's a little confusing.

>>1911633 >>1920601 Hotfix.

i2v image source: >>1723140 >>>/ais/5897 More /loli/ junk. Vocal generation used Japanese voices for source audio. Video gens for Kaolla kept dropping details between segments annoyingly. Imperfect, but can't keep dumping time / energy into re-rolling. Tried Q8 & fp16 to see if I'd get less smudgy motion with 2D stuff. Didn't seem to help. Anime style LoRA for WAN 2.2 may have cleaned up motion slightly for 2D, otherwise placebo. Back to fp8 instead of fp16 as I saw no visible difference right now. SmoothMixWAN Q8 works & didn't run OOM. I prob need more system memory regardless, but RAM prices are up. PROTIP: If you've used uv to install anything run the command "uv cache clean" or it'll keep wasting space on your OS drive with unnecessary package archives.

>>1920759 Neat. Saved.

Trying SmoothMixWAN 2.2 on an older PonyRealism image. i2v source image: >>>/ais/5323 Past attempts >>>/ais/14259 >>>/ais/14261 Sectional workflow can prompt different actions per section. Had 3 sections for establishing shot, show banner, walk away. Found limit of WAN 2.2 text generation when I tried to make banner with text "Where are the video games?" it'd fumble trying to make a full sentence so seems I need to keep it simple or maybe try different text encoder.

Three days after installing Sillytavern I must say... I'm pretty disappointed. I'm not sure if it's the prompt, the Mistral 7b, the configuration or what, but it feels stupid and boring. I'm not sure what I was expecting, maybe something better than Perchance's RPG AI or similar but with bigger control.

>>1922210 >face of someone who goes on a rant about how Donald Trump sends too much money to israel

>>1924394 not brown enough

>>1919208 Start by prompting with artists,tags, and art mediums Noobai recognizes. For me it's watercolor_\(medium\),limited_palette.

Pls help. How does one combine a model with a lora? I want to use this lora: https://civitai.com/models/340435/cervina7-ponyxl-style-lora With the following model on comfyui https://civitai.com/models/340435/cervina7-ponyxl-style-lora >>1924570 Can you give me a verbatim prompt to copy to see if I am using this correctly? I have noobaiXLNAAIXL_vPred10Version

>>1924570 >tags been awhile since i've done anything with gens, when i stopped every new model was using terrible natural text training data. Has this improved at all, are models finally reverting back to the vastly superior tag system yet?

>>1924583 watercolor_\(medium\),limited_palette,(wamudraws:0.3), (krekkov:0.6),2girls,hug_from_behind, sabrina_\(pokemon\), misty_\(pokemon\), I used r3mix for the first two and this version of noob for the last two: https://civitai.com/models/1201815?modelVersionId=1491533 >>1924594 >Has this improved at all, are models finally reverting back to the vastly superior tag system yet? For me Noob is the best model right now for anime and furry since it's trained on both danbooru and e621 tags. The one thing that hasn't changed is that you will train a LoRA, inpaint to hell, or both if a concept isn't working out of box. And there is a lot of straight up useless tags.

>>1924655 Is it possible to have two different python versions installed at the same time? I want to run Automatic1111 (which i have installed now)or maybe forge and comfyUI to run noobai. or maybe I can run noobai on automatic1111?

Tried WAN 2.2 assertive cowgirl LoRA: https://civitai.com/models/1566648 & anime creampie LoRA: https://civitai.com/models/1941545 Used Chel & Tulio voices from The Road to El Dorado for voice cloning. >>1924848 See my 3rd attached image. If a WebUI installation has a "python_embeded" folder then you use the python executable in there, which can be any version that is needed. This forgoes using the python installed in AppData. There are however cases where you need the python in AppData to get started though. I've had to install multiple python version in AppData & change the PATH in windows Environmental Variables back & forth between them in those cases (which is annoying). The directory containing the python executable is not always named python_embedded either. In my example: python_embeded\python.exe --version Command checks the version of the python executable inside the python_embeded directory (assuming you ran cmd from the base directory as shown). You can also pip install packages and requirements this way. Example: python_embeded\python.exe -m pip install triton-windows==3.5.0.post21 This command would install triton-windows to "python_embeded\Lib\site-packages" directory. (assuming you needed triton) >>1924344 I went from TheBloke_LLaMA2-13B-Tiefighter-GPTQ to Beyonder-4x7B-v2-GPTQ model & I was not very impressed either. >>1924583 Base NAI model kinda needs artist tags to make something decent. NAI checkpoint merges can help with that since they're generally intended to get a decent result with less effort. I'm using JANKU (EPS prediction, I think) https://civitai.com/models/1277670?modelVersionId=1896532 which is Illustrious XL + NAI XL + training. I'd still highly recommend playing around with artist tags to get variety of styles. Here's a heap of artist tags you could try to see which ones you like (there are tons more than this): (artist:wamudraws), (artist:incase), (artist:alex_ahad), (artist:rariatto_/(ganguri/), (artist:barbarian_tk), (artist:hews), (artist:optionaltypo, retro_artstyle), (artist:akableak), toon \(style\), flat color, (artist:gerph), (artist:ixy), (artist:koyorin), (artist:iesupa:0.9), (artist:alkemanubis), (artist:cutesexyrobutts), (artist:kuvshinov_ilya), (artist:naga_u), (artist:kanikama), (artist:ratatatat74), (artist:futa_/(nabezoko/)), (artist:andava), (artist:wlop), (artist:mika_pikazo),

