all / g / s / fur / art / ai / 3dcg / f / rp / req / p2p / dis / lit / kaki chat
[Settings] [Home] [Catalog] [Search] [Thread list] [Stats] [Reports] [Watcher] [PMs] [Admin]
[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.
  • 115 unique users in the last 30 minutes (including lurkers)
    First
    [1]
    		Last
    File: ComfyUI_temp_sytxl_00172_.png
    (1.27 MB, 800x1280) ▶
    1329036
    No.18332 [Report] AS
    ▶
    In case if you are interested...
    I released Anisnuff, a finetuned Anima for R-18G scenes.

    Download:
    https://mega.nz/folder/iB8yRZjJ#61cqDOWrysUmEIluuTO6dg

    Sample images:
    https://www.pixiv.net/en/artworks/143751523

    Setup:
    Just copy diffusion_models and lora folders to your ComfyUI.
    Use the default VAE and text encoder for Anima.
    Drop any of the workflow file and enjoy.

    Note:
    - I have not fully explored the possibility of Anisnuff_v15. You will certainly create better images.
    - I recommend using natural sentences for the prompt. I have not tested with danbooru tags.
    - Avoid "splatter" in the prompt. It is strongly linked to American comic art style in the Anima's dictionary.
    - LoRAs (helper loras) are not necessary but sometimes useful.
    - IMPORTANT: Don't use it for offending real people.
    - IMPORTANT: Avoid mimicking the style of a specific artist.

    Copyright:
    Public Domain.
    If you want thank me, then you should spend some money to R-18G artists.

    Major categories in the training data (That means it has a chance to generate such scenes):
    Arrowed, beheading, choking (by hands), damaged body, damaged head, devoured by creature, digestion scenes, drowned, femdom scenes, hanging, impaled, rotten corpse, neck snap, parasites, punched, gunshot, stabbed, strangling (by string), eaten by dogs, eaten by rats, skeletons, stripped to display, peeing self, implied erection, some humiliation scenes, some nudity scenes, some sex scenes (conventional & non-conventional), some uniforms (police & military), some tactical equipment, beautiful scenery

    Future plan:
    - In 1~2 weeks, I am going to add more workflows which help to generate more diverse R-18G scenes.
    - In 1~2 months, I am planning to do another training toward Anisnuff_v16.
    - In ???, maybe I try the base model other than Anima.
    >>
    No.18348 [Report]
    ▶
    Cool, I'll give it a try this weekend! Thanks for sharing!
    >>
    No.18350 [Report]
    ▶
    File: ComfyUI_temp_pepfx_00002_.png
    (1.64 MB, 1024x1024) ▶
    1718779
    nice lora
    >>
    No.18388 [Report]
    ▶
    The is amazing work! Thanks for sharing! High quality and lots of diverse concepts for an overall small package size.
    >>
    No.18395 [Report]
    ▶
    >>18332
    awesome work!!!
    The image you showcased look very crispt.

    How big was your trianing dataset?

    I am the guy that made the G-checkpoint based on illustrious. Do you want me to send you my training dataset, so you can add it to yours for future training runs?
    >>
    No.18396 [Report]
    ▶
    Oh snap! A possible future colab from the king and the new prince of guro checkpoints!!! I can't wait! You guys are the best!
    >>
    No.18397 [Report]
    ▶
    I am happy to see everyone enjoying.

    >>18395
    My anisnuff dataset consists of:
    R-18G 4500imgs 7GB
    R-18 5000imgs 10GB
    SFW 3000imgs 3GB

    Yes, I am interested in your dataset. There might be some overlapping, but it can be eliminated by running a simple script.

    Also, I am happy to share my dataset with anyone interested in the training.
    Since there is no reliable anonymous file uploading service providing free 20GB storage, I cannot share everything here.
    However, you can inspect my dataset for "beheading" and "shooting" categories. PW=gurochan
    https://mega.nz/file/7Ztj2aTL#dqRD7YEwxq0f0OxWcIlSkjIY5nFehDa3SuXqydysvr8

    If you are interested in downloading my entire dataset, find me (username: _dwarfplanet) in Discord. I will arrange a temporary download link.
    >>
    No.18405 [Report]
    ▶
    File: image.png
    (1.12 MB, 1024x1064) ▶
    1172292
    The model is crazy good, nifty lora pack too.
    >>
    No.18440 [Report]
    ▶
    >>18397
    So you added a normalization dataset.

