[Return] [Bottom]

Posting mode: Reply

BBCode
(for deletion)
  • Allowed file types are: gif, jpg, jpeg, png, webp, webm, mp4
  • Maximum file size allowed is 25000 KB.
  • Images greater than 255 * 255 pixels will be thumbnailed.
  • 103 unique users in the last 30 minutes (including lurkers)




    First
    [1]
    Last

    Starting a thread to share my testing and findings with the new Anima model, download here: https://huggingface.co/circlestone-labs/Anima

    My main goal and fetish is cooking and cannibalism so thats where my testing's focused on. Will be posting generated images.

    Details to know and why I'm interested:
    -Its a new model in testing that generates very fast without much hastle, expecting the full release to be slower but people already made checkpoints to speed it up.
    -Model advertises that can do both tags and natural language or a mix of both. With clever wording, you can get good results.

    Issues:
    -Its a model on preview and its not the full thing yet
    -Its a new architecture and its mpt compatible with flux or sdxl/illustrious loras, requires its own lora training.
    -Proper guro things like beheading, guts, blood, etc etc aren't implemented (or I haven't found the wording for them)
    -Doesn't work with controlnet, doesn't work with regional prompting.

    I'll be taking requests so I can explore different ways of cooking people. I can do women, loli and futa.
    >>
    >>
    Imapelemts seem to be very finicky and not doable. Will research on how to have a perfectly straight human on a spit.
    >>
    Boiling seems to work but I'll need a better color for the skin.
    >>
    Last picture for today. Frying seems to be good if you inpaint. Hoping this is useful to someone.
    >>
    How about bolting and throat-slitting? Would be curious to try turning some girls, or boys into slaughtered livestock.
    >>
    >>16946
    I tried throat slicing but it didn't work out at first, then I also figured out that most of my gens were "wrong" because anima's so weird that it didn't understand my concatenated prompt and was just using the raw one without the quality tags. Might need to find a guro artist to copy from that's on the database...
    >>
    I've also been testing this model, and I think it's really solid with a lot of potential for guro works.
    There's no ControlNet yet since the official base model hasn't finished training. However, there are already a few community finetunes up on Civitai. I personally like AnimaYume the best aesthetically.

    This model uses danbooru tags as visual anchors (What elements are presented in the image) and natural captions for element compositions (How are these visual elements organized).

    >>16946
    Here is one example for throat-slitting:

    masterpiece, best quality,@kase daiki,newest, nsfw, explicit,1girl, kneeling on floor, arm behind back, open kimono,floral print,large breasts, thighhighs, crying with eyes open, slit throat,blood, guro, 1boy, holding weapon, holding knife, blood on knife.In a brightly lit chamber, a young women kneeling on floor, her arms restrained behind her back.She wears an open kimono with floral print and black thighhighs, her large breasts exposed. Her eyes, wide and unblinking, showing anguish as her throat being sliced open, dark red blood spilling out, flowing down her chest and body, the red shows strong contrast on her pale skin. Behind her, a small boy clutches a knife with black handle in his right hand, his face twisted in cruel resolve as the knife stabs her neck, its blade disappears into her flesh,slitting her throat. His left hand grabs her head, as her head tilted to side.
    >>
    >>16950
    Uh, today I learned I guess. Just to check, danbooru tags are used like a base for things that are in the booru you want to add to the image, then natural language is used to make arrange those tags into the specific actions? Training a personal anima lora currently but that seemed to bring SOME results, will reroll and refine soon. I may be asking the moon for some preview model too lmao.
    >>
    Using AnimaYume v0.2 finetune with my recent stuff.
    https://civitai.com/models/2385278/animayume?modelVersionId=2753918

    I am finding it so much more versatile, both on the base and on finetunes, then stable diffusion based models like Pony or Illustrious. The expanded reasoning and prompt adherence of the text encoder allows you to do a lot more with just simple plain language, peppering in some tags.

    I tend to do a mix of generating and traditional digital work to get my finished pieces, but here is a cherry picked image I got from a batch of 10 with no modifications or loras, just the model and the prompt working.

    I don't know, you still have to do more work to get good results, but it's so much easier than the old days of 6 months to 2 years ago.

