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