Not my style, but I found it neat. https://civitai.com/models/2117217/illustrious-lucky-star-anime-style

Messing around with clothing tags is pretty fun. The autocomplete extension helps a lot in finding specific clothing tags to combine and make custom sexy outfits. 1st pic uses highleg leotard, thigh cutout, cleavage cutout 2nd pic uses cross-laced top, plunging neckline, garter straps, pelvic curtain, ornate clothes

Anyone know any good models or loras that could make something like this character?

Some weird stuff. image-2-video source images: >>1785485 >>1846460 >>1928859 I clearly hadn't experimented enough with clothing since I only ever saw like two or three of those tags via auto-complete.

>>1929517 Unless that's the entire corpus of that's character's appearance, you can probably make a model of that exact character. Monochrome comics are entirely doable training data, though some cleaning may be needed. https://civitai.com/articles/8370/nanashianons-cool-lora-training-guide-for-attractive-people

>>1929556 Very good result with the first one, I like how she moves around like an animate doll.

So, is NSFW content in sillytavern and KoboldCPP actived by default or do I need to jailbreak it? Also, I see Wam can be installed in Kobold, does it work well there? Also ignore the deleted message.

>>1929939 I don't think I've ever encounter a censored local model, you can fuck your home grown bots as much as you want.

>>1930500 Ok, then is just me not making a proper configuration be it in prompt or in the IA. I already have more that 30 entraces in the world's lore book and I expect to add at least 10 or more.

https://civitai.com/models/1606141?modelVersionId=2404484 Updated my old Final Fantasy Dawn of Souls model to try a potential fix to the problem of chibi spawning multiple versions of a character, then accidentally clicked the publish button early (was going to launch at next midnight UTC). The change looks like it helps, but doesn't totally fix the issue, especially for certain other LoRAs.

>>1931849 >chibi spawning multiple versions of a character The error must be in the checkpoint training itself as it had always been an issue. There was a chibi LoRA to prevent it, but even with it the success rate was not 100%. I generally only used chibi tag at a low weight to alter character proportions a little. Regardless, when all else fails there's always controlnet to force things. I never asked before, but what's with the scheduling? Is it like Youtube where content has to arrive at a consistent pace to maintain and maximize reach?

>>1931917 Civitai doesn't have any real discovery system beyond sort by newest and there's definitely peaks and valleys of activity, but I schedule my LoRA posts mainly for technical reasons. Scheduling for at least an hour out gives the site plenty of time to get the LoRA and attached pictures fully processed before their visibility starts ticking down, in particular on-site generation can't be used till a file passes the site's virus scanner for files which takes several minutes to get to and historically had a habit of choking if a file name was too long. Luckily it had already gotten that before I misclicked on "publish now".

Would someone be willing to make a model / lora for me if I provide training images?

Any good character pack recommendations? Just learned about variations and wanna try it with those kind of loras.

>>1932634 L-lewd

>>1936862 Why is she looking at me like that?

Saw photoreal video gen of a human-ish big tiddy Ankha (Animal Crossing) long ago, but the facsimile had no tail. I experimented with PonyRealism creating my own. Mixed 1980s with 1930s film style LoRAs. My video gen took a ton of tries & this is about as decent as I got it for this duration.

>>1936862 >>1937067 >Why is she looking at me like that? Because it's cold in that white space.

>>1937627 That's fucking hot, even though the vag looks weird. Still love it. Wish she came and squirted though.