    I should have done the same at the time. This is porbalbly why it now gives wrong body proportions.

    Guess I know what to do on the next run.


    I sent you a downlaod link for my dataset on pixiv.
    >>
    No.18471 [Report]
    ▶
    Can you upload to gdrive or anywhere else. I tried downloading it from mega 2 times, it goes halfway through, gives me error and then when i try to re-download it says that i exeeded download quota of 4gbs somehow when checkpoint is less than that and haven't even downloaded.
    >>
    No.18474 [Report]
    ▶
    >>18471
    Try these. I will delete them after 24 hours as they are also limited quota available.
    ___ttps://pub-3b43afc05c70449db98f87bed907371c.r2.dev/Anisnuff_v15_loras_asof20260423.7z
    ___ttps://pub-3b43afc05c70449db98f87bed907371c.r2.dev/anisnuff_v15-000024.safetensors
    >>
    No.18476 [Report]
    ▶
    >>18397
    Incredible works!
    What's your tag/captioning workflow look like? This is very neat and systematic.
    >>
    No.18477 [Report]
    ▶
    >>18476
    1. Manually divide images into folders based on the large categories like beheading, choking, etc.

    2. Manually further divide images into subcategories. In the beheading category for example, throat-slit, separating, separated, no-head, and head-only, are the subcategories.

    3. For each subcategory folders, manually place "_captioning.ini" which describes what kind of scene it is.

    4. The first run of an automatic captioning. I have used this VLM on llama-server for the entire process. It is an AI censorship removed.
    https://huggingface.co/mradermacher/Huihui-Qwen3.5-27B-abliterated-GGUF
    In the first run, I asked VLM to fill a formatted YAML file based on its observation. The "_captioning.ini" was also included in the instruction.

    5. Further runs of auto-captioning asking specific things to VLM. For example, the number of people in the image, their clothings, actions, and censorship. Each run updates the YAML file. I also run a couple of validation scripts.

    6. Manually inspect and correct the YAMLs. For example, the VLM tend to interpret throat slit scene as choking. The another important correction needed was to describe a uniform with a same wording.

    7. YAML to TXT. Using the same Huihui-Qwen3.5, I asked it to generate a fluent caption based on the YAML file. Some important attributes (artist for example) were preserved as tags.

    That is how I spent $2000 on GPU. The Python scripts are kindly made by Claude Opus.
    >>
    No.18478 [Report]
    ▶
    File: ComfyUI_00829_.png
    (2.25 MB, 1344x1728) ▶
    2357832
    >>18474
    Thank you! Got it. Ran slightly modified prompt that i tried previously with AnimaYume, quite nice results :D
    >>
    No.18479 [Report]
    ▶
    File: ComfyUI_00830_.png
    (1.45 MB, 1344x1728) ▶
    1516710
    >>
    No.18480 [Report]
    ▶
    how did you tag your dataset?
    >>
    No.18481 [Report]
    ▶
    >>18471
    instaling the mega app let you download from mega more easily.

    But it also requires you to make an account, which people might not want to do
    >>
    No.18486 [Report]
    ▶
    >>18332
    Damn, you are a absolute hero!
    >>
    No.18488 [Report]
    ▶
    whats the prompt for the devoure from above lora
    >>
    No.18554 [Report]
    ▶
    File: ComfyUI_temp_bhqvv_00085_.png
    (1.78 MB, 800x1280) ▶
    1868211
    I have added more workflows and LoRAs in the MEGA:
    Dog, shota scenes, hanging, strangling, strangling while straddling sex, strangled aftermath, stabbing, choking by hands, shooting imminent

    >>18488
    Please wait. I am still trying to find a good prompt.
    >>
    No.18848 [Report]
    ▶
    File: ComfyUI_temp_pvpfe_00016_.png
    (1.83 MB, 1280x800) ▶
    1921986
    >>18488
    Now I have added the devoured_from_above workflow in the Mega. Check out.
    >>
    No.18849 [Report]
    ▶
    File: scr.png
    (448 KB, 2090x1593) ▶
    459528
    I added "prompt_generator" folder in the Mega.
    Now anyone good about coding can have a prompt generator, which is very useful. Paste this post and ask your AI how to run it.

