Bl4ckSpaces/Animagen-XL-V3

Animagen XL v3

Animagen XL v3 is a specialized, anime-focused SDXL model created by merging 9 distinct 2D and anime models.

Rather than using a standard equal-weight merge, this model utilizes block-specific U-Net routing and targeted Text Encoder adjustments. The goal is to provide stable anatomy, consistent flat-color anime styling, and reliable prompt adherence using standard Danbooru tags.

---

🏗️ Technical Architecture & Merge Details

This model addresses common issues found in large-scale merges (such as prompt syntax conflicts and anatomical degradation) through the following routing strategies:

• Text Encoder Isolation: The Text Encoder from Pony V6 has been intentionally excluded (set to 0% weight) to prevent syntax conflicts between Pony-style prompting (score_9, source_anime) and standard Danbooru tags. This model relies entirely on standard quality tags (e.g., masterpiece, best quality).
• Deep Block Routing (Anatomy & Poses): To maintain structural stability in full-body generations and complex poses, models with strong anatomical foundations (NoobAI, Illustrious Janku, and Pony V6) are assigned higher weights in the U-Net's Deep Blocks.
• Mid-Block Routing (Faces & Eyes): Illustrious Wai and Hassaku are given priority in the Middle Blocks to ensure consistent, detailed facial features and clean eye generation.
• Merge Optimizations:
  • L2 Norm Rescaling: Applied to the Text Encoder to maintain signal integrity and prompt responsiveness.
  • Epsilon Sweeping: Tensor values below 1e-5 are zeroed out during the merge to reduce static noise and artifacts.
  • Dynamic Range Clamping: Custom clamping is applied based on the base model's natural limits to prevent color burnout while maintaining proper saturation.

---

⚙️ Recommended Settings & Resolutions

Generation Parameters

• Sampler: DPM++ 2M Karras, Euler a, or Heun Karras.
• Steps: 25 - 40
• CFG Scale: 7.0 - 8.5 (Note: CFG below 7.0 may result in desaturated or grayscale images due to the multi-model blending).
• VAE: Baked in (Inherited from Illustrious Wai, stabilized).

Supported Resolutions (SDXL Standard)

For optimal composition and anatomy, please use the following standard SDXL resolutions:

• 1:1 (Square): 1024 x 1024
• 3:4 (Portrait): 896 x 1152
• 4:3 (Landscape): 1152 x 896
• 9:16 (Tall Portrait): 832 x 1216
• 16:9 (Widescreen): 1216 x 832
• 21:9 (Cinematic): 1536 x 640

---

🎨 Prompting Guide & Examples

This model responds best to standard Danbooru tagging formats.

Example 1: Close-up Portrait

masterpiece, best quality, ultra-detailed, 1girl, close-up portrait, glowing mesmerizing eyes, intricate pupils, detailed eyelashes, long flowing silver hair, soft cinematic lighting, depth of field, looking at viewer, highly detailed face, masterpiece illustration.

Example 2: Full Body & Armor

masterpiece, best quality, ultra highres, 1girl, full body, dynamic pose, standing confidently, intricate fantasy knight armor, holding a glowing greatsword, detailed hands, perfectly drawn fingers, outdoors, ruined ancient castle background, epic sky, sharp focus.

Example 3: Dynamic Action Angle

masterpiece, best quality, 1girl, dynamic angle, from below, dutch angle, fighting stance, mid-air, casting fire magic, glowing particle effects, flying debris, vibrant colors, cyberpunk street at night, neon lights, motion blur, highly detailed clothes.

Example 4: Soft Flat Color Landscape

masterpiece, best quality, 1girl, sitting on a grassy hill, looking at a breathtaking starry sky, galaxy background, shooting stars, beautiful landscape, calm night, soft flat color anime style, studio ghibli aesthetic, anime coloring, peaceful, magical atmosphere.

Standard Negative Prompt

To maintain a clean output, use this baseline negative prompt:

(worst quality, low quality:1.4), monochrome, grayscale, zombie, text, watermark, signature, ugly, bad anatomy, bad hands, missing fingers, extra digit, blurry, mutated hands, poorly drawn hands, disfigured.

---

🖼️ Result Preview

---

Merged by Bl4ckSpaces.