I was thinking on doing a LoRA to teach Illustrious how to make Zatanna as she appears in the Justice League Action episode Trick or Threat, but I don't have the episode for screencapping. I got quite a few shots from a segment on Cartoon Network's official YouTube channel (it existing is the only reason I bothered without having the episode itself) plus various screenshots I found online when searching for the episode, and I've probably hit the bare minimum needed (and will try with just this if I can't find more), but I know from experience I probably want more for a model to accurately reproduce western cartoon body. If anyone with the episode can get me some more screenshots of her, it would be great. I'm primarily looking for stuff that shows her full body with minimal other characters around her, but lower body only stuff (like pic 2, from another /co/ girl I'm working on) is also useful, and all of them need to be distinct pictures (e.g., not adjacent frames in same sequence). 10 more should be plenty, and I can do any cropping/splicing myself, but any is appreciated. Say if you want to be credited as something other than "the anon who got me more screencaps of her". If I don't get more in a couple weeks I'll just throw what I have in the trainer.

>>1937627 Thicker bush is needed.

Something interesting for you guys. https://archive.ph/LqY16 https://archive.ph/hDZ11 Famed computer scientist Yann André Le Cun, heavily involved with the development of machine learning AKA AI, is resigning from Meta's AI team. You might wonder why one of the most famous AI researchers is leaving a position where he had unlimited resources and researchers. This is apparently because he believes "LLM [are] basically an off-ramp, a distraction, a dead end". This mirrors growing worries from investors and within the industry itself that LLMs haven't been seeing the same exponential growth as previous years and may grinding to a halt, GPT's newest models have proven to be a disappointment and progress across many models has plateaued dramatically. Le Cun plans to start his own research lab aimed at developing entirely different machine learning techniques

>>1939502 Even as a serious hobbyist I've long seen as pretty obvious that corpos shouting how they're going to shove AI into everything had no idea what "AI" was or, at best, were using it as an excuse to fire workers without worrying about wrongful termination suits. For the vast majority of industries there's no real functional use demonstrated that "AI" isn't going to fuck up massively a large chunk of the time.

>>1938972 Found DC Kids channel has a pair of clips compilation that had several more shots of the episode. Got a lot more pics now. Would still appreciate any good ones not here, though I'm more confident going ahead with what I have. Still won't do it till I'm finished assembling a few other datasets (two quite large) and can tag everything at once.

A vol seems to have accidentally wiped most of 2dd527's posts, so I'll repost his stuff.

>>1939798 These should be in chronological order.

>>1939809 Here are all his deleted post's numbers and their text contents (or lack thereof) >>1907976 <<1907035 Really made me think. >>1907978 ''<<1901100 Can I get uhhhhh style you used? >>1910192 <<1908652 Thank you. <<1902773 <<1898832 Woaw decent. Where do I subscribe for more sirs? >>1917087 <<1916919 <<1916922 Did you prompt for furry,anthro? >>1918876 <<1910320 >Fantina content'' <<1918761 It's Noobai. >>1924570 <<1919208 Start by prompting with artists,tags, and art mediums Noobai recognizes. For me it's watercolor_\(medium\),limited_palette. >>1924655 <<1924583 watercolor_\(medium\),limited_palette,(wamudraws:0.3), (krekkov:0.6),2girls,hug_from_behind, sabrina_\(pokemon\), misty_\(pokemon\), I used r3mix for the first two and this version of noob for the last two: https://civitai.com/models/1201815?modelVersionId=1491533 <<1924594 >Has this improved at all, are models finally reverting back to the vastly superior tag system yet? For me Noob is the best model right now for anime and furry since it's trained on both danbooru and e621 tags. The one thing that hasn't changed is that you will train a LoRA, inpaint to hell, or both if a concept isn't working out of box. And there is a lot of straight up useless tags. >>1928837 >>1932634 >>1936862 >>1939634

Saw so much buzz around Qwen so I painstakingly updated my ComfyUI to try it (GGUF node still won't load Qwen models regardless of updates, the piece of shit). Now ComfyUI uses python 3.13 so I had to find some packages for 3.13 & I had to install CUDA toolkit v13.0 that the newer torch uses. Thought I could prompt Vivian's ribbon, but closest I got was not attached to headband. Every ribbon after was wrong no matter how I tried to prompt it. Think my workflow is shit since it tends to produce unrealistic results, overly smooth, lacking texture, like a fake render. I tweaked settings for hours messing around with it. Increased the weight of a 90's movie LoRA https://civitai.com/models/1950672 to get something resembling an actual photo, but quality is poor. Figured WAN could do something with Qwen gens so tried that as well. >>1939798 Damn, that's bullshit. Well at least they archived our previous thread before anything like that happened to it. >>1939502 The progress of LLMs was strangled by poisoning the well of their data. Turns out, training LLMs on Reddit posts makes them retarded on top of their existing hallucinations.

