Club activities editionPrevious thread - >>3093930 # (OP) #*Request Thread: >>>/y/airt>To make a new thread, wait for page 8 at bump (and/or image) limit in catalog.>ONLINE GENERATORShttps://tensor.art (NO LONGER SUPPORTS NSFW)https://tensorhub.art (NSFW version of tensor, It's no longer free sadly)https://civitai.com (NO LONGER SUPPORTS NSFW, use the link below for that)https://civitai.red/ (Civitai new NSFW page)https://refineforever.com/ (Free but limited options and probably slow)https://tungsten.run/ (10 generations free every day, but the premium it's cheap)https://novelai.net/ (PAID ONLY)https://pixai.art/es (MAYBE?)>GOOGLE COLAB FROM LAST THREADhttps://colab.research.google.com/github/gutris1/segsmaker/blob/main/notebook/Segsmaker_COLAB.ipynb>METADATA & TAGS EXTRACTION/LISThttps://www.metadata2go.comhttps://rentry.org/airt-rs3https://rentry.org/povlist>/airt/'s RESOURCEShttps://rentry.org/airt-rs1 (FAQ, etc.)https://rentry.org/airt-rs2 (Setup)>QDIFFUSION (LOW-END PC)https://github.com/arenasys/qDiffusion>LOCAL FORGEUIhttps://github.com/lllyasviel/stable-diffusion-webui-forge>LOCAL REFORGE *recommended*GitHubhttps://github.comPanchovix/stable-diffusion-webui-reForge>LOCAL MODELShttps://civitai.red/models/2458426/animahttps://civitai.red/models/3005406/moss-imagehttps://civitai.red/models/2732656/krea-2-turboNaixl:https://civitai.com/models/833294/noobai-xl-nai-xl>RESOURCESTraining: https://rentry.org/59xed3 | https://github.com/derrian-distro/LoRA_Easy_Training_ScriptsTags: https://danbooru.donmai.us/wiki_pages/tag_groups | https://danbooru.donmai.us/related_tagAnimation: https://rentry.org/AnimAnonWildcards: https://rentry.org/NAIwildcards4chanX Catbox/NAI prompt userscript: https://rentry.org/hdgcbNAI API/prompting in auto1111: https://github.com/Metachs/sdwebui-nai-api>RELATEDFurry (Male) AI art: >>>/trash/msdgAi art request thread >>>/airt/
>>3096036>If you do make a thread post 5-6 images to get the thread going....
>>3096038I forgor I fell sleep
>>3096048I don't have a lot worth sharing rn tho
>>3096049
>>3096050Please tell me you have something with Ivan
>>3096052That was done with character reference on anima, so this is a test using art of that character
>>3096057
As mentioned numerous times in the last AI thread, I have been practicing with Krea 2 (having found MOSS to be too chaotic, and Anime being somewhat disappointing). I am going to start posting images created with Krea 2 starting Monday the 16th on my profile here:>https://civitai.red/user/gtgiygusThis will included an article with some specific tips and extensions I have found useful. I am adding the information (in brief) here, since not everyone bothers with CivitAI:Firstly, Krea 2 has been reasonably OK to work with. Using the turbo model, and being able to skip both the ADetailer and Upscaling steps, makes generating the image faster than with the older SDXL type checkpoints (like illustrious) while keeping the details and resolution high.The problem is that Krea 2 was not natively trained on either Anime or XXX material. You pretty much need a Lora for everything, and of course the Loras focusing on male anatomy are lacking. A common complaint with the straight male dominated gooners programing this stuff.But, I have gotten it too work reasonably well, good enough others might want to try.
>>3096065I ended up using the base turbo model, no fine tunes. The reason being that the finetunes are mostly created with real pictures and not anime, so it lacks the style I want.>https://civitai.red/models/2732656/krea-2-turboSo lets talk about Lora's. I have needed at least 3 Lora's for every picture. As many as 5 were used for some details, like the penis or anal sex, to look good. You need this Lora (or one like it) to remove the refusal encoded into the LLM, without it, nothing will work:>https://civitai.red/models/2742336/nsfw-krea2-low-vram?modelVersionId=3089248After that, a style Lora was used. Specifically, for what I will be posting, this one here at 0.7 strenght:>https://civitai.red/models/2810202/hes-my-man?modelVersionId=3176985On top of that, a basic male focused lora was used. There are a few, but these worked best so far:>https://civitai.red/models/2732041/krea-2-dicktator-male-focused?modelVersionId=3071591>https://civitai.red/models/2838670/twinks?modelVersionId=3204067This is a non-exhaustive list of some others I needed. Note most needed to be at 0.7-0.8 to not override the anime look.>https://civitai.red/models/2915731/anal-sex?modelVersionId=3448268>https://civitai.red/models/2883732/deepthroat-blowjob?modelVersionId=3259590>https://civitai.red/models/2962273/perfect-bussy?modelVersionId=3356054>https://civitai.red/models/2754165/penis-xpenis-krea-all-round-male-body-model-butt-anus-penis-cut?modelVersionId=3098796
>>3096066Once you have gotten your Lora's hooked up, there are some other issues to deal with. First, the camera viewpoint for Krea is very unrealistic. It almost always shows the character from above, with them looking up at the viewer. Even with strong prompting, you can hardly drag it away. Profile view also sucks, barely turning the character.The solution I found was not a Lora (I tried a few) but an extension for Forge-neo. Note that there is a ComfyUI node that does the same thing. This extension can be found here (you can copy URL into "Install from URL" on the extensions page in Forge-Neo:>https://github.com/flyfront/sd-forge-negpipThis fixes the LLM that Krea 2 uses to accept weights while prompting (like "from above:-3.0"). This is extremely useful, since turbo Krea 2 does not make use of the negative prompt (CFG set to 1 disables it). That makes generation much faster, with the obvious downsides of no negative. SD 1.5 and SDXL (illustrious) have this built in natively (coupled with different LLM architecture) It lets you use negative weights in the positive prompt to act as a poor mans negative prompt, while keeping the speed. The set of prompts I used to control the camera as follows:>(looking up:-3), (high angle:-3), (eye level:1.6)>(profile:3.0), (from side:3.0)This is not exhaustive, and can be used with any word. Vagina and Labia were both used at negative 3 weight in my generation elsewhere.
