AniSee V2: model card, config and galleries for Turbo / Aesthetic / Base

Documents all four checkpoints, adds the three-way comparison gallery and per-variant sample galleries, and removes a stray Thumbs.db.

.gitattributes

@@ -49,3 +49,71 @@ images/9.png filter=lfs diff=lfs merge=lfs -text
 images/anisee-workflow-cover.png filter=lfs diff=lfs merge=lfs -text
 images/cover.png filter=lfs diff=lfs merge=lfs -text
 images/Thumbs.db filter=lfs diff=lfs merge=lfs -text
+images/v2/aesthetic/g01.webp filter=lfs diff=lfs merge=lfs -text
+images/v2/aesthetic/g03.webp filter=lfs diff=lfs merge=lfs -text
+images/v2/aesthetic/g06.webp filter=lfs diff=lfs merge=lfs -text
+images/v2/aesthetic/g07.webp filter=lfs diff=lfs merge=lfs -text
+images/v2/aesthetic/g08.webp filter=lfs diff=lfs merge=lfs -text
+images/v2/aesthetic/g09.webp filter=lfs diff=lfs merge=lfs -text
+images/v2/aesthetic/g10.webp filter=lfs diff=lfs merge=lfs -text
+images/v2/aesthetic/g12.webp filter=lfs diff=lfs merge=lfs -text
+images/v2/aesthetic/g17.webp filter=lfs diff=lfs merge=lfs -text
+images/v2/aesthetic/g20.webp filter=lfs diff=lfs merge=lfs -text
+images/v2/aesthetic/g21.webp filter=lfs diff=lfs merge=lfs -text
+images/v2/aesthetic/g22.webp filter=lfs diff=lfs merge=lfs -text
+images/v2/aesthetic/g23.webp filter=lfs diff=lfs merge=lfs -text
+images/v2/aesthetic/g24.webp filter=lfs diff=lfs merge=lfs -text
+images/v2/aesthetic/g25.webp filter=lfs diff=lfs merge=lfs -text
+images/v2/aesthetic/g26.webp filter=lfs diff=lfs merge=lfs -text
+images/v2/aesthetic/g27.webp filter=lfs diff=lfs merge=lfs -text
+images/v2/aesthetic/g28.webp filter=lfs diff=lfs merge=lfs -text
+images/v2/aesthetic/g29.webp filter=lfs diff=lfs merge=lfs -text
+images/v2/aesthetic/g30.webp filter=lfs diff=lfs merge=lfs -text
+images/v2/base/g01.webp filter=lfs diff=lfs merge=lfs -text
+images/v2/base/g06.webp filter=lfs diff=lfs merge=lfs -text
+images/v2/base/g07.webp filter=lfs diff=lfs merge=lfs -text
+images/v2/base/g09.webp filter=lfs diff=lfs merge=lfs -text
+images/v2/base/g10.webp filter=lfs diff=lfs merge=lfs -text
+images/v2/base/g12.webp filter=lfs diff=lfs merge=lfs -text
+images/v2/base/g17.webp filter=lfs diff=lfs merge=lfs -text
+images/v2/base/g20.webp filter=lfs diff=lfs merge=lfs -text
+images/v2/base/g21.webp filter=lfs diff=lfs merge=lfs -text
+images/v2/base/g22.webp filter=lfs diff=lfs merge=lfs -text
+images/v2/base/g23.webp filter=lfs diff=lfs merge=lfs -text
+images/v2/base/g24.webp filter=lfs diff=lfs merge=lfs -text
+images/v2/base/g25.webp filter=lfs diff=lfs merge=lfs -text
+images/v2/base/g26.webp filter=lfs diff=lfs merge=lfs -text
+images/v2/base/g27.webp filter=lfs diff=lfs merge=lfs -text
+images/v2/base/g28.webp filter=lfs diff=lfs merge=lfs -text
+images/v2/base/g29.webp filter=lfs diff=lfs merge=lfs -text
+images/v2/base/g30.webp filter=lfs diff=lfs merge=lfs -text
+images/v2/comparison/01-seesee-elf-aerial-silk.webp filter=lfs diff=lfs merge=lfs -text
+images/v2/comparison/02-seesee-elf-library.webp filter=lfs diff=lfs merge=lfs -text
+images/v2/comparison/03-seesee-kitsune-shrine.webp filter=lfs diff=lfs merge=lfs -text
+images/v2/comparison/04-pixel-art-jungle.webp filter=lfs diff=lfs merge=lfs -text
+images/v2/comparison/05-dragon-woman-beach.webp filter=lfs diff=lfs merge=lfs -text
+images/v2/comparison/06-painterly-lighthouse.webp filter=lfs diff=lfs merge=lfs -text
+images/v2/comparison/07-rooftop-sunset.webp filter=lfs diff=lfs merge=lfs -text
+images/v2/comparison/08-tags-elf.webp filter=lfs diff=lfs merge=lfs -text
+images/v2/comparison/09-tags-fox-girl.webp filter=lfs diff=lfs merge=lfs -text
+images/v2/comparison/10-tags-chibi-magical-girl.webp filter=lfs diff=lfs merge=lfs -text
+images/v2/turbo/g01.webp filter=lfs diff=lfs merge=lfs -text
+images/v2/turbo/g03.webp filter=lfs diff=lfs merge=lfs -text
+images/v2/turbo/g06.webp filter=lfs diff=lfs merge=lfs -text
+images/v2/turbo/g07.webp filter=lfs diff=lfs merge=lfs -text
+images/v2/turbo/g08.webp filter=lfs diff=lfs merge=lfs -text
+images/v2/turbo/g09.webp filter=lfs diff=lfs merge=lfs -text
+images/v2/turbo/g10.webp filter=lfs diff=lfs merge=lfs -text
+images/v2/turbo/g12.webp filter=lfs diff=lfs merge=lfs -text
+images/v2/turbo/g17.webp filter=lfs diff=lfs merge=lfs -text
+images/v2/turbo/g20.webp filter=lfs diff=lfs merge=lfs -text
+images/v2/turbo/g21.webp filter=lfs diff=lfs merge=lfs -text
+images/v2/turbo/g22.webp filter=lfs diff=lfs merge=lfs -text
+images/v2/turbo/g23.webp filter=lfs diff=lfs merge=lfs -text
+images/v2/turbo/g24.webp filter=lfs diff=lfs merge=lfs -text
+images/v2/turbo/g25.webp filter=lfs diff=lfs merge=lfs -text
+images/v2/turbo/g26.webp filter=lfs diff=lfs merge=lfs -text
+images/v2/turbo/g27.webp filter=lfs diff=lfs merge=lfs -text
+images/v2/turbo/g28.webp filter=lfs diff=lfs merge=lfs -text
+images/v2/turbo/g29.webp filter=lfs diff=lfs merge=lfs -text
+images/v2/turbo/g30.webp filter=lfs diff=lfs merge=lfs -text