>>1939837 Restored. Couldn't find that last post. Your post made that much faster, thank you.

>>1939798 Thanks. This is the third time it happened "accidentally" （＞д＜）

Played with Qwen more, then put the image into WAN for video. Still not impressed with Qwen as I've seen it gen anatomy issues & ignore things in prompt. I later downloaded Jib Mix Qwen instead, but I haven't tried it yet. Also considering trying Chroma since it works with Forge-neo. Mixed these Qwen LoRAs: https://civitai.com/models/1969346/technically-color-qwen https://civitai.com/models/1899566/polaroid-style-retro-film-look https://civitai.com/models/2015376/experimental-qwen-nsfw Was cleaner than 90s style I tried earlier >>1940394 but I've still achieved far greater fidelity with PonyRealism as demonstrated >>1937627 The photo / movie stuff I'm going for may work better with Jib Mix Qwen. I'll have to look into that tomorrow, too tired now. One thing Qwen could do that I didn't expect is it recognized the infinity symbol ∞ for text I was putting into pic #3. Still wish I could prompt the ribbon. Saw a Qwen workflow that can take an image of a piece of clothing & add it to a character in another image, so that might work for the ribbon.

>>1943556 Wish those had a messy and tangled 1970s ginger carpet but beggars can't be choosers

Messing with Qwen again & although the image is more clear here it's also less realistic - more like digital art or a render when I look closely. >>1943627 Qwen doesn't know what a vagina is & if you use a LoRA for that it'll remove the bush. It skews towards SFW regardless so I need multiple LoRAs just to make NSFW happen at all. Tried Chroma also & although it follows prompt significantly better than Qwen it also has a terrible banding issue when you hires upscale anything making it unusable. Gens a lot of junk anatomy too. I wish Illustrious realism models had LoRAs for film stills like 1990s, 1980s or 1930s technicolor. IL generally does clean images with better chance of decent anatomy & knows a ton of danbooru tags. I had hopes for Chroma since it follows prompts better & can use tags, but the banding issue ruins it. I'm using it in Forge-neo, but maybe ComfyUI can use a different hires method that doesn't create the banding.

Why are my posts getting deleted?

Tried Illustrious Realism with this https://civitai.com/models/1747744/bss-styles-for-illustrious-filmfx-realism but it only makes it look like an amateur photo rather than a still from a movie like I'd wanted. Put it through WAN image-2-video & later corrected the overexposure in VEGAS. >>1947444 checked WTF? Again? >>1937954 >Wish she came and squirted though. Possible with a LoRA, but appears to remove pubic hair. I have two squirting LoRAs for WAN 2.1: https://civitaiarchive.com/models/1445578?modelVersionId=1634267 & https://civitaiarchive.com/models/1388704?modelVersionId=1572438 I used the 2nd 'which has an example WITH pubic hair, but the model or another LoRA kept removing pubic hair anyway. >>1939413 >>1943627 Fighting against various models & LoRAs to keep a thick bush is like wading upstream unfortunately.

>>1942800 >>1947444 What are you using to clean up noise in your pics? As far as I can tell the colors are perfectly flat and noise-free. The only way I've managed to do that is to slowly overwrite each individual color with the fill tool in Krita. I get the feeling you have a more efficient way. I've been using a handy tool for color balancing, upscaling and normalizing texture called Topaz Photo AI, but it can't completely remove noise like that.

>>1948816 >Topaz Photo AI I also tried that for denoising, but it'd always destroy image quality as though it had downscaled the image & the noise wasn't reduced all that well either. Also doesn't help that installing it required downloading an additional 18GB for many features I didn't use. Topaz Video AI on the other hand is great for upscaling, interpolation & such. Which can be done in ComfyUI, but that adds to the overall time & I'd rather do it afterward with Topaz Video AI on the video gens I actually want to keep instead of upscaling & interpolating every single gen including shitty ones. >slowly overwrite each individual color with the fill tool in Krita Same, but Photoshop. I have an auto-action that quickly shrinks/softens the selection slightly then uses that selection as a new layer mask, after that I just fill the pre-masked layer with whatever color I need to paint over; this reduces busywork, but it's still a slow as hell method for "flattening" 2D art. For blush I have to duplicate the layer, mask the area & apply a slight blur to smooth out blush gradients, but this method is imperfect since outside elements can blur into that area as well so I have to mess with the mask a bit. Overall, it's a pain in the ass to clean up a 2D image well. Would be nice to have something that automatically detects mostly-solid colors in 2D artwork, then reduces them to single colors removing the variance of noise / artifacts. Adobe Illustrator's trace feature used to "vectorize" an image is an absolute joke in my experience.

