Raehoshi Anima

Raehoshi Anima is an enhanced iteration built upon the Anima Base v1.0 architecture. This release focuses on elevating visual style, integrating extensive new concepts, and expanding character knowledge. The ultimate goal is to deliver a more polished, balanced, and visually stunning output while remaining faithful to the core strengths of the base model.

Installation & Requirements

Important Note: This model does not include a built-in Text Encoder or VAE. You must download these components separately to achieve the intended results.

File Placement Guide

• raehoshi-anima-v1.0.safetensors goes in ComfyUI/models/diffusion_models

• qwen_3_06b_base.safetensors goes in ComfyUI/models/text_encoders

• qwen_image_vae.safetensors goes in ComfyUI/models/vae

Recommended Settings

For the optimal experience and the highest quality generations, we recommend the following configurations:

• Sampler: Euler a or ER SDE

• Schedule Type: Beta or Normal

• Steps: 32

• CFG Scale: 4.0 – 5.0

• Resolution: Any resolution up to 1536 (ensure dimensions are divisible by 32)

• Positive Prompt: masterpiece, best quality, score_7, absurdres

• Negative Prompt: worst quality, low quality, score_1, score_2, score_3, artist name, blurry, jpeg artifacts, bad anatomy, bad hands, bad proportions, mutation, deformed, extra digits, fewer digits, missing arms, missing legs

Prompting Tips

• Tag Ordering: For the most consistent results, follow this structured prompt order:

  [Quality / Meta / Year / Safety tags] ➔ [1girl / 1boy / Character Count] ➔ [Character Name] ➔ [Series / Copyright] ➔ [Artist] ➔ [General Tags]

• Character Accuracy: Always include the official series/copyright tags alongside the character name to significantly improve details and accuracy.

• Hybrid Prompting: The model handles hybrid prompting seamlessly. Feel free to mix dan match danbooru-style tags with natural language descriptions (e.g., use tags for characters and natural language for background/action).

Training Details

Raehoshi Anima was trained using a custom personal fork of Diffusion-pipe across a comprehensive two-stage fine-tuning process. The dataset utilizes multi-level captioning with random selection and tag dropout to ensure flexibility.

Stage 1: Concept & Character Expansion

• Dataset Size: v1-(25k images), v2-(21k images)

• Trained Resolution: 1024x1024

• Hardware: NVIDIA RTX PRO 6000 (96GB VRAM)

• Batch Size: 32

• Gradient accumulation steps: 1

• Learning Rate: 1.5e-6 (LLM Adapter LR: 2e-7)

• Focus: Introducing new franchises, series, and character knowledge.

Stage 2: Aesthetic & Style Refinement

• Dataset Size: ~1k high-curation images

• Trained Resolution: Multi-aspect (1024x1024 & 1536x1536)

• Hardware: NVIDIA RTX PRO 6000 (96GB VRAM)

• Batch Size: Per-resolution batch size (24-1536x1536) & (48-1024x1024)

• Gradient accumulation steps: 1

• Learning Rate: 1e-6 (LLM Adapter LR: 0)

• Focus: Mitigating artifacts, balancing composition, and enhancing the overall visual style.

List of New Series/Characters Trained:

Expanded Knowledge Base v1 (Up to May 2026) for this series

• Zenless Zone Zero

• Wuthering Waves

• Honkai: Star Rail

• Genshin Impact

• Arknights: Endfield

• Neverness to Everness

Expanded Knowledge Base v2 (Up to July 2026) for this series:

• Honkai Impact 3rd

• Goddess of Victory: Nikke

For character trait details prompts, please refer to the Danbooru site for accurate tags and references.

Special Thanks

A huge thank you to GSlinux for providing the development support needed to make this project a reality.

Support the Development

If you love using this model and want to help funds for future update, consider supporting the project:

• ⚡ Send a tip of Yellow Buzz

• ☕ Buy me a coffee

License

This model is released under the CircleStone Labs Non-Commercial License.