@ragnaroook is right, it's actually just the non-tentacle subset of images from my main LoRA. This isn't just a LoRA for vaginas, it's a LoRA for making Qwen NSFW. Plus, if you've got Qwen running, you have to have a good sized disk. It takes so bloody long to train these, I'm not going to sacrifice accuracy in the name of disk space. Training the LoRA took up 315GB of disk space on my training rig.

@civitaiclimatic796289 Well, oddly, it shouldn't. So, a LoRA is basically a change made to a model, it doesn't make the model itself bigger. It essentially tells the base model "change this weight from 1.03 to 2.34, and this weight from 1.33 to 0.65".

BEFORE inference, the Low Rank matrix is expanded to the dimensions of the target parameter shape, and it permutes the parameters. So a larger rank LoRA will take more time to load the model, since there's more math to do. But once the model is loaded, it shouldn't take up more or less RAM.

But what I THINK is happening, is that this LoRA is a BF16 format. So when it's applied to smaller quants of the base model, like INT4, then it casts them up to FP8 or BF16, which uses twice or 4x as many bits. I don't see more VRAM usage on my machine when I use the base FP8 model (https://huggingface.co/Comfy-Org/Qwen-Image_ComfyUI/blob/main/split_files/diffusion_models/qwen_image_fp8_e4m3fn.safetensors). I haven't personally tried nunchaku, but when people complain about VRAM issues, they seem to mention that quant a lot.

Give me like a week, the training dataset on the main LoRA has around 10 000 images now, so to get through a single epoch it takes about 8h. I don't need as many epochs as other LoRAs, because of my huge dataset, but it needs a few to settle down and stop doing freaky things.

@ragnaroook I don't really mind my computer grinding away for 8h, I sleep most nights, so it can grind away while I'm resting. Qwen is crazy, it's actually too large to fit entirely on my GPU for training, so my CPU has to help with some of the layers.

All of the training data are images that I have personally generated. Which means that the main thing that consumes my compute time are actually generating images. It's much much easier now that I have LoRAs for each architecture I've loved, but initially, back in the days of SD1, there was no prompt for a yellow tentacle. So I had to hand-build the training data in Photoshop, but I wanted the entire dataset to be from my own art, not anyone else's. Obviously it's a fusion of my art and the base training data from the models I use to help, but when I'm making new training data (last week I was trying to teach it the concept of octopus-like tentacles with suckers) I have to keep Photoshop open, because new concepts that no model knows are not easy to generate with AI. I've got trained YOLO models for detecting faces, boobs, tentacles, feet, hands, and for finding when an arm stops being an arm and suddenly and inexplicably becomes a tentacle. I train a lot of LoRAs that never leave my computer, because they have like 10 training images, and they've been trained for like 3 epochs, so they can barely do the new concept.

BUT, on the flipside, generating new training data and sorting and labelling it involves looking at a lot of tentacle porn, so it's not soooooo awful, haha.

Sorry, I haven't figured out what's causing that. The source images I feed into the training look a lot more realistic. If you are using ComfyUI, you can take the output image, and pass it through PonyRealism (that's been my favorite semi-realistic to realistic converter) with just a denoise of like 0.1 or 0.2, and it will make it realistic. Hopefully in v2 I'll have that figured out.

@Omegaxx That's odd. A LoRA shouldn't take up VRAM. It permutes the source model, it doesn't add to it. I don't have this issue in ComfyUI. I'm seeing a recurring theme of people disliking the LoRA size, I'll look into reducing it.

@Omegaxx I just double-checked with ComfyUI locally, and the LoRA does increase my RAM usage, but not my VRAM at all. Generating an image without any LoRA, and with the LoRA takes just as long. The whole concept of a LoRA is to be able to fine-tune without fully fine-tuning. Maybe nunchaku is doing it strangely? I got a LoRA to consume VRAM, but only when I generated two images in a single workflow, with Illustrious, presumably because ComfyUI realizes that it has the space in VRAM to two copies of Illustrious, so rather than re-applying the LoRA every time the workflow executes, keeps two copies around. I'm just using all standard built-in nodes.

OH! I bet I know what's happening. This LoRA is intended to permute OG Qwen, and Nunchaku is a quantized model. I'm betting that since this LoRA is BF16, it's taking your FP4/INT4 or otherwise quantized model, and it's blowing it up to BF16, so like 4x-ing the RAM it needs. Does it still eat VRAM when you use a non-nunchaku model?

The OG Qwen has very traumatizing hallucinations of the nethers.

Sorry, I didn't have any crazy stuff in the training set. I want to keep it realistic, nothing cartoonish, but if you have some sample images as example that represent what you'd like it to do, let me know and I'll make a handful for the next version.

