YOUR Penis (Retracted) 🍌 by Huggy

💡 Search "huggy" to find all my models on civitai.

💡 This lora only generates retracted foreskin. A phimosis version is planned later this year.

🍌 Get Involved in the Next Lora (👮 keep reading )

😫 It’s been 7 months since the 4.3 update (which was very overfitting). By now I should’ve trained a fixed version using the training methods I’ve learned since then, but current 200 samples I’ve got are not enough.

🙅 Random selfies from the internet don’t really work as training data — they’re more about vanity than natural anatomy. Proper training data should be shot in natural daylight, framed from below the navel to above the knees.

🙏 I’ve been asking bros on Civitai to donate photos, which makes me feel a bit crazy, but it’s the only way to build a proper dataset. Please DM @huggy on Civitai.

With 1 particular pubes type, a full set of images from a single person should include total of 160 images (16 states x 2 poses x 5 angles).

🍌 16 States (4 x 4)

• Shriveled (Saggy & Tight scrotum) (Pulled Back & Semi Covered foreskin)

• Natural Flaccid (Saggy & Tight scrotum) (Pulled Back & Semi Covered foreskin)

• Semi-erect (Saggy & Tight scrotum) (Pulled Back & Semi Covered foreskin)

• Erect (Saggy & Tight scrotum) (Pulled Back & Semi Covered foreskin)

🍌 2 Poses & 5 Angles

• Standing: POV (from above and avoiding feet), floor view (from bottom), front, profile

• Squatting: POV (from above and avoiding feet), front, profile, floor view (from bottom), back view (from behind)

🍌 4 Types of Pubic Hair

• Shaved

• Trimmed

• Natural

• Wet

🍌 Ethnicity

• African

• Latino

• Asian

• Caucasian

🍌 Submission

Chuck your photos into a zip file, upload it to one of these storage sites, and DM @huggy the link on Civitai:

• drive.google.com

• https://mega.io

• www.dropbox.com

---

🍌 Prompt Bleed:

Prompt bleed is something to be cautious of when writing captions for training data. A classic example is training a lora of a kiwi bird. Using the trigger "kiwi bird" seems natural, but the model's association with "kiwi fruit" is so strong it can totally stuff up the results, creating a bird-fruit hybrid. Switching to a unique trigger like "kiwi_bird" with an underscore solves the problem.

Another case is using "curtain wall" for a lofi lora, which will often just generate a wall with a curtain. A plain phrase like "a giant window from ceiling to floor" is a much safer bet.

It shows how hard it is to teach a base model a new concept with a lora without causing confusion, and some models get confused more easily than others. A good process is to test captions as prompts across different models first. If there's already prompt bleed without a lora, it's best to use a synonym. For instance, with the prompt "red lighting," many illustrative models will generate a character with red skin instead of a lighting condition—an issue that seems unique to the colour red.

This makes a project like training for a penis a massive challenge. You're dealing with 16 states (4x2x2), so it would be much easier to train each state as an independent lora with its own unique trigger word to avoid bleed. For the penis shaft alone, you've got 'Shriveled' (like after a cold shower), 'natural flaccid,' 'semi-erect,' and 'erect.' Then for the scrotum, it could be 'very saggy' or 'tight,' and the foreskin might be 'pulled back' or 'semi covered.' That’s where the 4x2x2 calculation comes from, giving at least 16 initial states to classify.

Then, there are also 4 types of pubes (shaved, trimmed, natural, wet) and the 4 major races, which require sub-classes. It would be a total disaster to repeatedly use the word "penis" in the captions. Multiple unique trigger words is likely the only workable solution.

---

🍌 Version 4.4 - Overdue:

1. To balance style fidelity and generalization, the following adjustments will be made:

• Use epochs to 26-81 to prevent excessive memorization.

• Lower Num repeats to 3.

• Shuffle Captions: Enabled.

• Reduce UNet LR to 0.00025 - 0.0003 for more stable training.

• Increase Text Encoder LR from 0.00005 to 0.000075 and 0.0001 to strengthen TE influence while keeping it balanced.

• Lower Min SNR Gamma (5 → 2-3) to reduce aggressive sharpening.

• Adjust Network Dim to 48, Network Alpha to 32 to avoid capturing too much detail.

• Enable Flip Augmentation for better generalization.

• Reduce Noise Offset to 0.06 - 0.08 to minimize artifacts.

• Switch LR Scheduler to "cosine" for smoother learning progression.

2. Increase training dataset to ??? photos, boosting versatility.

3. Use multiple trigger words prevent prompt bleeding.

4. Might split this lora into small penis and big penis versions for better control.

(Reference from this article and this one)

📒 SDXL

I would love to bring all my Loras to SDXL/Flux, especially the 'You Penis' Lora. However, it’s important to note that SDXL and Flux were not fine-tuned on data containing genitalia or detailed anatomical features, unlike models like Pony, which were fine-tuned using a dataset of over 10,000 full-body images with detailed anatomy. Training SDXL/Flux to generate accurate genital features would be much more challenging due to the lack of relevant training data, requiring a large, specialized dataset to ensure accuracy. This means I would need to prepare a new set of training files, including high-quality full-body images, to properly fine-tune the model.

• Learning Rate: 1e-5 (0.00001)

• Exact Value: 3000–5000 steps

• Exact Value: 4–8 images per batch

• LoRA Rank / Scaling: Rank 4 with a scaling factor of 0.8

• Image Resolution: 512×512

📒 Training dataset (for SDXL / Flux)

This range usually provides enough variation in poses, lighting, and contexts, ensuring that SDXL 1.0/Flux learns the nuances without overfitting.