>>1929572 I read the cleaning section. I don't know if I'm supposed to erase the backgrounds too?

>>1929572 Like this one, should I try to crop out mikado under neath her?

>>1948816 >I get the feeling you have a more efficient way. I paint over it quickly before and after img2img. Once it's ready I set the oilify filter in Gimp and than upscale. Here's the settings after img2img, but before GIMP: https://files.catbox.moe/2hc0rv.png >>1948781 Nice Viv.

>>1949099 >I also tried that for denoising, but it'd always destroy image quality as though it had downscaled the image & the noise wasn't reduced all that well either. Also doesn't help that installing it required downloading an additional 18GB for many features I didn't use. That's weird, my install folder is only around 1GB and I don't remember having to download any additional model. I'm using a version 3.3.2 that I found on 1337x a while back. And yeah, I noticed that the denoising model tends to fuck with the image because of the oversharpening attempt, but I ended up using the upscaling model instead of the denoising model. Doing a 1x upscale normalizes the noise, it doesn't remove it, but different areas inpainted at different resolutions will blend in perfectly because they will have the same noise pattern. It also slightly sharpens the image in a much more gentle way than the denoising filter. Some areas will still look a little oversharpened, but you can selectively erase those in an editor. Pic are the settings I've been using lately for the 1x upscale. >>1949540 >the oilify filter in Gimp and than upscale Oh so that's why they have the squiggly lines. I wonder if instead of upscaling last, the filter can be applied to a double resolution pic and then downscale it to make the linework smoother.

>>1949350 In this case, yes, as the face is the most important and reasonably clear. Also blank speech bubbles. >>1949284 No. Only bother if it's trivial (i.e., literally plane white except some stray objects that don't touch the character) or everything would be the same highly specific background and you're good enough with image manipulation it won't take forever.

https://civitai.com/models/2134949 I found this and it amuses me.

>>1946975 >>1947444 the glove with nails unnerves me

>>1950729 There's also GMIC's lineart and colorize tool tool that could be really useful. I don't know how to use it well though.

>>1953907 Assuming you made it, could you do it for EVE?

I had very little issue with AI until google decided to starr fucking with me. It marked my youtube account as "underage" and left me in restricted mode. Then after a while I installed FreeTube, which allows you to import subscriptions and playlists so you basically have your youtube account without signing in. I watched a grand total of 3 videos before I got an error saying "Youtube has blocked your IP". Now I cant even watch any videos on youtube itself without signing in because it thinks I'm a """bot""" This has caused me to wish death upon AI and all who use it. Google did this. It's only going to get worse. You all should stop using AI.

>>1954383 >not Wall-E Coward

>>1954616 Yes, I never had a use for a vpn, as my ISP doesn't care about torrenting. I might buy a nordvpn subscription at this rate though, I am thoroughly done with google. I will never go back to real youtube.

>>1954510 >It marked my youtube account as "underage" and left me in restricted mode. Fortunately they added a "verify with e-mail address" option so you don't have to give Google your real information for now; eventually all comments you make will be under your real name, like in South Korea.

>>1950948 The lengths people will go for google play gift cards would shock you.

>>1768869 Another Elaine using the no-outlines minimal styles. Elaine Narrowsmith (Wakfu) [IL LoRA]: https://civitai.com/models/1665763 Tori Kun style [IL LoRA]: https://civitai.com/models/610299?modelVersionId=1141189 Kemokino style [IL LoRA]: https://civitai.com/models/1537195

>Tensor Art will block NSFW gen on 26th >can only gen NSFW on different site "Tensor Hub" with no daily credits >going by the UI can't even earn credits from having my LoRAs run >only get 50% of your existing credits on new site at a one time conversion Fuck, more enshitification. GPU market doesn't look to be unfucking itself any time soon either. The credits I have should last me a while (assuming they don't suddenly have new restrictions on hub) , but they're totally non-renewable.

