SexGod PinkCherry v1.6 (see updates below)
** July 26 2026: Uploading GGUF low vram models on the v1.7 alpha model for those that want/need lower vram options for their cards
** July 25 2026: I have posted my alpha version 1.7 model, so those that want to beta test while I continue to train and improve it I would appreciate the feedback as its hard to test and find gaps or problems on my own (I can only watch so much porn haha). audio sync should be fine in v1.7 alpha. the workflow im using is saved in each video, feel free to drag into comfyUI.. FYI im just working on training a distil model for this base to help with prompt adherence and video quality, ill let you know when I have something!
**July 22 2026: bf16, fp8 and int8 convrot versions for v1.6 are all under the same v1.6 dev model download, select the model from the dropdown
LTX 2.3 NSFW Checkpoint I2V/T2V
** Updated workflow and distil lora. The previous distil lora I was using caused audio and video issues and less fluid looking motion (artifacts etc). See below for the updated workflow (or use your own), use the corrected distil lora below please.
Model type: checkpoint
Training: full fine tune LTX 2.3 dev base
** July 25 2026 Distil Lora (better audio/motion): https://huggingface.co/Lightricks/LTX-2.3/blob/main/ltx-2.3-22b-distilled-lora-384-1.1.safetensors
Heretic Gemma3-12B: https://huggingface.co/DreamFast/gemma-3-12b-it-heretic-v2/tree/main/comfyui
**July 14 2026 Updated Workflow v1.5
(modified Rune workflow gives much better quality video/audio outputs and has a fixed distil lora than previous workflow):
Heretic Gemm3-12B is a uncensored version of the regular LTX 2.3 text encoder, much better for being naughty!
Official distil lora is used designed for the base model LTX 2.3
most LTX 2.3 based work flows work normally with these models, dont have to use mine version (I made changes to my workflow to improve quality though).
Supports: I2V, T2V, V2V
Workflow available at:
https://huggingface.co/SexGod1979/PinkCherry_NSFW_LTX23/tree/main
I would try the official LTX 2.3 384 distil lora first (try different distil loras to compare motion/audio):
https://huggingface.co/Lightricks/LTX-2.3/blob/main/ltx-2.3-22b-distilled-lora-384.safetensors
ill continue to work on this model and add more niche content and styles/actions (I am taking requests!). Will continue training this model with more style/actions, let me know what you want!
version 1.6 Training update;
The v1.5 training dataset is large, things I have tested so far:
blowjobs/pussy licking/oral sex/tongue
double blowjobs (two women)
POV style, high angle, low angle
cock/ball licking/sucking etc
doggystyle/cowgirl/reverse cowgirl/anal/missionary/spooning etc
double penetration/group sex
threesomes
lesbian/kissing/making out
breast touching/squeezing
hairy/shaved/trimmed/bushy pussy
spreading pussy, fingering
wet stuff
clothing related actions
handjobs/stroking
solo/mutual masturbation/rubbing
breast play
moaning/heavy breathing/sexy talk
Stuff I have not tested yet but the dataset contains, but may not have enough representation to swing the model towards those learned actions yet:
dildos (its probably ok but havent tested yet)
facials
ass licking
fisting
fingers in ass
pegging/strap on play
cum/creampies
bdsm
ass slapping
cuckold
strong orgasm response
theres a lot more stuff in the training data
Description
v1.6
FAQ
Comments (54)
C'mon pal, I just started to test your fabulous 1.5 checkpoint. Phew, so much stuff to download and test... you are just too fast. :o) - Good job, so far -
yeah I know, im sorry about that. I just keep wanting to mess around with it. its an addiction
A simple t2v workflow?
there are probably many T2V workflows that work fine, the Rune one I modified has a simple bypass toggle to enable T2V mode instead of I2V mode.
Have you considered making a model that comes with distilled fusion built-in? Using an external distilled LoRA is usually slower. Anyway thanks for your fantastic work!
Is there a recommended lora weight? You're really just using a weight of 1?
Thanks
what lora are you talking about?
@sexgod1979 distilled.. anyway I tried the model with distilled lora at 1 on both passes and it seems fine, I just wondered if it had the same issue as ltx eros.. This model seems to be way better.
