PipelineP prvpns v9
Training corpus 18+
Technical documentation · 18+

prvpns
Krea2 training set

Creator-owned reference stills used to fine-tune the prvpns anatomical LoRA on Krea2. Below: how the set is organized, real example frames per class, and the captions used at train time.

106 training stills Trigger prvpns Base Krea2 Rank 16 · 2200 steps · done
Pipeline

Training pipeline

End-to-end path from curated stills to the shipped adapter — illustrated with real training images (not AI mockups).

01 Curate → 02 Caption → 03 Fine-tune → 04 Export

Details

What each stage does

Image on the left, explanation on the right — one row per pipeline stage.

Penetration · training progress

Mid-train samples for the dedicated contact prompt (sample index 3). Left = sample image, right = step metrics + what improves.

Ejaculation · training signal

No separate mid-train sample prompt was set for fluid — the class is learned from dataset stills mixed into the 2200-step loop. Cards show real train stills + what they teach.

Reference

Training stills (106)

Compact catalog for audit — full-res + caption on click.

Filter
Train

Fine-tune statistics

Run metrics from the completed Krea2 LoRA job (2200 steps).

Status
Completed
final checkpoint saved
Steps
2200 / 2200
~6.9 h wall time
Final loss
0.0832
min 0.0173 @ step 1633
Learning rate
1e-4
adafactor · bf16 · flowmatch

Hyperparameters

Productprvpns_penis_krea2_v9.safetensors
BaseKrea2 · Qwen-Image / flowmatch
NetworkLoRA rank 16 / alpha 16
Dataset106 captioned stills · 768×1024
Batch1 · gradient checkpointing
Triggerprvpns
Caption dropout0.05
Checkpointsevery 250 steps (kept: 250…2000 + final)

Dataset mix · with sample stills

Each row is a training class: count, share bar, and real stills from that class. Click a thumb to open full-res.

Loss curve

Window averages — start (1–100): 0.08765 · mid (1000–1100): 0.05919 · end (2100–2199): 0.0766

Train loss · every 25 steps 0.6630.4970.3310.1660.00005001000150020002200

Training progress · sample vs metrics

Mid-train samples (same prompt seed family) next to step / loss / progress bars — layout from the older exhibit style.

Class cards · still + stats

One real training still per class with share bars and extra samples beside the metrics.

Docs

Methodology

How to read this page as model-card documentation.

Purpose

LoRA adapter for stable anatomical identity under viewpoint and contact. Multi-state stills (erect / soft / contact / fluid) prevent pose lock-in.

Main token: prvpns.

How to read

  • Pipeline — compact process sketches.
  • Stills — the 106 fine-tune frames.
  • Click — full-res + training caption.
  • Filters — by anatomical class.

Classes

  • Erect / semi / flaccid — erection state.
  • Penetration — contact junction.
  • Ejaculation / toy / condom — controlled variants.

Access

Private technical documentation for the published model card. Do not redistribute stills. Creator-owned reference material only.

Train

Specifications

Productprvpns_penis_krea2_v9_training.safetensors
BaseKrea2 · flowmatch · bf16
NetworkLoRA rank 16 / alpha 16
Steps2200 · ai-toolkit · adafactor · lr 1e-4
Resolution768 × 1024
Triggerprvpns (identité)
SourceCreator-owned stills · 106 curated JPG
Typical weightcouples 0.45–0.55 · solo 0.55–0.65
Mixsolo identity + contact stills + flaccid · 0% IA showcase

Adult technical material (18+). Private consensual source photography. Reproducible fine-tune documentation — real corpus only.