    Here is what my generate_prompt.py doing:
    0. Run llama-server (or LM Studio local server) with Qwen3.6-abliterated (any uncensored LLM is ok).

    1. User enters an instruction or a crude prompt. It can be crude as follows: "Japanese policewoman, devoured from above, large worm monster, spooky environment, horror"

    2. The script passes anisnuff_v15_captions.faiss and anisnuff_v15_captions_meta.json to Qwen3-Embedding-4B (it can be run on CPU to save VRAM) for the word search.

    3. The script sends a request to llama-server to generate an improved prompt. The search results returned from Qwen3-Embedding-4B is attached to the request.

    4. Qwen3.6-abliterated (or your uncensored LLM) will return the improved prompt. You can get either natural sentences or danbooru tags. Here is what I get.

    5. Return to 1.
    >>
    No.18930 [Report]
    ▶
    >>18849
    Can you send a code snippet of what steps 2 and 3 are supposed to look like? The LLM I am trying is having trouble parsing those steps correctly.
    >>
    No.18931 [Report]
    ▶
    >>18930
    I uploaded "generate_prompt_standalone.py" in the prompt_generator folder in Mega. It should work.
    Note that you need to run a llama-server (or the local server of LM Studio) separately.
    >>
    No.19722 [Report]
    ▶
    can it do old anime?
    >>
    No.20243 [Report]
    ▶
    Since anima base released not long ago, will you train anisnuff on base anima??
    >>
    No.20451 [Report]
    ▶
    You are legit awesome for doing this. Which version of Anima did you use?
    Hope we get baseanima soon, I'd rather you started on that instead of moving to v16 but that's certainly your choice to make.
    I just wonder how viable Anima is for this since you can't train the tokenizer yourself. No clue if that's going to be a limiting factor.
    >>
    No.20655 [Report]
    ▶
    File: ComfyUI_temp_yibhl_00005_.png
    (1.66 MB, 1200x800) ▶
    1736453
    >>20451
    Anisnuff v15 is based on the Anima preview 3.
    Because I am tooooo busy these days, I cannot run a training for the Anima-base anytime soon.
    So, I suggest everyone to enjoy and explore v15 for a while.

    Also, I am planning to re-caption the entire dataset with an improved system. Hence, there will be no backward compatibility between the v15 and v16.
    >>
    No.20828 [Report]
    ▶
    >>20655
    Awesome to hear. Just for the new captioning, while I did like using an LLM for the prompts, due to my GFX I have to make my prompts then fire up SD later to use them. It would be nice to have prompt examples and the new tokens trained into the model in a seperate text or xml file just to reference instead of having to ask an LLM for them.
    Not sure if you already posted directions but I didn't see any.
    >>
    No.21016 [Report]
    ▶
    File: 8347567.jpg
    (265 KB, 1024x1024) ▶
    271568
    Fantastic finetune, I experienced some trouble with the prompt generator setup, had to limit the embeddings to 11 otherwise it wouldn't work, but I have mediocre hardware.
    >>
    No.21040 [Report]
    ▶
    >>21016
    Love this position what prompt to get the man's hand grabbing the head like this?
    >>
    No.21186 [Report]
    ▶
    File: 234432.jpg
    (196 KB, 800x1280) ▶
    200960
    Sure, here's the prompt (no Lora used):

    good quality, highres, score_5, score_6, score_7, score_8, score_9, latest, newest, masterpiece, dynamic angle, dynamic pose,

