NAI-ANIMA 2d accelerated

An open Anima merge made to cut the glossy "AI sheen" typical of AI-generated images and push toward a flatter 2D look. Full recipe disclosed. Drive it like NovelAI: @artist tags + danbooru tags. Final output varies a lot with your prompting skill/techniques.

This is not a good-quality model — it's meant to share insight into LoRA training & model merging, and to encourage you to merge and post your own.

- (same seed) Paradoxically, the difference in output image style seen in 2dac also illustrates that, by the same measure, the styles the original model was capable of producing can be lost.

- Don't struggle to blend two or more artist tags — one LoRA, and one @artist tag. That's the best combination for maintaining quality. Since this model already has LoRA merged, If you want to add LoRA, please proceed in the Base Model, not this Model.

- On the off chance you do like this model, don't train on top of it. It's likely to produce a biased model.

🧪 Merge recipe

0.5·anima-base-v1.0 + 0.35·novaAnimeAM_v15 + 0.15·animayume_v05 → then bake the personal style LoRA:0.5

• Tool: Chattiori-Model-Merger · all bf16 · LoRA baked faithfully (no guard).

• The baked LoRA is itself a blend of two snapshots — the 2310-step and 1154-step checkpoints, each extracted at a partial weight and merged — then applied at 0.5. It's reflected as a full style with no trigger word (just prompt normally; the style is always on).

• LoRA Training spec:

  network_dim 32 · network_alpha 32 · network_train_unet_only true · module_dropout 0.05
  optimizer AdamW · lr 1e-4 · lr_scheduler constant · max_train_steps 2310 · batch_size 1
  resolution 1024x1024 · enable_bucket true · bucket_no_upscale true
  min_bucket_reso 512 · max_bucket_reso 1536 · bucket_reso_steps 64
  mixed_precision bf16 · save_precision bf16 · gradient_checkpointing true
  cache_latents true · cache_text_encoder_outputs true
  shuffle_caption false · keep_tokens 1 · caption_dropout_rate 0.1
  attn sdpa(torch) · images 233

• ✅ Dataset prep checklist

  • Generate with NAI Diffusion 4.5 in a style you like — keep only images with clean, well-formed hands & feet.

  • Make sure backgrounds are free of clutter/junk.

  • Apply proper tagging for training.

  • Spend paid Anlas to secure enough image buckets at ~1.8MP (Anima's max resolution).

  • Keep this in mind — Garbage in, Garbage out. You should be stricter about the quality of the AI-generated images you use for training.

🧭 New to training/merging?

Even if you don't fully understand LoRA training or model merging, paste this page into Claude or Codex and it can give you a solid step-by-step guideline. (This model was built with Claude Code + Opus 4.7, xhigh effort.)

LoRA Training tool: Trained with a small homemade Forge (Neo) extension I simply generate with claude — it's just a convenience wrapper around kohya sd-scripts (Anima fork), nothing custom in the math. So the spec above reproduces 1:1 with stock anima_train_network.py; you don't need my tool to redo this. Your agent will be make it easily.. (Repo: https://github.com/kohya-ss/sd-scripts.git), Pinned commit: 068bcd7 (full: 068bcd7ffe76b2cd5012fb680a2c94e295398bbc)

🎯 Recommended prompts & params

Quality tags (positive):

newest, year 2024, (best quality), score_8, highres, absurdres, {your prompts}, (perspective, foreshortening, dutch angle:0.75), (dynamic facial expressions), exaggerated and dark environment, violent composition, (low-contrast, muted color, watercolor \(medium\), highly aesthetic Pixiv style illustration, clean composition, view focus concentrated on the character with blurry background, high-quality digital art.:0.75)

Sample negative prompt:

ai-generated, 3d, (worst quality), low quality, (score_1), score_2, score_3, realistic, furry, furry female, anthro, unfinished, work-in-progress, absurdly detailed composition, blank, blank background, letterboxed, blurry, jpeg artifacts, mutated, mutated digits, missing fingers, extra digit, fewer digits, artistic error, unusual anatomy, watermark, patreon username, web address, patreon logo, weibo username, (artist logo, twitter username, signature), watermark, (multiple views), distorted anatomy, english text, anatomically incorrect, doodle on background, bad perspective, high contrast, cool colored, glitch, distortion, colorful, neon palette, detailed background, (vignetting, shiny skin, shaded face, face in shadow, underexposed face, underexposed body, dark body, body in shadow, low-key lighting, cast shadow, diagonal shadow, shadow across face, shadow across torso, harsh shadow, dramatic lighting, spotlight, rim light, split lighting, chiaroscuro:0.85), loli, aged down, child, chibi

Params: Well-Known Options — All

• take these values and reuse them as-is on Anima base v1.0. With prompting alone, a lot of the model's output changes.

• Another Quality+ Tags example: oekaki, flat color, low-contrast, muted color, watercolor \(medium\), traditional media, ...

📄 License

Non-commercial only. Inherits the CircleStone Labs Non-Commercial License + NVIDIA Open Model License (Anima is a derivative of NVIDIA Cosmos-Predict2-2B-Text2Image; all derivatives stay non-commercial). Commercial licensing: contact CircleStone Labs ([email protected]).

🤝 Credits & sharing

Ingredients:

• anima-base-v1.0 — CircleStone Labs × Comfy Org

• Nova Anime AM v1.5

• AnimaYume v0.5

Post your own merges & recipes — improve and share together.

TL;DR

Oh, we worship thee, Circle Stone bestow upon us the next version of ANIMA.