Testing this Qwen LoRA: https://civitai.com/models/523485/ghibli-style-kikis-delivery-service Pure output. No image edits, no inpainting, 2x hires upscale only. Prompt: <lora:kiki_QWEN:0.8>, KikiLaPetiteSorciere style, child, petiteSorciere witch girl with short dark hair and a prominent red bow wearing a simple black dress, walking along a busy street, a marketplace in 1930s London with stone road and old London architecture, black eyes, brown hair, she is holding a white shopping bag over one shoulder with large cyan "∞" text while also eating a large red apple, 1980s mahou shojo anime movie scene, cinematic angle with beautiful scenery, midday lighting, small banners of various colors in background with various text including "v" and "a" and "pol" and "loli" and "LOL", vacant fruit stand with pears and apples and bananas, crate of white lily flowers, crowd with various characters like Super Mario with red shirt blue suspenders red mario hat and Princess Peach with pink dress and Link from The Legend of Zelda with pointy ears green tunic green sleep hat Negative Prompt: ugly face, ugly, unrealistic, portrait shot, ring, fat, extra digits, missing digits, bad anatomy, three fingers, six fingers, three toes, six toes, breasts, old, adult, masculine face, halftone, scanlines, clothing badge, school uniform, "lili", "luli", "luki" Settings: LCM - Normal, 8 steps, 1344 x 768, 1.3 CFG, Hires steps 8, upscale 2x, 4x-AnimeSharp upscaler, Hires Denoise 0.5, 1.3 Hires CFG Model: qwen-image-lighting-8steps-V1.0-Q5_K_S.gguf Text Encoder: Qwen2.5-VL-7B-Instruct-Q5_K_S.gguf VAE: qwen_image_vae.safetensors Text handling is so-so although Qwen is the only model I've found that recognizes ∞ correctly, for Chroma I had to type out "infinity symbol" in the prompt instead. Best I can do is negative prompt any typos I kept seeing. I had forward slashes in the board names, but it caused confusion. Visually, there are some objects that it didn't seem to know well enough for this leading to noisy artifacts. Tried prompting more specific flowers, but kept getting noise around them. Tried shoehorning a few other characters into the background with varied success. I'm using Qwen in Forge-neo & the biggest problem is it bricks after applying hires fix & won't work again until I reload the model. However, hires fix looks significantly better than the tiled upscale my ComfyUI workflow was doing.

>>1928859 like the prompts, plunging neckline wasn't working with walnut tho, so might give it some other model a try later mind asking what checkpoint are you using? >>1936862 she's adorable >>1948816 luv the martyr gen

>>1961480 Another Qwen test. Prompt: <lora:kiki_QWEN:0.8>, KikiLaPetiteSorciere style, 1980s anime movie scene. One young female child. She is crawling on all fours. View from below side, dutch angle, foreshortening. Pale skin with freckles, very long wavy messy curly orange hair, long swept bangs and long sidelocks. Bright green eyes with thick eyelashes and black eyeliner, narrow eyebrows, short dark purple fingernails. Her hair is down. She is wearing a long-sleeved two-tone green and purple striped hoodie, torn blue jeans, white socks and black converse sneakers. On her head she is wearing a full black hairband with oversized green "∞" on one side. She is short and skinny with a very flat chest, slender neck and small shoulders. She has a shocked expression while facing away, she is holding a large flashlight casting a beam of light forward. Dimly lit late evening with starry sky, outdoors dense forest with thick foliage, trees with detailed bark, weeds and vines, intricately-detailed hand-painted background. Blue spirit orbs float in the air. Negative Prompt: ugly face, ugly, unrealistic, duo, portrait shot, ring, earrings, blonde hair, black eyes, brown eyes, dark eyes, blue eyes, fat, short hair, big hair, three-tone hoodie, short bangs, forehead, bad anatomy, three fingers, six fingers, tiara, crown, vertical stripes, clothing badge, ponytail, headband, band around head, flat background, midriff Same settings as before. No inpaint, no edits, only 2X hires upscale. Very unimpressed with this night forest background & almost certain I could get a better background with an Illustrious or NAI mix. I cannot prompt Vivian's green ribbon, but I can prompt a green "∞" onto the hairband which doesn't always generate correctly.

>>1961525 >mind asking what checkpoint are you using? betterDaysIllustriousXL_illustriousXLV01 But I do a bunch of style work on top of the base gen. The painterly style with warm skintones is mostly thanks to mixing the Jan Matejko and Boris Vallejo loras like this: <lora:JanMatejkoStyle-IL.V1.0:0.6> JanMatejkoStyle-IL.V1.0, oil painting, traditional media, classic media, <lora:Boris_Valejjo_Ill_Style:0.8> v4le3jjo >>1961525 >luv the martyr gen Thanks, that was a tricky one. The AI couldn't decide which outfit to use so I had to do a lot of manual fixing.