    Prompt:
    ---
    realistic, sketch, greyscale,

    @adler robert, @chifudoon, death metal cover, low light, black background, star_\(sky\), full_moon,

    scene: a satanic rite - Background - cultists gathered solumnly chanting in shadow to watch the rape and murder of a young girl.

    Scene - In Focus - An adult man and child (loli) are nude on a sacrifical altar and the man is raping the girl vaginal. She dead eyes and mouth open and her chest is covered in blood and she has a chest wound. The man is gleeful and thrusting his pelvis, his large penis in her vagina, the adult man is holding a knife, the knife is covered in blood and is dripping blood. He is kneeling on the altar with her hips raised to meet his, her legs spread and her back is on the altar flat, arms hanging over the side.
    ---
    >>
    >>16956
    Yeah, I'm enjoying the step up in technology. I recently joined the slop game and I can't say how bad it was early to generate something but I've noticed how limiting sdxl is with using the specific tag and not deviating much then rerolling a lot to get what you wanted, and if there wasn't a lora you had to get fancy with it. I saw that a second preview's been released, unsure what changed but it felt the same. Also, unsure if you're using aura flow but its been a game changer, high number for more adherence, low number for more detail.
    >>
    >>16961
    I have it but I don't use it like I should, kind of set and forget, be useful to keep that high/low thing in mind for different phases. high for t2i and low for i2i.

    Cheers~
    >>
    Yeah, feels like it has a lot of potentail.
    >>
    here is a tip for you - don't use loras, when you use lora you hope it will do everything for you, but the reality is on first gen the head is good on second gen the legs is good, etc, hard to get everything right instantly, by trying do everything right in one click you're dramatically reducing your chances for good result. instead learn how to do inpainting, you can even roughly combine successful parts from good gens in photoshop, then do masked inpaint over connecting parts and it will fix for you. or you can regenerate only a certain part multiple times then once result is good save image and add again to inpaint another part.
    >>
    >>16978

    Facts, and Anima does inpainting really well with the InpaintModelConditioning node. Makes editing a almost there gen into something good so easy.

    This combined with a lora turned on for the express purpose it covers is actually pretty OP with Anima. Get a good base, turn on your lora, inpaint the area you want it to fix up, and you get to keep the rest of the image intact as is.
    >>
    >>16981
    Yeah, I saw that the inpaint was great. I need to get better at it but this was doable because of it. Also, 2 girls does wonders for decapitation + body since it considers the body one girl and the head another.
    >>
    Anybody having problems getting it to show conventional gore and blood: Make SURE your booru-style tags have the underscores replaced with spaces. It gets MUCH gorier when the gore tags have spaces and not underscores.
    >>
    >>17012
    What inpaint model are you using for anima?
    >>
    >>17101

    I don't use a special model myself. On ComfyUI I just use a node that you hook up. Your image and the mask you're using hook up along with conditioning, then you hook up the latent and conditioning to your sampler. Just mask out the area you want with the mask editor and keep the denoise below like .70 for most tasks, or be ready to describe the area specifically. Works great most of the time with Anima.
    >>
    >>17101
    No real specific inpaint model, just the basic model conditioning this person has posted >>17119
    no fancy stuff needed,

    >>17092
    I don't tag with underscores but also I can't seem to get good amputations
    >>
    >>17123
    Anima can't do amputation I think. I've tested with many prompts, but didn't get any good amputations.
    >>
    >>17147
    >>17123
    What have you been trying? What I've found works is some combination of guro, blood, gibs and a description of the effected areas.

    ---
    best quality, score_9, guro,

    1girl, solo, sailor jupiter,

    pinup, happy, contrapposto, right hand on hip, looking at viewer, blood, gibs, left arm severed and the wound bleeding,
    ---

    This was with Anima Preview v2
    >>
    >>17147
    It can.
    >>
    >>17154
    I mean Anima cannot do the incision of amputation.
    >>
    Anima is very promising
    >>
    >>
    >>
    >>17263
    you mean like stump insides? I haven't tested it extensivelly, but quick prompt looks already promising, i feel like if you refine it and try different wording it might be even better
    >>
    >>17281
    also this one is single gen, i feel like you can get really amazing results buy generating base amp image -> inpainting the meaty insides of the stump -> inpainting the bone in the middle
    >>
    >>17270

    can i ask what prompt about hanging by neck?