>>3096067Another issue was prompting itself. Krea 2 (and to an extent MOSS and Anime) seem to work best with detailed and dense natural language paragraphs. In fact, Krea 2 can accept up to 300 tokens, and still functions OK with twice that many. SDXL starts getting confused after 75, topping out at around 150.I quickly found that all the time saved in image generation speed was taken back writing a short story to describe the pictures I wanted.My solution was to lean on a downloaded uncensored LLM model, run on a program called "Unsloth", to make my prompts for me. The following link contains the Qwen LLM model, modified to be completely uncensored:>>https://huggingface.co/DavidAU/Qwen3.8-27B-TURBO-Fable-Cold-Fusion-735-882-Heretic-Uncensored-NEO-CODER-MAX-MTP-GGUF/blob/main/Qwen3.8-27B-TurboFCFusion-735-882-Here-Uncen-NEO-CODER-MAX-LOW-MTP-IQ4_XS.ggufThis runs reasonably fast on my 16 gig vrm/ 32 gig system ram setup. You have to start by creating a "skill" in unsloth to help it understand how to prompt for Krea 2(or anima, or MOSS). I had grok make that for me, and just copied pasted it. After that, just ask what you want. For example:>use @krea2-prompting to create a male focus image prompt. The style is realistic anime. The subject is a 19 year old male with undercut blond hair and blue eyes. He has a slim but toned body and is wearing a black chest harness, black crotch harness, black armbands, black thigh strap and black speedo type underwear. He is erotic pole dancing on a stage with stage lighting. The background is faceless crowd cheering. Use natural language and aim for 175-200 words. Do not use em dashes and skip any technical generator information for Krea 2.You can also expand on the prompt, asking for a series of poses or camera angles, or adding characters in. Then just copy and paste into Forge-Neo. Note that the LLM uses as much ram as image generation, and I could not run them together.
>>3096068Even with all that information being gathered and tested, I have to admit to being underwhelmed. Krea 2 can do a lot of stuff, and I am going to keep trying.However, illustrious (and homosimile checkpoint) are so effortless when it comes to beautiful male anatomy. It understand so much about anal sex and blowjobs that Krea 2 doesn't. The tag based prompting for SDXL, while imprecise, is so much easier than huge paragraphs. I love forcing my characters to play with dildos. Krea 2 can barely make a dildo look normal. Krea 2 still tries to put vaginas on men, and the heads of cocks leave a lot to be desired. Cum is messy, and it puts it everywhere for no reason when added to a prompt. Lora's to fix these issues are imperfect, and trying to figure out the right combination with the right strength is tedious, if not impossible.All in all, I am looking forward to the MOSS anime checkpoint they are developing, and in the meantime I might try to do more with Anima instead. I just don't see a long love affair with Krea 2 the way I have enjoyed with illustrious.I fully admit I am imperfect, and I might be doing something wrong, or missing some vital checkpoint or Lora. Any input is greatly appreciated.
>>3096050>>3096052>>3096057Piers please, there is so little GS porn
short lazy moss prompt experiment:1. pasted https://i.imgur.com/xZY2wo4.jpeg into chatgpt and said "caption everything visible"2. said make it male (without long hair) and make it more risque (thrice)3. pasted into google translate -> chinese (simplified)chatgpt is way more willing to make sexual stuff than I thought, and it wouldn't have taken much effort to make it explicit (just used the prompt 1:1). anatomy weirdness is just the lorastep 3 is because I am pretty sure that, since moss uses a chinese text encoder, chinese will be at least slightly better than english if you wanna bother translating any prompt. it seemed to have better prompt adherence for me at least
What are you staring at?
>>3096050Holy fuck please do more. Especially Piers.
>>3096180Have you any good frontal or good quality pic of that character? Put it here and I look what I can do, sadly I played the game a long time ago and that character is not one that I remember very well skjdf
>>3096181Not perfect but I love the character reference node,
>>3096180I hope this is the correct dude
>>3096183
>>3096185