SexGod PinkCherry v1.8 (see updates below)
Updated: Join our discord server: https://discord.gg/GCrr4Cj3D
-Mini Max H3 Pink Cherry Alpha now in testing at https://huggingface.co/SexGod1979/PinkCherry_MiniMax-H3/tree/main/alpha-0.2-testing
Version 1.8 is out, the entire model was reworked and trained with a new training pipeline. An updated workflow with negative video/audio prompts is available at
https://civarchive.com/models/2828422/pink-cherry-ltx-23-uncensored-workflow
GGUF, fp8_scaled, bf16 and int8 versions are available for download.
This is a great krea2 lora for creating realistic nsfw key frames https://huggingface.co/lodestones/Kroma/tree/main
distil lora is available at https://huggingface.co/Lightricks/LTX-2.3/blob/main/ltx-2.3-22b-distilled-lora-384-1.1.safetensors
uncensored gemma-3-12b text encoder is available at: https://huggingface.co/DreamFast/gemma-3-12b-it-heretic-v2/tree/main/comfyui
Model type: checkpoint, supports I2V and T2V. fine tuned from LTX 2.3 base model.
Description
v1.5 fp8
FAQ
Comments (61)
My favorite checkpoint right now! Any plans on updating the int8 version to 1.5?
I just posted an int8 convrot version for v1.5, running some samples through it now
Thanks for uploading it :)
@sexgod1979 Results are really good, thanks!
Do you have a preferred way to receive tips? Incredible work!
No tip needed, just happy to hear comments that people are enjoying it. Thanks!
1.5 is a really big update, and it's much better than 1.0. Good job.
What actions are already available on the model? Please add breast sucking if it's not included yet. Thanks!
great question. The dataset is large but I have to test to see what the model learned and didnt learn, I am listing what I have actually tested in the model description now and will update it as I further test
@sexgod1979 Thanks a lot for the added description! Very helpful.
Would you consider adding the action I asked?
@shukutan yes I can work on adding breast sucking
@sexgod1979 Yeah... I want that sucking action too (a man sucking a woman's boobs),,, Even from up close
Workflow v1_5.json does not seem to have added LoRa functionality.
Is there wil be gguf vetrsion?
+1 for this, please! Have mercy for lower VRAM user!
What is the minimum VRAM?
Thanks for the int8 love!
Very impressive good work
Question, and i apologize for any stupidty this may possilbly convery on my end, but I truly think you're model might be one of the best if i can get past some flaws. I'm using the INT8 version? A LOT of video noise compared to my usualy fp8s? audionoise overlaps as well? almost like the distillery is week? is this a normal thing? or maybe i'm unaware of certain settings that go along with INT8?
are you using my updated workflow and the correct distil lora?
version 1.5 of the workflow (completely different from v1.3) is at https://huggingface.co/SexGod1979/PinkCherry_NSFW_LTX23/blob/main/workflows/PinkCherry_LTX23_NSFW_Workflow%20v1_5.json and you will need the correct gemma and distil lora from my huggingface page. I dont get anything unsual from the int8 version, make sure comfyUI is updated and you have the correct int8 nodes though I think like comfyui-kitchen. fp8 and int8 are not the same beast. If the int8 gives you issues just use the fp8 version instead. int8 is pretty new for comfyUI. are you doing i2v or t2v?
I'm using the in8 version and it's really good, I didn't compare directly but I think it's much faster than fp8, I'm using the distilled-lora-1.1_fro90_ceil72_condsafe with 0.6 strength and 13 steps/sigmas
@sexgod1979 Thanks for the response. Even using your workflow i got some interesting stutter, HOWVER, when i LOAD the INT8 model into comfyui there are a ouple of settings to change and I'm not very fluent with them, its the load difussion model INT8 (W8A8) node. I was starting to wonder if i was using the correct loader node.
@jdo7496517 yeah I dont use that W8A8 node, or least I have not tried it. I just use a normal comfyUI load checkpoint node for bf16, fp8 and int8 PinkCherry models. COuld be that node doesnt play nice with something, I just ran the int8 on an I2V prompt (gave it an image and a prompt using the load checkpoint node and my workflow) didnt run into anything odd looked great for me
Subtle things in the training data is also important, like nipple play, posing sexy, undressing in a sexy way, groping and fondling. Otherwise porn stuff will override everything.
yes you are correct
20,000 videos of gangbangs only, got it. Im kidding.