It took FOREVER. I eventually got it to work with Musubi tuner. The actual training only took like 16h, on my RTX 5090 with 32GB of VRAM, but GETTING IT TO WORK was the literal worst. I had to drop from Windows to WSL, I had to buy a Patreon sub to SECourses to get the files for his Qwen image training tutorial, and his settings absolutely obliterated my model's coherence, because his learning rate was far too strong. He was training on just 30 images, so he didn't have the kind of dataset I wanted to throw at it. I have around 1000 coochie images I used to train this, but I also needed to feed it non-sexual images and text images to keep it stable. Qwen is super picky. If you just give it 10 images, you'll make a completely inflexible LoRA that absolutely will recreate that dude's face, but if you want it to be smart, to be good, to understand what you MEAN, you need to follow the same training process the OG devs used, which involves having Qwen-VL describe your images, and Qwen-VL has no idea what sex is. You need long form QWEN SOURCED descriptions, you need to train at 1.8 megapixels, it's so finicky. Which is super sad, because the base model is so so capable. It also needs my PC to have NOTHING running, or the training goes from 3s/it to like 87s/it.

The key is that you need Qwen to make the image descriptions, and then you edit those long form descriptions. It can't handle tags like other models, it just goes crazy.

@Lady_Valeria Qwen is super picky about training labels, I've been struggling for so long to get it working, I eventually read the entire Qwen-Image paper, and they trained it with descriptions from Qwen-VL. The way I finally got it to work is with Qwen/Qwen3-VL-4B-Instruct, I give it each image in my dataset, and ask it to describe it, here is my current prompt below. But after Qwen has labelled them all, it gets everything the LoRA is for wrong, haha, because Qwen-VL hasn't been trained on what it means for a tentacle to be in a vagina, so I then go through every caption manually and correct each caption, and cut out the headers before training, it's pretty funny to see the output of Qwen-VL, since it uses analogies and euphemisms to describe things, presumably because the training data never got to see the word "nude" or "naked" or "vagina". It's like "a glossy tendril is near her lower torso and is excreting a thick white syrup". But Qwen-VL handles describing most things very nicely, like the background and her face/skin and most clothing:

```
Describe this image in these sections. Be succinct, each section should be a maximum of 30 words long. Avoid negative descriptions like "no shirt" or "without fear", and use positive descriptions like "topless" or "brave". TONE: The aesthetic tone and lighting (ex. moonlit ethereal misty forest ambiance) SETTING: The location the image is set in (ex. A brightly lit opulent hallway with clean marble floors) SHOT: Shot scale (ex. three-quarter shot, bust shot, dutch angle wide shot, or close up shot) POSE: The pose of the central subject, include arm and leg positions (ex. standing at attention, legs together, one arm in a military salute, the other holding a rifle) CLOTHING: The clothing, or lack thereof, of the central subject (ex. wearing a white crop top and blue pleated skirt with long black leggings and combat boots), if the subject is partially nude, describe that (ex. topless with green shorts). EXPRESSION: The facial expression of the central subject (ex. gently smiling) NATIONALITY: A one-word nationality for the central subject (ex. Canadian), always pick a nation, not a general class. DETAILS: Without repeating earlier descriptions, include any other important information here. This section should be at least 40 words long. (ex. She is surrounded by thick blue tentacles that are wrapping around her legs and torso. One tentacle is in her vagina. The cave walls are wet and slimy. A small stream of muddy water flows from the background, pooling at her feet.) SUBJECT: Without repeating earlier descriptions, include a full description of the central subject. (ex. A cute Neko catgirl with 4ft long red hair and pale skin with vibrant red lipstick. She is filthy from playing in the mud.).
```

I'm hand-labelling training data tonight, and I generate all the training data myself, and label it myself, and I learned early along, that if an image is "almost good enough" then it is NOT GOOD ENOUGH DON'T PUT IT IN YOUR TRAINING DATA. And I was just labelling her pose, since I'm trying to teach the model the word "clitoris", and my label is:

"Legs parted, she masturbates with one hand, fingers adorned with a silver diamond ring. Her middle finger is rubbing her clitoris. Her pointer finger and ring finger rest on her labia. Her pinky curls away with pleasure. Her thumb is above on her thigh and FUCK THE THUMB IS ON THE WRONG SIDE"

This image sadly needs to be discarded.

It should have basic A capabilities. It's just the non-tentacle data from the main LoRA, it's not actually specifically a vagina dataset. I'm not just personally an A man myself, and I hand-label all my training data, so I just don't feel like hand-labelling an A-specific dataset. But there are plenty of training images with A.