Pony Pregnancy — Pregnancy & Postpartum Anatomy Model NSFW & SFW

A fine-tuned Pony checkpoint focused on pregnancy and postpartum abdominal morphology. It’s designed for clinical/realistic depiction (not stylized anime) with expressive tagging to control the shape and features of the stomach. In addition, postpartum conditions are captured for a consistent before, during and after transition, which includes wrinkling, overhang, diastasis, popped navels etc.

What's new in v1.0?

The latest release is much improved on overall visual fidelity and responsiveness to prompts. There are also a lot more images in the data set non-pregnancy related which helps to ensure concepts are not lost or overridden if they aren't well represented in available pregnancy imagery.

✅ Key Tags

Important

Phase-locked mandatory control with two anchors:

• phase_pregnant - use to generate pregnancy material

• phase_postpartum - use to generate postpartum material

Ethnicity

• chinese ,european - more refinement to come

Pregnancy

• Trimester control with:

  • first_trimester, second_trimester, third_trimester

    • Tip: control overall size with more refinement by weighting phase_pregnant between ~0.5 - 1.0

• Textural control with:

  • gigantic belly

  • stretch marks

  • flat navel

  • slit navel

  • deep navel

  • outie navel / huge outie navel

  • linea alba / hyperpigmentation

• Expanded Attire control with:

  • shirt overhang

  • dress overhang

Postpartum

• Postpartum taxonomy for nuanced outcomes:

  • pp_deep_creases - heavy wrinkling

  • pp_crepey_skin - subtle loose skin, noticeable affected skin texture

  • pp_panniculus- belly apron, skin that hangs and folds over the groin area

  • pp_hidden_navel- navel occluded by loose skin folding over

  • pp_frowning_navel - downturned/hooded look

  • pp_flat_navel- navel that didn't return to an innie and is flat along the stomach

  • pp_diastasis_recti - ab separation causing a soft squishy belly

  • pp_dr_mild/moderate/severe- severity of the ab separation

  • stretch marks - future improvement will differentiate between fresh stretch marks and faded out ones

📸 Recommended Settings

• Sampler: DPM++ 3M Karras / Euler a

  • Each has slightly different characteristics, use your personal preference

  • With Forge, I have found DPM++ 3M to give the most realistic appearance and skin texture, but does add a bit of fuzz like you'd get from amateur photography (as is most of the dataset out of my control). Euler a will give you sharper results, but slightly less realistic

• Steps: 25–35 (Euler a) · 28–40 (DPM++ family)

• CFG: 4.5–6.5

• Resolution: 1024×1024 (train res), or 832×1216 / 1216×832 if composing full height

• Hires fix / Upscale: optional 1.2–1.6× with a gentle denoise (0.2–0.35) to sharpen textures

Working with niche characters?

You can a LoRA of the character you want to generate, and if from an anime source, set to a CLIP strength of 0.5. You'll still capture their details decently, pairing with (big eyes) in negative prompt to stay grounded in reality.

🧩 Prompt Guidelines

Captioning rules:

• Prefix every pregnancy prompt with: phase_pregnant, trimester_X, X navel, ethnicity ...

• Prefix every postpartum prompt with: phase_postpartum, pp_X, pp_Y, ethnicity ...

Note: Existing pregnant tag is unnecessary.

Belly size:

The belly size is very flexible, but takes a little bit of tweaking to get it perfect. As mentioned above, you can control overall size with more refinement by weighting phase_pregnant between ~0.5 - 1.0 in the prompt. First trimester in some cases may still show a larger belly, but you can reduce it with the phase reduction.

As an experiment, I found that gigantic belly gives much larger pregnant bellies than what is currently in the dataset with realistic properties still retained. Try this for triplet + bellies.

Please also check out the metadata of the images provided above on this model.

Why not a LoRA?

I chose a full Pony fine-tune instead of a LoRA because the goal was precise, clinical anatomy without fighting the base model’s priors.

• Fidelity over flexibility. LoRA’s low-rank adapters are great for styles/characters & concepts, but they often miss fine-grained surface detail (wrinkling, creasing, skin pigmentation, navel nuances) unless you crank ranks/weights—and even then textures can “mush” under different seeds/resolutions with oversaturation. Full fine-tune lets the UNet itself learn these micro-features.

• Avoiding anime-like priors. Many base checkpoints implicitly learn idealized or stylized (often anime-skewed) pregnancy shapes from the internet. A LoRA tends to add a direction on top of that, so you end up battling the base with weights/negatives. Full fine-tune rewrites those priors so realistic pregnancy/postpartum anatomy is the default when you use the phase tokens.

• Trade-offs (accepted). Yes, a full model is larger and costlier to train than a LoRA, and it’s less “plug-and-play” with other adapters. But for this use case—accurate, phase-specific abdominal anatomy—the consistency and detail were worth it.