Anima Hyper Muscle & Penis Concept LoRA

# Anima Hyper Muscle & Penis Concept LoRA

A concept LoRA for Anima focused on furry hyper-muscle and hyper-genitalia.

Trained from scratch on curated e621 high-score data.

## What it does

- Amplifies furry male anatomy toward hyper muscle / hyper penis territory

- Works with kemono / furry / anthro subjects

- Best used as a standalone effect amplifier

## What it doesn't do

- Realistic / photo styles — Anima itself is illustration-focused

- Female / chibi / cub subjects — these were excluded from training

## ⚠️ Usage notes (please read)

- Standalone: works well, clear amplification at weight 1.0

- Stacked with style/character LoRAs: muscle and penis growth is mild,

and you may notice some style bleed from the training data.

→ Use weight 0.6 - 0.8 when stacking, then push tag weights higher instead.

- This is a v1 release. Feedback welcome.

## Recommended prompt

Positive:

masterpiece, best quality, score_8, explicit, kemono, furry,

male, anthro [species],

hyper, hyper muscles, hyper penis, hyper genitalia,

huge muscles, huge penis, huge balls,

veiny penis, vein, pecs, abs, biceps,

[your pose / scene tags]

lora:THIS_LORA:1.0

Negative:

worst quality, low quality, score_3, score_4,

(female, 1girl, pussy, breasts, wide hips:1.2),

ai-generated, ai-assisted,

muscular female, feminine,

comic, monochrome, sketch,

bad anatomy, extra digits

## Tags that work

Trained captions use plural forms (all from e621):

- ✅ hyper muscles (plural)

- ✅ hyper penis, hyper genitalia, hyper balls

- ✅ huge muscles, huge penis, huge balls

- ✅ veiny penis, vein

⚠️ Avoid these — they were removed during caption cleanup and may

confuse the model: muscular, muscular male, big muscles, big penis,

big balls, big pecs.

## Tested base model

- Anima base1.0 base

Base Model: Anima basev1.0

Type: Concept LoRA

Dataset Size: 329 images (positive samples only)

Source: e621, score-based filtered, manually curated

Network:

Rank (Dim): 32

Alpha: 32

Network Dropout: 0

Optimizer:

Type: AdamW

Learning Rate: 2e-5

Betas: 0.9, 0.99

Weight Decay: 0.01

Epsilon: 1e-8

Scheduler:

Type: Cosine

Warmup Steps: 100

Training Schedule:

Max Steps: 10000

Global Batch Size: 4

Save Every: 800 steps

Mixed Precision: bf16

Resolution:

Mode: Mixed-resolution

Resolutions: 512, 1024, 1536

AR Buckets: 9

Min/Max AR: 0.5 / 2.0

Anima-specific:

Freeze TE: Yes (llm_adapter_lr = 0)

Timestep Sample: Logit Normal

Sigmoid Scale: 1.3

Flow Shift: 3

Dataset Config:

Num Repeats: 1

Caption Prefix: (none, full caption in .txt)

Caption Extension: .txt

Caption Dropout: 0

Tag Dropout: 0

Shuffle Captions: No

Flip Augment: No

Regularization: None

Caption format:

- e621 native tags (spaces, not underscores)

- Artist names with @ prefix

- Rating tag at start (safe / sensitive / explicit)

- "kemono, furry" as fixed prefix

- Tag count limited to ~30 per image

- Low-frequency tags (<5 occurrences) removed

- Diluting synonyms removed (muscular*, big muscles, big penis, etc.)

- VLM-generated natural language description appended

## Dataset construction

Built with a custom e621 scraper:

- Source filter: score:>= adaptive ladder (200/150/100/50), favcount:>=100

- Sort: composite score (post.score + favcount), descending

- Quality gates:

- Manual exclusion: cub, gore, ai_generated, female, comic, sketch,

monochrome, traditional_media, 3d, photorealism, watermark, censored,

eyeliner, makeup, multiple_penises, animated

- Auto exclusion: B&W/lineart detection, comic-panel detection,

dHash-based duplicate removal

- Species balance: each common furry species (wolf/fox/dragon/tiger/

lion/equine/shark/deer/bear/canid/felid) has ≥10 samples

- Three sub-batches for tag balance:

- dual hyper (muscle + penis): 150

- hyper muscles only: 90

- hyper penis only: 80

- species top-up: ~9 additional

- Captioning:

- e621 native tags (rating + species + artist + characters + general)

- Qwen3-VL appends one natural-language morphology description per image

- Caption cleanup: deleted diluting synonyms, low-freq tags, capped at 30 tags

## Known limitations

1. Stacking weakness: when combined with other LoRAs (style/character),

the hyper amplification is reduced. Compensate by lowering this LoRA's

weight to 0.6-0.8 and increasing tag weights:

(hyper muscles:1.4), (hyper penis:1.4).

2. Style leak: noticeable e621 art style influence even at weight 1.0,

stronger at 1.5+. If you want pure form change without style shift,

keep weight at 0.7 or lower.

3. Plural-only tags: hyper muscle (singular) does NOT trigger this LoRA.

Always use hyper muscles (plural).

4. Weight ceiling: weights above 2.0 break the image (geometric distortion).

Sweet spot is 0.8-1.3 standalone.

5. Body symmetry: trained without flip augmentation, so consistent

left/right symmetry but slightly less pose flexibility.