@sexgod1979 Hahaha! i just prefer sexy subtle stuff, and your amazing model can do this so far, so yeh, hope future updates also can and maybe even better with the stuff i mentioned before.
@AiCrafter yes I agree, I also enjoy the playful fun sexy time stuff
Nearly indistinguishable quality from fp8, lower VRAM usage, slightly smaller file size, faster inference for older video cards. Sometimes actually slower on newer cards, but usually as fast as fp8.
Great model, int8 working really well with zero extra loras, only the distill one at 0.8str. will post videos soon
great, looking forward to seeing the videos!
If I use the workflow which is listed here I only get garbage after the upscale. what may be wrong? And is there a way to use the int8 model without distilled lora and what settings are to made then?
maybe show me some screenshots of outputs/workflow by hitting me up on chat, hard to tell what the issue is without seeing it
@sexgod1979 it looks like this https://photos.app.goo.gl/dYBbZm8XXsEyXLWp9
In my other workflow I used for Sulphur the model works, but I guessed maybe your workflow gets better results. ;)
@meilenstein79277 ive never seen that before, I meant I need to see the workflow nodes, something isnt right.
@meilenstein79277 also I would test the fp8 version to make sure its not something odd about int8, might lead us to what the cause is.
Awesome WF+model (1.5).Provide excelent quality out of the box without need to understand loads of "spaghetti" and dozens of "WTF they are doing" nodes. Rare nowadays.
Really great model and WF. Thank you! I2V works out of the box. I have a question, when I switch T2V to true i.e. no ref image, I'm getting a ton of errors with nodes missing inputs.
I could go through and bypass the nodes and whatnot, just wondering if there's something I'm doing wrong.
using the latest workflow v1.5 on my huggingface page? any missing node errors?
@sexgod1979 But thats an I2V workflow. Any way to convert it into a T2V one?
@123sirako123521 it has a toggle/bypass option for text to video near the bottom
an excellent model, in terms of results and artifacts, everything is good and stable, unlike eros, but in terms of body movement and liveliness of the scene, eros will be better, just try to generate the video with the same seed and promptom on your model and on the eros model, you'll understand what I mean, tweak if possible the life of the characters and the model will be perfect, well, my personal desire is spanking) for some reason, all ltx models have a weak spot) I write with a translator, I apologize if something is not clear)
if you haven't already try adding the furry nsfw lora and JoyAI-Echo-content_r256 (instead of LTX-2.3-OmniNFT-RL-Lora_bf16) you'll get much more dynamic movement and prompt adherence. or at least it did for me with the Int8 and FP8 versions.
@tanowski1mh360 Haven't found the Lora you're writing about, is there a link?
@tanowski1mh360 I haven't tried EROs since v1, do you know if its worth trying? Int8 versions
love people who just reupload others wfs and call them their own without any credit lmao
I modified runes workflow slightly, but you are correct. If I dont offer a simple workflow already setup for them I get a lot of questions. I didnt think Rune wanted to be associated with this type of content. I can definitely mention it on the model description and will do so. Thanks
Working well for me, but like the other versions, action can be a little lifeless. I use the distilled LoRA you specified, but is there another regular LoRA I can add to help with sexual positions that you might recommend?
This is the best int8 NSFW LTX model. Awesome work!
Any way to improve motion and face expressions? it's very poor on the 1.5v
try v1.6, I think it will be better
Ok. I need help. The Fp8 Version seems unbeatable in terms of speed (sageatt_qk_int8_pv_fp8-cuda++) but some of you said that int8 is actually faster? please share some info. I'm getting 6.45s/it in FP8 and 19s/it in Int8.
Blackwell architecture here. I2v only (never used T2V)
i think if you have cuda 13.1 or higher and the right pytorch (Idk if that architecture or whatever), the int8 will properly go faster. I mean sageatten might be giving you alot too, and it might actually be faster.
Looks like we don't have an active mirror for this file right now.
CivArchive is a community-maintained index — we catalog mirrors that volunteers upload to HuggingFace, torrents, and other public hosts. Looks like no one has uploaded a copy of this file yet.
Some files do get recovered over time through contributions. If you're looking for this one, feel free to ask in Discord, or help preserve it if you have a copy.