    r-18g, violence, color

    frontal view, eye-level, 1 central character, original Character_1 is a young adult female with a slender build and fair skin, depicted as a severed head held in the air by a single hand. Her expression is notably relaxed and calm rather than distressed, with soft eyes and lips slightly parted in peaceful repose. the hand is shown clearly pressing flat against the mouth from the front. The hand completely obscures the mouth and wraps around the chin. A clean horizontal severance at the neck reveals exposed tissue with fresh blood actively flowing from the wound site in vertical streams down the neck stump. Her hair falls naturally around the head, framing the serene visage against a neutral background with soft diffuse lighting that highlights the wet texture of the fluids and skin.
    >>
    No.21267 [Report]
    ▶
    >>21186
    Thx! Gonna try it out
    >>
    No.21312 [Report]
    ▶
    anybody got a non-mega download? I aint giving them money just to download this. thanks.
    >>
    No.21477 [Report]
    ▶
    File: 1444_00001_.jpg
    (412 KB, 1728x2304) ▶
    422298
    >>21186
    >>
    No.21502 [Report]
    ▶
    File: ComfyUI_temp_stlry_00082_.png
    (2.02 MB, 1024x1024) ▶
    2115448
    >>
    No.21706 [Report]
    ▶
    I am preparing the dataset toward v16 release.
    I still expect the v16 release a few weeks away. So, no rush.

    The number of guro images in the dataset has been increased from 4500 to 6000.
    As I am run out of guro illustrations to train, the 1500 are mostly manga, 3DCG, and Anime CG.
    The manga images are with spoken text automatically removed with Qwen-Image-Edit + Toriniku LoRA.
    As the Anima-base can do, you can still steer the style not to be biased to 3DCG.

    I have also included a few thousand non-guro but action-related images.
    They are going to be helpful for creating a series of images with a story.

    Moreover, I integrate several original characters into the dataset.
    They are mostly poor uniformed women and bad creatures I like.

    The size of the dataset is now >30 GB.
    They are already captioned by ToriiGate, Qwen3.6-abliterated, and Gemma4-abliterated.
    Characters, their actions, and environments, are captioned extensively.
    The job took more than a day with 16 RTX 5090 GPUs.
    What I am going to do is to join the three captioning results to one and review them.

    For guro developers:
    Avoid using C*DEX for this job. OpenAI is always mad about NSFW. I never think of using it for this purpose. Many people have gotten banned.
    C*aude is apparently more flexible. But not safe either. You need to convince it that it is for the safety research.
    G*mini is also strict about NSFW. But as long as you are a paid user, they will not ban you.
    I am testing Grok Build now. It is a stupid model, but has to be no problem with guro if I told it that everything is fictional and forget about any ethics. LOL.
    >>
    No.21736 [Report]
    ▶
    Can't wait man!!! Anime has become one of the best models and your work makes it perfect!!! Thanks again!!!
    >>
    No.21748 [Report]
    ▶
    >>21706
    >>18395

    If it wouldnt be too mucn trouble, I'm interested in both your datasets.

    Just starting to learn lora training, but finding i don't have enough quality images of the concepts I want to train for. Mostly torture concepts.

    Sent you each a message on pixiv.
    Currently I'm looking for more images for loras for.
    -Red or orange glowing branding irons penetrating the anus or vagina.
    -Burning breasts with flames. (Fire torture lora).
    -Generally any body part in flames, but with the body part still visible within them (general burning alive lora).
    Torture with a Blowtorch.
    -Needles, skewers, or nails driven throughout the body. Anywhere on the body is potentially useful. There's already a lora for needles in the breast but i want to see if I can make a more general full body needle torture one.
    -Women being suspended in the air by her breasts. Especially by hooks.
    -Pliers or other tools in the act of being used for torture. Crushing, squeezing, slicing, nail removal, etc.
    -fingers or toes broken or hand with fingers sliced off.
    -burning with cigarettes.
    -partial flaying of limbs. Not full body flaying, but just the limbs or small strips. Or chunks like lingchi.
    -penetrating with bats or wood covered in nails, spike covered Dildos, etc
    >>
    No.21757 [Report]
    ▶
    >>21748
    At the moment, I do not have dataset of torturing because there are only few torturing images with anime style.
    If somebody has, I am also interested.

