Sagging [Flux]

Dataset:

Greater variety, but lower number in images, better captions, unfortunately again fragmented into tags due to the incorrect use of the onsite trainer from Tensor.art.

Known issues:
The subject does not always know where the front and back are. As a result, the anatomy is sometimes considerably disturbed (bottom on the knees or stomach, etc.).

Due to the incorrect captioning/tagging, the people are sometimes generated too large, too small or with two buttocks.

The model has also incorrectly learned to generate doors, sockets or towel rails on the wall.

Stay tuned. v0.4 - v0.9 will be published within the next few hours.

Images: 30

Captions: natural language, destroyed by TA

Epoch 35 · Steps: 5250 · Loss: 0.228

Trigger words: -

Repeat: 10

Epoch: 35

Save Every N Epochs: 3

Clip Skip: -

Text Encoder learning rate: 0.00001

Unet learning rate: 0.0001

LR Scheduler: constant

Optimizer: AdamW8bit

Network Dim: 16

Network Alpha: 16

Gradient Accumulation Steps: 2

Noise Offset: 0.03

Multires noise Discount: 0.1

Multires noise iterations: 10

conv_dim: -

conv_alpha: -

Batch Size: -

Sampler: euler

Please post one of your created images here, in the model's gallery, using "+ Add post". Thank you :-D
New and improved in v1.0:

Realistic clothing: Highly detailed fabrics, seams, and folds.
Wide pose variety: Poses learned from a diverse dataset to ensure natural sagging and accurate fabric wrinkling.
Natural lighting: Consistent and lifelike illumination throughout.
Reduced overfitting: Thanks to improved and concise captions during training.
Flawless results: No plastic-like skin, broken nipples, or distorted faces.

Recommended settings:

Strength: 1.0 (Other values not extensively tested yet.)
Guidance: 2 for optimal results, but up to 3.5 is possible. Even at higher values, there are no issues with nipples, plastic-like skin, or faces.
Resolution: Resolutions of 768x1024 or higher are strongly recommended to avoid potential moiré effects.

Known issues: belts, belts everywhere, too many leather patches and Nike sneakers. New dataset in progress.

Triggerword optional, best combinations for prompts include:

sagging [details] [pants] (Enhance prompts with phrases like '(slightly) pulled down'.)
reveal(ing) [details] [underwear].
exposed/exposing [underwear].
[details] waistband/text/brand.

Examples:

"A photo of a young woman viewed from the front. She wears a black blouse and sagging black jeans, revealing white boxers with a green waistband displaying the yellow text 'I ❤️ SAGGING'."
"A man with dark-toned skin is standing with his left side facing the camera, wearing red sneakers, a light-blue hoodie, and sagging light-purple sweatpants, revealing black boxers."

For more Examples check the uploaded images or take a look at the captions (Dataset).

Targeted Sagging Control:
This LoRA allows you to direct the sagging effect to specific individuals or clothing items in your scene.

Example:
"The person on the right wears blue jeans and a black hoodie, while the other person is wearing sagging dark grey sweatpants and exposing bright green briefs."

Additional notes:

Combination potential: This LoRA integrates well with Briefs and Bulges - FLUX
Waistband customization:
- For custom text on waistbands, you can use emoticons like 😍🍆😈.
- To remove text or design, simply include "with a plain [optional:color] waistband" in your prompt.

Model Type	LORA
Base Model	Flux.1 D
Published	1/4/2025

Details

Download Files

About this version

Model description

Images made by this model