V4-MEGA — The most complete version yet (Trained on 5000 images)

V4-MEGA is a major leap forward in every dimension. Built on the lessons of all previous versions, this release doubles down on consistency, precision, and versatility — with a dataset nearly double the size of V3-MEGA and the most rigorous curation to date.

If V3 was the transitional version, V4-MEGA is the definitive one.

✨ What's new compared to V3-MEGA

Significantly stronger visual identity across all character types
Much more stable anatomy and proportions, even on complex poses
Richer and more consistent color handling
Better generalization — performs well outside trained distributions
Smoother results at standard LoRA weights, no need to push hard
Reduced noise and artifacts even on tricky prompts

📦 Dataset

~5000 high-quality images
Carefully curated — no junk, no duplicates, no weak generations
Covers dragon, anthro, and furry content in depth
Manually refined captions throughout

Prompt used: (A powerful anthro dragon blacksmith with heavy muscular definition and iridescent obsidian scales. He is captured in a dynamic pose, wielding a heavy smithing hammer in his raised right hand, while his left hand firmly holds a molten, glowing runic sword steady on a massive anvil. Cascading embers and sparks illuminating his focused expression and sharp horns. Wearing a weathered, sweat-glistening leather apron. Industrial fantasy forge background with steam pipes and glowing furnaces. Dramatic chiaroscuro lighting, embers floating in the air, 8k resolution, cinematic composition, hyper-detailed textures)

(New Improved Version – Better than V1-MEGA)

This LoRA is focused on dragon, anthro and furry characters, trained with a large and carefully curated dataset and tested extensively during training.

This version is a clear improvement over V1-MEGA, with:

better consistency
stronger identity
cleaner shapes
more reliable results at normal LoRA weights

✨ What’s improved compared to V1-MEGA

Stronger visual identity at 0.8 – 1.0
Less overfitting / less noise
More stable anatomy and proportions
Better response to simple prompts
No need to push high LoRA weights

This version was validated during training with regular sample checks instead of blind overtraining.

📦 Dataset

~2600 high-quality images
Only images that already worked well in previous versions
Mixed dragon / anthro / furry content
Clean captions, no junk prompts

🔑 Trigger word

furry

V3 is a noticeable improvement over V2, with better consistency, improved color handling, and more flexibility when generating dragons and anthropomorphic variants.

That said, V3 is still a transitional version. While it performs well on most learned concepts, it can still struggle when pushed far outside its trained distributions.

If you are looking for the most polished and consistent experience, V4 is worth waiting for. V3 is recommended for testing, experimentation, and feedback.

LoRA trained on ~1200 images (V2)

This is V2 of my dragon LoRA, rebuilt from a cleaned and refined dataset.

Compared to V1, this version is significantly more stable, with:

much better color consistency
fewer broken generations
improved overall structure and anatomy

The dataset includes anthro dragon and furry concepts, properly separated during training.

Trigger word:
furry (works equally well for a dragon, wild dragon, furry anthro)

Recommended settings:

Lora weight: 0.9 – 1.0
Resolution: 768×768 (works up to 1024×1024)
Sampler: DPM++ 2M / SDE
Steps: 8–10
CFG: 1

Training details:

Dataset size: ~1200 images
Training time: ~8h30
Rank: 32

Captions were written manually and refined (no auto-tagging tools used).

V1:

Lora trained on 1160 images for V1

This is my first created of Lora.

This is V1, and this version is very unstable in terms of color.

The optimized resolution is (768x768 and 1024x1024).

Generation time: approximately 3 hours.

Rank 32

For image subtitles, the prompt was used. #I’m lazy to use Taggui with WD3

The word to activate Lora is: furry

Search AI Models & Addons

very few results