V1.0

This version uses ZImage Base as the base model. It was trained on a larger dataset and has a much higher success rate. Two optional trigger words have been added: Bulg3_TF and Bulg3_LF.

These trigger words help the LoRA understand how to render bulges on clothing types that were not present in the training dataset.

Use Bulg3_TF for fabrics that tightly wrap around the penis and testicles (tight jeans, underwear, etc.), and Bulg3_LF for fabrics that loosely wrap around them (such as regular pants, loose shorts, sweatpants, etc.).

All preview images in this post were created in ComfyUI using the standard ZImage Base workflow from the Templates.

Settings: CFG: 4, steps: 25, sampler: res_multistep, scheduler: beta.

The initial prompt was enhanced using the official Prompt Enhancer from the ZImage developers: link

Example of initial prompt:

A man with a visible erection bulge in his shorts. The bulge is directed to the left and clearly defined.

A medium shot of a lifeguard sitting on a lifeguard tower, looking through binoculars. He is wearing short yellow shorts. In the crotch area, the outline of an erection bulge is clearly visible. The contours of two rounded testicles and the ridge of the shaft are strongly defined beneath the stretched fabric.

The lighting is bright and natural, emphasizing the contours of the bulge and the man’s physique. The background is a beach. The mood of the photo is erotic and intimate. The image appears to be taken with a professional camera.

v0.9

Note:
This release is labeled v0.9 because the LoRA can still be somewhat stubborn when generating bulges under certain parameter combinations, and there is clearly room for further improvement.

What’s New

1. No custom trigger word required
   The LoRA no longer relies on the custom trigger word 3r3ctC0ckBulg3. It can now be activated simply by using the word “bulge” in the prompt.
   That said, prompts using “erection bulge” tend to produce more consistent and stronger results.

2. Recommended prompt preamble
   While the LoRA can work without it, using a structured preamble is strongly recommended:

   A [man/woman] with an erection bulge in [type of clothing: pants, underwear, wrestling singlet, shorts, etc.]. The bulge is directed [upward, to the left/right, downward, downward along the left/right thigh]. The bulge is [slightly, clearly, strongly, etc.] visible. [YOUR REGULAR PROMPT]

   This helps the model lock onto anatomy, direction, and visibility early in the generation.

3. Wider clothing support
   This version works with a broader variety of clothing compared to earlier releases.
   Not every clothing type has been tested yet—if you encounter a specific clothing type that consistently fails to produce a bulge, please mention it in the comments so I can include examples in the dataset for the upcoming v1 release.

4. Improved direction handling (with natural limits)
   The LoRA now has a more flexible understanding of bulge direction.
   However, each clothing type still has natural limitations. For example:

   • Underwear most naturally supports left or right directions.

   • Upward bulges in underwear are still possible, but usually require very detailed prompting and may only produce okay results at best.

5. Better performance on black fabrics
   Compared to v0.5, this version has a significantly improved understanding of black clothing, allowing for much clearer and more reliable bulge definition under dark fabrics.

6. Experimental support for women
   The LoRA can now also generate bulges for women, although this functionality has not yet been thoroughly tested and should be considered experimental.

7. Important: Most of the training dataset consists of close-up images of bulges. As a result, the LoRA tends to generate cropped compositions by default, often excluding the face. To obtain full-body shots, the face (or facial features) should be explicitly described in the prompt.

What Remains the Same

Providing detailed anatomical descriptions can still noticeably improve results, for example:

“A massive, anatomically detailed bulge is clearly visible — the outline of large, rounded testicles and a thick shaft pressed against the underwear, with the lower curve of the testicles and the ridge of the shaft distinctly highlighted.”

Describing lighting interaction continues to be beneficial:

“Sunlight casts sharp highlights and shadows, emphasizing the bulge’s contours, especially the lower curve of the testicles and the shape of the shaft.”

v0.5

This is an alpha version of a LoRA I trained to experiment with LoRA training on Z-Image. It was trained on ~40 images of varying quality, so do not expect excellent generations, but it can produce decent bulges in most cases.

The trigger word is 3r3ctC0ckBulg3, but it can also work without the trigger. You just need to describe the bulge in the prompt, for example:

“A massive, anatomically detailed bulge is clearly visible.”

Sometimes, providing a more detailed description helps, such as:

“A massive, anatomically detailed bulge is clearly visible — the outline of large, rounded testicles and a thick shaft pressed against the underwear, with the lower curve of the testicles and the ridge of the shaft distinctly highlighted.”

It can also help to describe how light interacts with the bulge, for example:

“Sunlight casts sharp highlights and shadows, emphasizing the bulge’s contours, especially the lower curve of the testicles and the shape of the shaft.”
Recommended parameters:

• Sampler: res_2s

• Scheduler: beta57

Known issues:

1. It can only generate bulges for underwear and singlets (see example images).

2. It struggles to add bulges to completely black clothing (likely because it relies on highlights and shadows, which are harder to produce on black fabrics).

3. It may reduce overall image quality; in that case, use a LoRA weight of 0.7–0.9.

4. It appears to generate only cis men with bulges and does not work for trans women.

5. In some cases, the model cannot decide which side the penis is tucked, resulting in a double-bulge / double-shaft artifact. In such cases, explicitly specifying the direction of the curve (e.g., curved to the left or curved to the right) usually fixes the issue (see the image with the man in a singlet).

I hope to find some time in the future to improve the training dataset and address the issues listed above.