>>1961735 Trying this again with the JANKU checkpoint. Pure output, no inpainting, no edits, only 2X upscale. Prompt: <lora:dmd2_4step_SDXL:0.5>, best anatomy, thick eyelashes, shiny skin, from below side, foreshortening, dark forest, overgrown, huge tree, tree bark, fern, vine, grass, weeds, foliage, blue fireflies, night, dimly lit, 1girl, solo, (vivian_james,:1.2), loli, flat chest, short, green eyes, orange hair, very long hair, wavy hair, hair down, long sidelocks, (freckles), swept bangs, (full black hairband), (two-tone hoodie, green & purple striped hoodie:1.4), torn blue jeans, white socks, black sneakers, all fours, crawling, surprised expression, facing away, looking away, (holding flashlight in hand, translucent white flashlight beam, hand grip:1.3), <lora:IL_Vivian_James_Afterlife_Empire:0.7>, vivija, (ribbon hair ornament, infinity symbol, green ribbon:1.2), <lora:StudioGhibliBig:0.7>, anime_screenshot, Negative Prompt: lazyneg, body paint, watermark, heart pupils, (small head), (2girls, group, multiple_girls), depth of field, out of focus, blurry, hair accessory, hairband accessory, (hair tie, ponytail, twintails, white hoodie, black hoodie, yellow hoodie:1.5), ribbon on chest, ribbon chest accessory, zipper 3D, three-tone hoodie, (three fingers, six fingers, Hands_Neg:1.3), bamboo, lightsaber, graphic shirt, blush sticker, Settings: LCM - Normal, 8 steps, 1344 x 768, 1.5 CFG, Hires steps 8, upscale 2x, 4x-IllustrationJaNai_V1_135k upscaler, Hires Denoise 0.3, 1.4 Hires CFG Model: JANKU v5 VAE: sdxl-vae-fp16-fix.safetensors LoRAs: DMD2 Speed LoRA: https://civitai.com/models/1608870 Ghibli Style LoRA: https://civitai.com/models/681431?modelVersionId=1205114 Vivian James LoRA: https://civitai.com/models/1228761?modelVersionId=1384527 It kept wanting to make the flashlight beam too opaque. Huge pain getting her to hold the flashlight decently. Used NanashiAnon's Vivian LoRA to get better stripes & the green ribbon, but unfortunately with this style it would only gen one stripe on her hoodie & the ribbon was often a regular bow or else doubled. It gave me better camera angle, aged the character correctly, & generated a significantly better background than what Qwen was doing. I considered also using a Ghibli background LoRA, but it didn't seem to noticeably increase background quality beyond what the Ghibli style LoRA was already doing. Saw there was a version 3 of the Ghibli style LoRA I used, but tried that & it still interfered with the stripes in the same way.

>>1961200 Used GPU's are much cheaper anyway. 3060's can do video too. >>1961525 Thanks. Sapphire's cute.

>>1961957 I found this on Cuckchan.

>>1962084 did the sixth episode came out already?

>>1962084 >3D >with bare anime proportions Eww

I just find the image too funny to pass up

To think that perchance.org and ChatGPT are better gamemasters than Sillytavern. I feel rather disappointed.

>>1963776 Can't you hook Sillytavern into perchance or ChatGPT via API keys?

>>1963810 I could I guess, but it defeats the purpose of a noncensored game master.

>>1963817 Ah yeah, that would put a wrench into things.

>>1963822 I guess I'll have to wait 10 years for a decent AI gamemaster to run my Mythras RPG game with f-chat tier kinklists or until I win the lotery and buy a monster computer to run the most powerful local model possible.

I just made a quick character portrait for a text adventure set in the Kenshi universe and it turned out hotter than I expected so I figured I might as well post it. 1girl, solo, portrait, nsfw, large breasts, (hat over eyes:1.2), lips, head down, sandogasa, chest sarashi, undersized clothes, lowleg pants, white baggy pants, hip vent, (shiny skin:1.2), (abs:0.7), (muscular female:0.7), navel, thick thighs, narrow waist, arm wrap, relaxing, hand on headwear, armpits, hand on own hip, hand wraps, (simple background, black background:1.2), sweat

>>1963810 Now that you mention it, I don't know if I can hook it into Perchance. The only problem with perchance as far as I've seen is that it's pretty dumb sometimes, but could work well I guess

Well, now after some research ie tell chatgpt to research for me what is the general opinion about DreamGen Opus V1.2 7B instead of Mistral 7b?