Great job
@MrTitsworth oh I use the official LTX 2.3 lora (the big one) at 0.6 strength
@sexgod1979 can u provide the weight link from the huggingface for this one ''LTX 2.3 lora (the big one)'' thanks
@singularity_matrix I use 0.60 for the official distil lora 1.1
Absolutely amazing model. Any anus spreading or winking in the dataset to be trained on in future iterations?
its not well represented in my dataset, ill work on that for the next version!
@sexgod1979 Hey, I may aswell just ask here but might you add anal fingering and dildoing to your data set for the next version? Just some suggestions if you're open to them
Great work so far, I think this is better than Sulphur and 10eros tbh
@MrTitsworth yep I can work on that for the next version!
This works so well. LTX was giving me so much trouble lately with recent comfy updates, I almost didn't feel it was worth it. It's worth it. This model should give people hope.
wait until v1.7, should be amazing
The results are amazing, but with the ComfyUI backend I'm not getting results that match the examples shown for this checkpoint. I'm also getting poor generations overall, and the model often fails to generate decent videos when there are more than two subjects in the scene.
On average, only 6 out of every 10 videos turn out well. Could someone guide me on how to achieve results similar to the examples on this page? Are there any recommended workflows, settings, or best practices that I'm missing?
I think most of us see similar results in 1.6. The "trick" is having higher resolutions and upscaling. Many of my videos I include a "zooms in to capture face details" and that generally will get the resolution high enough to support the proper syncing. But yes it seems worse in 16. the creator says it should improve again.
@makiaeveli thanks gotta increase the resolution
@makiaeveli btw im new and implementing comfyui as a backend and deploying on the Modal if possible can i share the code with you? currently implementing text to video ?
@makiaeveli youll have to send me some prompts or show me some example videos so I can see, I dont think I had this issue with bf16 model with v1.6. makes me think its specific to fp8 or int8 (which I tested only a few videos to make sure it 'worked'), otherwise I only use bf16. Could be related not sure, again will need to see the prompt and resultant video/workflow. Could be distil lora youre using as well, so many possible variables. Havent heard of any real problems other than audio sync or encountered any myself
@singularity_matrix you can DM, sure, but it's not gonna be all that much more than setting up a real workflow and exposing the right variables via comfy-cli, right?
@sexgod1979 i am using int8. It's more like if a face is too far away when it speaks, the mouth and face won't move right. Compared to base bf16 LTX, where the faces will move even if they are quite small, using this model the face will be quite stiff or not move at all at distance. If I make sure to have the camera closer on the faces when they speak it works well. I'd imagine if I tried the bf16 version I'd get closer face movement. The size difference is just really nice, I crash so often running LTX workflows atm, especially with any loras.
@makiaeveli if you can hit me up on chat, and send me a file link to an example generated video that would be great because then I can see the prompt and the effect and reproduce it, as well as test it in v1.7 which im working on now to compare. That lets me determine where the flaw is
@sexgod1979 after downloading your bf16 model and messing around more -- the int8 and bf16 are giving similar outputs -- it's just my own workflow is bad. The cop example you uploaded really shows what I mean. Like his face barely moves as he talks. But in that higher resolution you can see his lips match. My other workflow is at a quite low res (like 600x600) and less steps (cause I used the Normalizing Sampler), and most of your examples are at least twice that. So using your workflow with increased resolution my outputs are looking much more like your examples. It's definitely my image preprocessing and low steps in my workflow lol. Im also downloading the qwen model to see if it improves any further. Maybe other ppl are using their own workflows too, and need the extra steps.
@makiaeveli I think 600x600 is way too low a res for LTX 2.3, its a high res video model. For instance I run my outputs at 1500x1000. And my training dataset is high resolution as well, there has to be enough "real estate" for LTX 2.3 to do what it needs to do to generate a video, if its low res its going to struggle. It does much better at high resolution. Normally the easiest way to test is take a good starting image in I2V, and prompt heavy with audio and motion.