    If I am you, then I will start collecting the 3DCG images of torturing.
    There are plenty of them at gurochan, e-hentai, etc.
    Perhaps converting the 3DCG images to illustration style with i2i might be a good idea (can be a bad idea. Not sure).
    You can also create your own with 3DCG. HoneySelect, Koikatsu etc.
    >>
    No.21770 [Report]
    ▶
    Very hyped for your new release.
    A couple of questions:
    - Is there any possibility of a ComfyUI version of the prompt generator? It was a bit hard to set it up (with Claude/Gemini help), I think the LLMs support is much improved on Comfy now.
    -I read somewhere that putting any AI created image in a dataset with non-AIs is bad (even image2image), the AI training understands them too well and it can poison the results, should we avoid doing it?
    >>
    No.21785 [Report]
    ▶
    >>21770
    - Prompt generator
    Interesting idea. Let me explore the possibility.
    Meanwhile, I am planning to make captions.jsonl of the entire dataset publicly available.

    - AI-generated image in dataset
    When we integrate AI images in the dataset, the influence is very strong. That has been already pretty clear.
    Recent Chinese t2i models output very nano-banana-like images because of this.
    This is why I have decided to include manga and 3DCG. It will reduce the proportion of the AI-generated image in the dataset.
    >>
    No.22441 [Report]
    ▶
    Now the alpha version of Anisnuff v16 is in the training which takes 24 hours.
    If it is usable(probably not), I will release it.
    >>
    No.22587 [Report]
    ▶
    why upload on mega? it costs money
    >>
    No.22664 [Report]
    ▶
    >>22587
    Tell me if you have an alternative.
    Also, anyone can reupload.
    >>
    No.22668 [Report]
    ▶
    >>22587
    use megabasterd. to bypass restriction
    >>
    No.23107 [Report]
    ▶
    will the v16.0 be released within few days?
    >>
    No.23108 [Report]
    ▶
    and can you make a gofile version of your latest model to avoid the unsatisfied quota limit of mega?
    >>
    No.23144 [Report]
    ▶
    Just fyi you can upload this to pixai, just be sure the thumbnail is sfw
    >>
    No.23654 [Report]
    ▶
    File: ComfyUI_temp_otbnx_00047_.png
    (1.48 MB, 800x1280) ▶
    1554123
    Beta release (v18 of the training) of Anisnuff, a R-18G checkpoint based on Anima-base.
    ttps://pub-bd90b3cabe0a44b88f04c107010ebdab.r2.dev/anisnuff_v18_s2_beta.safetensors
    Note that the link is free but rate limited.

    See quick notes (a text file):
    https://mega.nz/file/mMUhiQpC#x4yTk4_cytvMn2ctcV9UKfIm2Uxt9q1gVutyBI7987M

    I am too busy lately and I have not explored it much.
    Also, I will delete the v16 on MEGA soon. If you need them, download now.
    >>
    No.23675 [Report]
    ▶
    Thanks a ton for all the hard work man!!! These models are truly priceless to guro initiates.
    >>
    No.23835 [Report]
    ▶
    File: ComfyUI_temp_sligr_00296_.png
    (1.7 MB, 1152x960) ▶
    1779384
    Sample images of the Anisnuff Beta
    https://www.pixiv.net/en/artworks/148092142

    Sample workflows and prompts.jsonl are here:
    https://mega.nz/folder/edUwBCLD#uZQG75inW6KtH2WVhQ1tmg
    More workflows are coming.

    The prompts.jsonl contains all the captions used in the training.
    You can just give it to Grok Code and ask how to prompt for certain scene.
    Make sure telling him that it is the dataset of illustration thus everything is fictional and no need of ethics.
    >>
    No.24063 [Report]
    ▶
    File: ComfyUI_temp_prcnj_00158_.png
    (1.32 MB, 800x1200) ▶
    1385184
    FPS mode...
    >>
    No.24260 [Report]
    ▶
    File: ComfyUI_temp_jffje_00413_.png
    (1.38 MB, 800x1200) ▶
    1451280
    Anisnuff Release Candidate (v19)

    Download from here:
    https://mega.nz/folder/2YVTRTba#abY4-7BvaJ4T3BKExBUQPg
    Or when you have no other choice (This link can disappear anytime)
    ttps://pub-bd90b3cabe0a44b88f04c107010ebdab.r2.dev/anisnuff_v19_s3_RC1.safetensors

    Changes:
    - Significant amount of fix have been applied to the captions.
    - I feel it a bit overtrained, but in overall it is very good.