• Close-Up Shots (50%):

• Medium Shots (30%):

• Full/Long Shots (20%):

📒 Flux Dev

• The noise scheduler settings are consistent with SDXL defaults.

• Gradient accumulation steps if your hardware is limited, but many users stick with 1–2 gradient accumulation steps.

• Optimizer: Often AdamW with weight decay of 0.01 is used.

• Learning Rate: 1e-5

• Steps: 3000 to 5000

• Batch Size: 4

• LoRA Rank: 4

• Scaling Factor: 0.8

📒 Issues with V4.3 and V4.2:

When training an SDXL LoRA for a pony style, the goal was to replicate a specific style/character from the training dataset. However, using Repeats 5, Epochs 40, Network Dim 64, Network Alpha 64 led to overfitting (V4.3), causing glitches when used with high-quality merged base models or other LoRAs. On the other hand, using Repeats 2, Epochs 10, Network Dim 64, Network Alpha 64 resulted in underfitting (V4.2), producing a "half-baked" style that lacked proper detail.

Unet block weights graph (V4.3E21) with uneven distribution indicates that some layers have much stronger activations than others, which means certain features are being exaggerated. The training process might be amplifying certain layers too much, leading to inconsistent outputs across different settings.

🍌 Version 4.3 - high fidelity.

V4.3E21 can generate realistic images with 5 times fidelity than V4.2 theoretically. New tagging method allows better control on penis size, pubes and balls. All LoRAs were trained with PONY based model and should only work with selected PONY based models, ignore the SDXL / illustrious label as I'm testing the compatibility. See below for 54-checkpoint reviews.

📒 Checkpoint recommendations:

🍌 Limitations of SDXL

• Checkpoint like Duchaiten Pony Real produces images with stunning details which is overwhelming when using additional lora, SDXL unable to effectively allocate sampling steps across different elements. 

• Unlike Flux, in SDXL, full-body generations struggle because the latent U-Net doesn’t allocate enough feature space early in the denoising process.

• Both Duchaiten Pony Real and Your Penis V4.3 are sharing limited denoising U-Net sampling steps during generation, resulting missing limbs, deformed face and mutated dick.

• This explains Your Penis V4.3 works flawlessly with PVC Style base model, as PVC material is easy to render.

• Duchaiten and Penis V4.3 are good on its own, for SDXL the denoising steps aren't being distributed evenly across complex images. This leads to artifacts that are amplified by highres fix. Offline inpainting could work though. Feels like I've hit pony's limit. Time for illustrious and flux.

📒 Important instructions:

• V4.3E21 (5 times fidelity): When working with other Loras (not more than 2), start with a lora weight of 0.5 and a low CFG scale 3.5 (a higher CFG scale results in sharper outputs).Best for inpainting.

• V4.3E41 (10 times fidelity): This ultra-fidelity version only works as a standalone. The output quality surpassed that of images from flux. Best for inpainting.

• V4.3E17 and V4.3E11: These are fine-tuned version, which should works with most Pony checkpoints with subtle effect.

📒 Two uses of this lora, create real life penis with triggering word "penis", anime style dominated checkpoint requires additional triggering word "realistic" .

📒 Enhance realness of a photo, especially on hairs without using triggering word.

• Use the Lora with epoch below 7 for a stable result, like V03E06, though lora with higher epoch like V03E10 has better resemblances to the training source; (this does not apply to V4 or later）

• USE fewer triggering word each time for best result;

• Use CFG between 4.5 (complex scene) to 7.5 (simple);

• For anime start training steps from a lower value like 19 with Euler a sampler, if you are satisfied with the result, use hi-res fix / face fix (and set Euler to 29) to increase over all quality;

• For realistic photos, use DPM++2M Karra, 30+ steps for close up portrait, 35+ steps for full body shot.

• There's no universal prompt / lora weight for every generation. If you're using anime style / character lora with realistic checkpoint, starting with a lower lora weight like 0.7. Ignore the lora weight in sample images, due to a recent bug, image generated online will display lora weight of 1 regardless the actual value.

• Step number setting (which is the weight for all tags) is closely associated with the complexity of the prompt, if your outcome has extra objects like fingers or limbs, try lower the step value (and the lora weight too); if your outcome has objects disappeared, lack of quality, texture and reality, try raising the step value.

🍌 Changes in Version 4.3 - high fidelity:

1. Use tag with underscore.

2. Trained with "slow baking" technique with proper num repeat and epoch ratio.

🍌 Changes in Version 4.2 - low fidelity:

1. New racial ratio: 7Afr | 3Lat | 87Cau | 71Asi

2. New state ratio: 60Flaccid | 38Erect | 66Semi-erect

🍌 Changes in Version 4.1:

1. Some of the source photos from V3 include limbs that needed to be cropped out or removed. They make V3 loose focus on penis, creating distortion and diphallia which never happened on V2.

2. Remove resources with weird foreskin

3. Reduce training resource quantity down to ???, as this is a common problem across most of my loras, too many training images reduce the outcome quality, the quality of the resource images is more important.

4. Increase training resource resolution.

5. Reduce training steps down to 1000, as I found lower training epochs lora got stable results.

📒 Mirroring

I’ve noticed a few sites, like Modelslab and yodayo, using bots to mirror my models. SeaArt in particular shows big numbers for downloads and generations.

1. I don’t mind if you mirror my models — sites like civitaiarchive have always done that.

2. I do mind if you block users from downloading the models and force them onto a paid online service.

3. I do mind if you claim copyright or give users ownership of the generations. These are open-source models, trained on open-source bases with limited-ownership data. No one — not even me — should claim copyright. Credit is all that’s needed.