@sexgod1979 attaching some of the gerneration im getting hardly any good generation(most probably skill issue from my side but it is also generating text, watermark, missed up hand and some time bad genetals); the examplke also include full HD full hd is also generating bad video im able to genrate good video but gte success rate is very low rn (all the attaced example si for text-to-video): for more context im using comfyui as a backend code and using distill lora 1.1 (big one) and this is the resolution
"16:9": (1280, 720) and example also contain full HD Example u eill able to tell which is one by watching the exampleshttps://storage.scintai.com/r2-ex/video-827.webm
https://storage.scintai.com/r2-ex/video-830.webm
https://storage.scintai.com/r2-ex/video-834.webm
https://storage.scintai.com/r2-ex/video-835.webm
https://storage.scintai.com/r2-ex/video-837.webm
https://storage.scintai.com/r2-ex/video-839.webm
https://storage.scintai.com/r2-ex/video-842.webm
https://storage.scintai.com/r2-ex/video-844.webm
https://storage.scintai.com/r2-ex/video-847.webm
https://storage.scintai.com/r2-ex/video-848.webm
@singularity_matrix have you tried image to video mode? pure text 2 video has always been a sort of weak spot with LTX 2.3, i2v mode gives much better results and most people dont bother with t2v mode, just my perspective. You end up wasting valuable generation time, instead try giving a good krea/flux starting frame and the results will be vastly different. for instance that hot police video where the woman is sucking the guy off in the gallery im pretty sure that started as a high resolution starting image and isnt text to video. Could also be the distil lora, as the official one doesnt know what nsfw concepts are really (ill have to create a distil lora, but there are nsfw distil loras out there to use). works fine for me in I2v mode, I always get great generations. Try different distil loras but honestly I dont bother with t2v because its always soft/plastic looking versus i2v I get crisp amazing videos. Most of the good videos others have posted are likely all I2V. for instance if you look at the breast massage video I posted thats an I2V from a generated close up of a woman's breasts, then I just get the model to animate it and it always works out great for me. not saying t2v cannot be decent with the right distil lora and workflows but not an expert on it, and all models I have tried with t2v in LTX 2.3 (even happy videos like dance videos with clothes on) to me look like crap, once I give it a strong starting image the game changes drastically because the model doesnt have to come up with a complex scene all by itself and then also animate it and apply crisp textures to everything. I just ran a text in v1.7 which I haven't released using a simple text to video prompt, and here is the result
for instance on your last video of the doggystyle if you gave it a high res starting image of that scene I bet you would find the quality and detail would go way up.
@sexgod1979 this t2v has simillar texture, audio, feel and quality as mine so i think you are right may be its t2v problem let me try image-to-video and i will let u know. Thanks for you help
@singularity_matrix let me put it this way. To generate a high res flux2-klein image you using a model of similiar param/file size, thats to generate one single image, no video. Asking LTX 2.3 to generate incredible t2v only videos is a huge ask, im not saying you cannot get good videos in t2v mode, but when you give the model all the high level details in an image the model only has to generate the parts and movements so it doesnt need nearly as much capacity. You can test this easily with the base ltx 2.3 model and something safe like 'A woman with a red dress dancing and singing a song', it looks good sure but never as good as if you gave it a starting image. So I always use starting images and I think most of us that have been using LTX 2.3 for thousands of generations do the same. Like the v1.7 test video I just posted of the woman giving two guys a blowjob with cum spilling on the floor just started as a single high res krea2 image, that I feed the video model and say "animate this thing". You can see how clear and crisp the results are because the model doesnt have to work nearly as hard trying to figure out an entire scene on its own
@sexgod1979 Just implemented the Image to Video and result is amazing still audio and i think lips syncing need more optimization nonetheless this is the current result https://storage.scintai.com/r2-ex/%23-0.6-20-3381811338-07af1f411a19fa76.mp4 2. https://storage.scintai.com/r2-ex/%23-0.6-20-1609556495-5f78787eb311db0e.mp4
@sexgod1979 so i tested it ; the quality is crazy good, the motion and the audio with lips syncing is quite good also; but most of the generation feel like 1.5x or 2x the speed i dont know why this is happening, tested with FPS 24, 25 and 30 ; would like to know your opinion on FPS.. in my opinion 30 is good for complex scenes i might be wrong; while generating the example it is possible that prompt is not fully matched with the first frame image bcs of this some example is not follwing the s*x scene properly; Examples file name format - {prompt-number}-{strength}-{fps}-{random-token}.mp4 and let me know your opinion; this is the drive link, drive contain both good and bad generations Thanks https://drive.google.com/drive/folders/1vI8XmEWyauhwUSgPtk9OJZfY3z0FXhc4?usp=sharing
@singularity_matrix my generations look different than yours, I use 24 fps and I trained the model at 24 fps. Keep it at 24 fps for testing for now. Is there a way I can see your workflow or are they in the videos? yes your videos seem like they are in 2x the speed (mine dont do that). if youre using comfyUI you can save one of my videos and load the workflow to compare.