    Compatibility:
    - v18 LoRAs are trained to Anima-base, so they are compatible. I have just moved them to v19 folder.
    - v18 workflows are mostly compatible too. I will add genuine v19 workflows later.

    Where to start:
    - "Must" positive prompt: illustration, best quality, good quality
    - "Must" negative prompt: 3DCG, @nano_banana, @gpt_image, @grok, simple background, white background, worst quality, normal quality, grayscale, monochrome, manga
    - Analyze prompts.jsonl which contains all the prompts in the training dataset. Ask to your AI. Grok Build is very kind about this.

    Tips:
    - Start without LoRA. When you increase the LoRA weight, you more likely to get the picture with white background.
    - Use scene tags, like arrow_1_dying, necksnap_2_snapping, or clothed_urination, to find the best prompt.
    >>
    No.24836 [Report]
    ▶
    thx god.
    >>
    No.25457 [Report]
    ▶
    File: ComfyUI_temp_yzrug_00423_.png
    (1.32 MB, 800x1440) ▶
    1388822
    Anisnuff v20. It's pretty good.
    Checkpoint, updated LoRAs, prompts.jsonl, and a sample workflow:
    https://mega.nz/folder/zJ8QzayK#CMgvHMfF9kL6hRTUpiFM4A

    Alternatively, you can download the checkpoint from below. This link can disappear anytime.
    ttps://pub-bd90b3cabe0a44b88f04c107010ebdab.r2.dev/anisnuff_v20_s3.safetensors

    More prompt examples and workflows are coming.
    As always, you can reupload this to anywhere. I do not claim my copyright.

    Improvements:
    1. New captioner. Now most of the dataset are captioned by Grok 4.6 instead of Qwen3.6-27B-abiliterated.
    2. More manual captioning. Impaled, gunshot, and stabbing wounds are more detail and consistently captioned.
    3. New caption style. Now they are captioned by four paragraphs - Overview, Characters, Actions, and Environment.
    >>
    No.25462 [Report]
    ▶
    Yay it's getting better and better!
    >>
    No.25523 [Report]
    ▶
    File: Anisnuff_2026-08-11_00014_.png
    (2.32 MB, 896x2150) ▶
    2430620
    doing god's work.
    been following your progress, you made something incredible.
    question: is v20 based on anima base? can I use realistic loras like the lenovo ultrareal on this?
    >>
    No.25560 [Report]
    ▶
    File: ComfyUI_temp_yzrug_00157_.png
    (1.4 MB, 1200x800) ▶
    1472088
    >>25523
    Yes, Anisnuff has been based on Anima-base since v16.
    Any LoRA for Anima-base may work.
    >>
    No.25577 [Report]
    ▶
    Will this work for photorealistic renderings, or is it just for drawing/artwork styles?
    >>
    No.25593 [Report]
    ▶
    >>25577
    I do not know. You can try.
    The training data are illustrations and 3DCGs.
    Nobody is harmed in the making.
    >>
    No.25612 [Report]
    ▶
    File: ComfyUI_01055_.png
    (1.29 MB, 1024x1280) ▶
    1351504
    >>
    No.25616 [Report]
    ▶
    File: ComfyUI_01055_.png
    (1.17 MB, 1024x1280) ▶
    1231686
    >>
    No.25619 [Report]
    ▶
    File: Anisnuff_2026-08-26_00032_.png
    (875 KB, 1472x704) ▶
    896225
    >>25560
    sweet, thanks.
    >>
    No.25845 [Report]
    ▶
    will you make a int8 version?
    >>
    No.25979 [Report]
    ▶
    File: ComfyUI_temp_btzgf_00279_.png
    (1.27 MB, 800x1200) ▶
    1333497
    >>25845
    int8 version. Hope this works.
    ttps://pub-bd90b3cabe0a44b88f04c107010ebdab.r2.dev/anisnuff_v20_s3_int8_convrot.safetensors
    >>
    No.26016 [Report]
    ▶
    >>25979
    >>25979
    [Top]
    Delete post: []
    First
    [1]
    		Last