🧪 PINK ALCHEMY ANIMA🧪

THIS. IS. ALCHEMY! AND THE PARAMETERS ARE AT THE BOTTOM :3

YOU NEED THESE:

• https://huggingface.co/circlestone-labs/Anima/tree/main/split_files/vae

• https://huggingface.co/circlestone-labs/Anima/tree/main/split_files/text_encoders

REAL TALK!

Hello everyone! I just released 2.9 and I'm excited to share that after much testing we have arrived at a very stable space. We have achieved a strong base style with lower style drift and higher variance. While a simple merge, it was crafted very carefully, multiple times over, etc. Someday I may make an XY plot or something but all I can say for now is, just trust me bro? Have I steered you wrong thus-far?

I've had skilled artists like @freckledvixon, @haiti5, @CHESHIRE_OS, @Sophorium, @MLGnom, and a great deal of others test this and they confirm that v2.9 is the best iteration and works extremely well with their clothing and style loras alike.

Artist styles are indeed affected. I consider the effects of the checkpoint very pleasing as it works with rather than replaces the artist's style. They work better with the tag traditional media alongside the artist tag. E.g. @zerodiamonds \(voice actor\), traditional media

I'm extremely pleased with the results and I'm sure you will be too. ENJOY

LLM SLOP AND TIPS!

Pink Alchemy Anima Base

An anime/illustration checkpoint built on a real next-gen architecture — not another fried SDXL merge.

TL;DR

Pink Alchemy Anima Base runs on Anima-Base v1.0, a 2B Diffusion Transformer derived from NVIDIA Cosmos-Predict2 — a genuinely different lineage from the Pony/Illustrious/NoobAI SDXL world. That means real structural reasoning, a 16-channel VAE that actually keeps your detail, and an LLM-backed text encoder that understands prose and booru tags. If you've been waiting for an anime base that isn't held together with merge scar tissue, this is it.

Download, drop into ComfyUI, paste the settings below, and go.

⚙️ Why It Hits Different

This isn't marketing — it's the architecture, translated into what you'll actually see in your outputs:

• Flow matching, not epsilon/v-pred. It solves a clean noise→image trajectory instead of stacking SDXL-era noise residuals. Translation: far less of the burnt, over-contrasted "fried merge" look. Clean lines, neutral color, no mud.

• Cosmos-Predict2 DiT backbone with 3D RoPE. Real priors for depth, lighting, and object permanence — and it generalizes across resolutions instead of falling apart the moment you leave native.

• 16-channel Qwen-Image VAE (vs SDXL's 4). Four times the latent bandwidth, so faces and fine texture survive the encode/decode round trip instead of getting crushed.

• Qwen3-0.6B encoder + LLM adapter. Your prompt runs through a small language model, not a bare CLIP. Write actual sentences, hard-lock concepts with tags, or both — it follows the plot.

• Trained on real images. No synthetic slop. Millions of real anime images plus ~800k non-anime art. Anime knowledge runs current to September 2025, so recent characters and artists are in there.

✨ What This Checkpoint Adds

PRETTY STYLES THAT DON'T OVERWHELM THE THINGS THAT MAKE ANIMA SPECIAL, LIKE ARTIST STYLES OR LORA COMPATIBILITY!

🚀 Quick Start (Copy-Paste)

Drop the files in ComfyUI:

anima checkpoint  → ComfyUI/models/diffusion_models
qwen_3_06b_base.safetensors → ComfyUI/models/text_encoders
qwen_image_vae.safetensors  → ComfyUI/models/vae   (Qwen-Image VAE — you may already have it)

Recommended settings:

CFG:        4.0 – 5.0   (range 3.0–5.5; past ~6 it starts to burn)
Steps:      35 – 45     (clean as early as 24–32 with er_sde / res_multistep)
Sampler:    er_sde      (neutral, sharp, reliable default)
Scheduler:  Simple / Normal / beta(57)
Resolution: 1024×1024   (or ~1MP: 896×1152, 1152×896, 832×1216)
⚠ Do NOT use Karras schedulers. Generic DPM samplers are weak here.
This is a flow-matching model — those assume the wrong math.

Sampler quick-guide:

Sampler Look er_sde Neutral, flat color, sharp lines. The default. euler_a Softer, thinner lines, slightly hazy/2.5D. dpmpp_2m_sde More creative variety; can go wild on loose prompts (THIS IS MY PREFERRED). res_multistep Fast, clean, great er_sde alternative.

🗣️ Prompting (Copy-Paste)

The encoder is an LLM, so feed it like one — hybrid prompting beats a tag wall:

1. Natural language first. Describe the scene, lighting, and action in real sentences.

2. Booru tags second. Lock characters, series, and named concepts with Danbooru tags.

3. Syntax: tags lowercase, no underscores (except score tags), @artist_name to invoke a style.

Positive baseline:

masterpiece, best quality, score_9, score_8, score_7, newest,

Negative baseline:

worst quality, low quality, score_1, score_2, score_3, artist name, blurry, jpeg artifacts, lowres, censor, (bad quality:1.15), (worst quality:1.3)

Write detailed prompts. Sparse prompts are where any base model drifts — give it direction and it rewards you.

🎯 What It’s For (and What It Isn’t)

Straight talk, because you should know before you download:

• ✅ Anime, illustration, stylized and artistic work — the detail and prompt adherence are a real step up from SDXL finetunes.

• ❌ Photorealism — intentionally not the goal. Don’t expect it.

• ❌ Long text rendering — a word or four, maybe. Not sentences.

• ℹ️ Heavier than SDXL per image (it’s a dense DiT)

• ❌ Turbo lora sucks, don’t use it, it ALWAYS fucks up loras.

If you want stylized art that holds together without endless ADetailer/hires babysitting, you’re in the right place.

📜 License & Usage

Built on Anima (CircleStone Labs Non-Commercial License) — a derivative of NVIDIA Cosmos-Predict2 under the NVIDIA Open Model License.

The part that matters to you: the non-commercial restriction applies to the model weights, not to the images you generate. Your outputs are yours to use commercially — concept art, game/VN assets, client work, whatever. What’s gated is running the weights as a paid hosted service. Generate freely.

Drop a like, post your gens, tag your LoRAs. Show me what you build with it.