    "hanged" or "hanging" prompt is not work well in anima preview ver2
    >>
    masterpiece, best quality, score_9, explicit, @ebifurya. Vertical full shot of two slender, pale-skinned girls suspended by rough ropes tightly around their necks in a brightly lit room with window blinds casting grid shadows, hanging straight in the air, them feet don't touch the ground. Their hands are tied behind their backs. The girl on the left has black hair styled into long braids tied with black ribbons; The girl on the right has long blonde hair in a single thick braid; The most striking detail is the gruesome trauma to them torsos: their abdomen has been sliced open horizontally, revealing raw pink intestines and internal organs hanging out. Both girls eyes are closed, looking as far down as possible; them chin is resting on them chests, them mouths slightly open and them tongues hang out. Them faces look pale and relaxed.

    They are nearly nude, showcasing extremely emaciated physiques with prominent rib cages and hip bones clearly visible against their pale skin. A trail of dark red blood runs down from their necks, continuing across their stomachs and down their inner thighs. The blacked-haired girl is bare below the waist except for black Mary Jane shoes with white socks on her feet. Soft, diffused daylight streams in from the right side, highlighting the sheen of moisture on their skin and casting soft shadows to the left, emphasizing the texture of the ropes and the wet appearance of the blood stains. Flies are flying around the girl's abdomen
    >>
    >>17285
    thank you sir
    >>
    >>17281
    Damn, what a genius you are, thanks for sharing!
    >>
    slightly different artstyle prompt and yume v3
    >>
    >>
    headshot seems to work too
    >>
    A bit more of a brutal one but I had an idea and thought to do it. I'm sure there's a niche somewhere that'd love to know that you could do small bites and missing pieces.
    >>
    Been awhile but I've learned a lot. First concept lora: Spitroasted. https://www.mediafire.com/file/udne8btzcrxnmdu/Spitroasted.safetensors/file
    Tags: Spitroasted (base)
    Impaled side, impaled front, impaled back, (angles)
    Cooked, dark skin helps if she looks raw.
    >>
    >>18407
    Nice! Now I can finally cook something with anima
    >>
    Hey, been a bit. I've been inspired by Mrik's checkpoint and loras so I've been working on a multiconcept lora for my specific fetishes. The lora includes:
    -Spitroast
    -Cooked bodies
    -Cooked breasts
    -Grilled penises
    -Seasoned bodies
    -Carrot in anus/pussy

    Its not finished and I might be doing things wrong. If someone has has experience on them I'd love to read about it. A preview on the cooked breasts.
    >>
    >>18972
    Cooked breasts are ana amzing fetish


    I would reccoemnd doing separate LoRAs for each of those concepts. It usually works better this way
    >>
    >>19015
    Thanks. I'll try doing separate loras in the future. Meanwhile I work on the loras, I'll post the current "big lora" I made so people can use it in the mean time. It is not perfect, it breaks a bit, but it should produce results.

    Download link:
    https://www.mediafire.com/file/f2d6xlt2wpu1yru/CanniMix.safetensors/file

    The general trigger word is "cannibalism" and more specific ones are:
    -"cooked_body" For cooked people, may need "dark skin" if the cooking isn't happening.
    -"cooked_breast" For cooked breasts, as before may need to specify "dark skin breasts".
    -"spitroasted" "impaled" "all the way through" for spitroasts. I add "steel skewer" and similars to specify.
    -"penis on grill" and "grilling penis" for dicks that go on grills. Doesn't work that well but it provides cooked dicks. May need to specify.
    -"dumbbune" an artist I really enjoy and I've tried to get the style from.
    -"seasoned" for seasoned bodies. May need to use "spices" or "sauce" for extras.
    -"carrot in anus" to stuff you meals with, also works with "carrot in pussy".

    If you still want to see full tags used or forget a tag and want to check them, I use https://xypher7.github.io/lora-metadata-viewer/

    I recommend doing some tags before finishing with a proper natural language prompt, I'm getting better results that way.
    >>
    >>19027
    terrific
    >>
    The people behind kohya_ss have started working on providing anima with controlnet. I'm seeing some decent results, but of course its still in the works and the base model isn't even fully released. Links:
    https://github.com/kohya-ss/ComfyUI-Anima-LLLite
    https://huggingface.co/kohya-ss/Anima-LLLite

    [Top]

    Delete post: []
    First
    [1]
    Last