Upload 3 files
Browse files- JD3_qwen-edit_clothes.safetensors +3 -0
- JD3_qwen-edit_cum.safetensors +3 -0
- info.txt +73 -0
JD3_qwen-edit_clothes.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:37d608c8713da575ee77d44f17823c28a343854020129c783848b08c3ad534d4
|
| 3 |
+
size 590058840
|
JD3_qwen-edit_cum.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:2fc9473cb9113e9871c0b7eca1ff9175c82acc9282911984dfdd456851820bca
|
| 3 |
+
size 590058840
|
info.txt
ADDED
|
@@ -0,0 +1,73 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
JD3's Qwen-Image-Edit Clothes Remover + Cumshot LoRAs
|
| 2 |
+
|
| 3 |
+
Hello everybody! Today I have two new LoRA models for you. I'll talk about both models in more detail below but here are the downloads and quick prompt info if you want to immediately jump in and don't care. Example images are once again using my character-LoRA output.
|
| 4 |
+
|
| 5 |
+
|
| 6 |
+
Clothes Remover LoRA:
|
| 7 |
+
|
| 8 |
+
Download: https://drive.google.com/file/d/10ZAWLUWYp9fNmFfie2sb-FFZPno4rbol/view?usp=sharing
|
| 9 |
+
|
| 10 |
+
Example Image: https://prnt.sc/U_2n-AAiAgb0
|
| 11 |
+
|
| 12 |
+
Baseline Prompt: "Transform the image into an amateur nude photo. Completely remove the subject's clothes to reveal their breasts, nipples and clean shaven vagina. Keep the subject's facial features and hair style identical."
|
| 13 |
+
|
| 14 |
+
|
| 15 |
+
Cumshot LoRA:
|
| 16 |
+
|
| 17 |
+
Download: https://drive.google.com/file/d/1R1u_Ty2uhyx6EFBDDMzBxCtnZ_jc3MTF/view?usp=sharing
|
| 18 |
+
|
| 19 |
+
Example Image: https://prnt.sc/mmdi3UXQIj0o
|
| 20 |
+
|
| 21 |
+
Baseline Prompt: "Cumshot, a small load of semen is now covering the subject's face and clothes."
|
| 22 |
+
|
| 23 |
+
|
| 24 |
+
I used the basic "Qwen-Image-Edit" workflow template that comes with comfyui.
|
| 25 |
+
|
| 26 |
+
|
| 27 |
+
========================= MODEL INFO ==========================
|
| 28 |
+
|
| 29 |
+
|
| 30 |
+
Clothes Remover:
|
| 31 |
+
|
| 32 |
+
The dataset consists of 80 partially synthetic image pairs. I gathered a set of nude photos, ran them through Qwen-Image-Edit to add bikinis to cover NSFW body parts, and then used Gemini/Nano-Banana to fully cover their bodies with more complex outfits.
|
| 33 |
+
|
| 34 |
+
An example from the dataset: https://prnt.sc/mSHrS_uHoBHF
|
| 35 |
+
|
| 36 |
+
More images would likely make a big difference but I am very limited with the amount of GPU runtime hours I can use since my local rig can't handle training/inference and I'm too broke to afford more cloud compute lmao.
|
| 37 |
+
|
| 38 |
+
|
| 39 |
+
There are some adjustments you can make to the Baseline Prompt I provided above:
|
| 40 |
+
|
| 41 |
+
- Removing "Transform the image into an amateur nude photo." will make the model stick more closely to the original image's pose, though still not perfectly. This will also make it stick more closely to art-styles if using non-photography images.
|
| 42 |
+
|
| 43 |
+
- If you have an upper-body shot, remove the mention of vagina: "Completely remove the subject's clothes to reveal their breasts and nipples."
|
| 44 |
+
|
| 45 |
+
- If you want to influence the breast-size you can add big/medium-sized/small in front of "breasts" in the prompt. When omitted the model automatically adjusts if it's discernible from the image, adding it can however make breast shape more consistent.
|
| 46 |
+
|
| 47 |
+
- Replace "the" subject's with "all" subject's for multiple subject images.
|
| 48 |
+
|
| 49 |
+
|
| 50 |
+
I was not able to test parameters extremely thoroughly due to aforementioned brokeness but I got the most consistent results with LoRA weight 1.1, Euler sampler, Simple scheduler and cfg 4. There are surely better combinations so feel free to share if you experiment.
|
| 51 |
+
|
| 52 |
+
|
| 53 |
+
_______________________________________________________________
|
| 54 |
+
|
| 55 |
+
|
| 56 |
+
Cumshot:
|
| 57 |
+
|
| 58 |
+
The dataset has 151 image pairs, mostly photos but also including some different art-styles. Again this one is partially synthetic. I gathered a dataset and used Qwen-Image-Edit to remove any cum from the images, making sure no zooms occur, like the model normally likes to do with the template workflow.
|
| 59 |
+
|
| 60 |
+
You also have a few adjustments you can make to the Baseline Prompt:
|
| 61 |
+
|
| 62 |
+
- Remove "Cumshot" from the start of the prompt to make the model stick more closely to the cumshot volume defined afterwards.
|
| 63 |
+
|
| 64 |
+
- Replace "small" with "big" or "medium-sized" to control volume.
|
| 65 |
+
|
| 66 |
+
- Add other areas to cover: " ... is now covering the subject's face, hair and clothes."
|
| 67 |
+
|
| 68 |
+
- If you want cum to drip down from their face (sometimes this happens anyway, this makes it more prominent) add: " ... is now covering the subject's face and hair, dripping onto their clothes."
|
| 69 |
+
|
| 70 |
+
The default parameters of the template Qwen-Image-Edit workflow already work sufficiently, experimentation might still be worth it.
|
| 71 |
+
|
| 72 |
+
|
| 73 |
+
Enjoy!
|