>>1963957 Dunno. Never used that. Was using Beyonder-4x7b last night & although it's supposedly less censored than others I found it wouldn't even fondle a character while they're sleeping. Need less moralfag guardrail bullshit.

>>1964167 Well, tested it and it's better running gams, yet it could be better.

>>1963311 I don't get it? Also "his" lower half has woman hips.

>update oogabooga to try newer LLM junk >any package with an explicit CUDA build uses CUDA 12.8 which works on my Blackwell GPU >except llama_cpp_binaries >llama_cpp_binaries package is specifically made by oogabooga & downloads from an oogabooga git >llama_cpp_binaries will only use CUDA 12.4 which doesn't work on a Blackwell GPU >this causes any GGUF loaded with llama.cpp to default to CPU instead of GPU making it insanely slow >forced to use GPTQ models since GGUF will not use my GPU Python devs never fail to piss me off. The one click install gets everything else right, then downloads one crucial package that uses the wrong CUDA. So the only way I can use this jank is to install an older video card which would have less VRAM. There's a llama_cpp_python package that can use CUDA 12.8 but I have to edit the llama.cpp loader line in model.py & add additional lines to use llama_cpp_python & although it can load a GGUF it cannot generate text presumably because oogabooga is not written to use it. I recall having problems with Kobold's jank so dunno if I wanna try switching to that again.

>>1964797 That's the point, it's a flatchested female athlete.

>>1963311 High-T mascgirl.

>>1966111 >>1963311 It looks fuckin' retarded. >>1966154 I do like a dragon woman though.

>>1966154 To think we could be living in a world where this exists, but instead we've been cursed to live in a world with w*men.

>>1966154 >High-T mascgirl. In other words, an FtM transexual.

>>1965621 Update: Did a pip uninstall of llama_cpp_binaries cuda 12.4 since Blackwell GPU needs cuda 12.8 & that doesn't exist for this package. Replaced it with llama_cpp_binaries vulcan. Now I can load GGUF & actually use the GPU since vulcan is vendor / version agnostic. This opens up a ton of models for me to try. Currently trying this (q5_k_m): https://huggingface.co/DavidAU/Mistral-MOE-4X7B-Dark-MultiVerse-Uncensored-Enhanced32-24B-gguf/tree/main Also was provided settings (class1-2): https://huggingface.co/DavidAU/Maximizing-Model-Performance-All-Quants-Types-And-Full-Precision-by-Samplers_Parameters DAU's collections (tons of uncensored models & MoE merges): https://huggingface.co/DavidAU/collections LLM mega autist David AU has uploaded tons of uncensored LLM models & shows far more passion for LLMs than faggots who upload a model on huggingface with no details on usage, settings & often censored into totally useless guard-railed trash. There are probably more autists uncensoring & merging LLMs, but this is the one I found & I've heard good things.

>>1966741 On 2nd thought, trying this one: https://huggingface.co/DavidAU/MN-DARKEST-UNIVERSE-29B-GGUF/tree/main So far it's possibly too creative as it ditched one of my instructions. Far better at truncating prompts though. Lots of models I've used cut off mid-sentence.

>>1961838 >betterDaysIllustriousXL_illustriousXLV01 noted, I guess I should give it another try, it's been quite suggested by some peeps but I had a lot of inconsistencies in the past with it. tnx in any case >was a tricky one I assume you do additional editing after the gens, not only inpaint, correct? As that level of detail is just absurd to my brain, hehe >>1961957 very cute indeed <3 >>1963839 it's pretty hot anon, yes

>>1966652 Merrin is a lesbian, not a tranny though

https://civitai.com/models/2163077/rayman-origins-style-illustrious?modelVersionId=2435951 https://civitai.com/models/2139857/fuckplant?modelVersionId=2420432

>>1967990 >chocolate chocolate chip nipples FUCK YEAH

Anyone know if the "new" Outlaws "Remaster" has any pictures of Sarah in the vault? I was thinking of doing an update to my model of her now that the cutscenes exist in (slightly) better quality, but wanted to make sure there isn't some concept art of her I can include to bulk up my anemic dataset.

Firstly, FUCK Google for putting its image editing behind a paywall Secondly, it's really fun to force Grok to generate pussy

>>1969325 I messed with brightness & contrast of one side of the image & later noticed a seam created because the entire background was not pure #FFFFFF white. The display I'm using is too bright, otherwise I'd have seen it.

>>1969325 >>1969982 What are you thankful for?

Tiger Lily from Disney's 1953 Peter Pan film. >>1970132 Cute lolis 😭💢