@sexgod1979 yeah thats the problem im not using comfyui ; i took a comfyui source code and using there backend part/api only as my pipeline code ; and there is none to zero resource on the internet and it is causing lot of problem; but let me try to fix it and if possible can u share your workflow file for 1.7
I've really had issues with audio syncing up to movement..people talk without moving their lips an moaning being out of sync with movement. Is there a solution for that?
in version 1.6? that's likely on me, I lowered the audio weight training and I think it learned a weaker audio/video sync because of it. ill get that fixed in v1.7. I posted a few test videos of v1.7 for instance a woman getting her ass slapped and you can hear that audio syncs up correctly. Sorry about that, my fault.
Well Im glad I caught this reply. I just stopped the download dead in its tracks.
@randombrowser1234 audio is fixed in v1.7
beautiful model,great job,i have a question,this work on WAN2.2GP in pinokio ? Thanks for reply me.
I would assume it would because its just the standard LTX 2.3 model fine tuned, if LTX 2.3 normal model works this one would as well
@sexgod1979 I tried, but I'm getting a mismatch error and a whole lot of others.
@darkness1it with the bf16 model? which one? if you have the exact errors I can see what they mean
@sexgod1979 Your model is not appearing in the list of available wan2.2gp models, so it cannot be used.
@darkness1it I would have to look into how I get "my model" on wan2gp as I dont know
@sexgod1979 thanks i apreciate this
I have posted my v1.7 alpha model. I would appreciate any feedback or issues you find with this model so I know where I need to direct additional fine tune resources as I continue to train this model further. my goal is to figure out what the weak areas are so I can direct further training at those areas that need improvement
i will try myself maybe tomorrow(never did that) but how about merging a checkpoint with vbvr and DMD/condsafe lora?! DaSiWa merges his checkpoints so it works with no additional loras needed.
@hartweizen yes good idea. One possible hiccup is that this model being fine tuned is no longer the same as LTX 2.3 base in terms of the weights. distil lora are basically teacher student models where the student (distil lora) learns from the teacher (base model). Thats the issue with same the official distil light tricks lora (which I do use) is it doesnt actaully know this base so it can mess things up. THere are various other issues with the official distil lora as well as im sure you are aware. So the question is does the DaSiwa DMD lora/condsafe actually work with this base, does JoyAI work with this base? Honestly I havent tested them yet. So any feedback on this would be appreciated. Another way to deal with it is for me to train my own distil condsafe lora on this base. Good feedback though, any information or tests you run please do share in this regard
1.7 looking great, just posted a video.
Haven’t tested DMD on this model but haven’t had too great of results with others, but probably worth a look. Also check out the new JoyAi merge, tested with some nsfw Lora’s and had some interesting results, motion and sound were spot on, so might be worth a look into making it a Lora and maybe doing a merge. If I get some time I might convert it and see how it behaves.
@rogertimes90834 interesting, lots of good ideas here.
I did try DmD v2 from Dasiwa and vbvr v4+motion (sulphur base) and it works.i did try only few times.i will try tomorrow maybe.
Im using dasiwa workflow but i change few things (1stage instead of 3) because the distil lora only works on the first stage.
I will try tomorrow maybe merge both lora.i newer did that on my own.
Maybe i have to use the bf16 version.merge and convert to fp8.
@hartweizen cool, im working on my own distil lora for this base now to see it improves prompt adherence and quality