README.md

@@ -15,183 +15,208 @@ tags:
 - diffusion
 - comfyui
 - fine-tune
+- lokr
 - safetensors
 - anisee
 ---
 
-<div align="center">
-
 # 🎨 AniSee
 
-### Personal Anime Fine-Tune of Anima Preview3 Base
+**A personal anime model built on Anima.**
+Two generations · Four checkpoints · Tags + natural language · ComfyUI-ready · LoRA-friendly
 
-**Full Fine-Tune • Clean Anime Aesthetics • Tag + Natural Language • Anima-Compatible**
+AniSee started as a full fine-tune of *Anima Preview3 Base* and has since moved onto the final
+*Anima v1.0 / v1.1* releases. Both generations live in this repo, because they are genuinely
+different things and some people will prefer the older one.
 
-**Diffusion Model • 1 MP Native • LoRA-friendly**
+Everything you need is here — the models, the Qwen text encoder and the Qwen-Image VAE.
+No hunting across repos.
 
-<br>
+---
 
-<a href="https://civitai.red/models/2628747/anisee">
-  <img src="https://img.shields.io/badge/CivitAI-AniSee-EC4899?style=for-the-badge&logoColor=white" alt="CivitAI">
-</a>
-<a href="https://anisee.anisee.workers.dev/">
-  <img src="https://img.shields.io/badge/🎨_Sample_Gallery-Explore-D63384?style=for-the-badge" alt="Sample Gallery">
-</a>
-<a href="https://huggingface.co/circlestone-labs/Anima">
-  <img src="https://img.shields.io/badge/Base_Model-Anima_Preview3-FFD21E?style=for-the-badge&logo=huggingface&logoColor=black" alt="Base Model: Anima">
-</a>
-<a href="#-license">
-  <img src="https://img.shields.io/badge/License-Non--Commercial-A02060?style=for-the-badge" alt="License">
-</a>
+## ⬇️ Which file do I download?
 
-<br><br>
+| File | Built on | Steps / CFG | Pick it if |
+|---|---|---|---|
+| **`AniSee-V2-Turbo.safetensors`** ⭐ | anima-turbo-v1.1 | 12 / 1 | You want speed. ~6 s per image. **Start here.** |
+| **`AniSee-V2-Aesthetic.safetensors`** | anima-aesthetic-v1.1 | 40 / 4–5 | You want the best look out of the box |
+| **`AniSee-V2-Base.safetensors`** | anima-base-v1.0 | 40 / 4–5 | You want maximum variety, or you train LoRAs |
+| `anisee.safetensors` | Anima Preview3 Base | 40 / 4.5 | You are already on the classic Preview3 setup |
 
-<img src="https://huggingface.co/SeeSee21/AniSee/resolve/main/images/cover.png" alt="AniSee Cover" width="100%">
+All four are **diffusion model files**. You also need the text encoder and the VAE from this
+repo — both are included below.
 
-</div>
+> An **all-in-one** build that packs model, text encoder and VAE into a single checkpoint
+> exists for V1 on [CivitAI](https://civitai.com/models/2628747/anisee). A V2 all-in-one is
+> planned.
 
 ---
 
-## 🖼️ Preview Gallery
+## 🖼️ Preview — the three V2 variants, same seed
 
-Browse the full curated set of sample images on the dedicated gallery page:
+Each image below is one prompt rendered by all three V2 variants with the **seed locked
+identical**, so the only variable is the checkpoint. Order is **Base · Turbo · Aesthetic**,
+left to right.
 
-<div align="center">
-  <a href="https://anisee.anisee.workers.dev/">
-    <img src="https://img.shields.io/badge/🎨_Open_Sample_Gallery-anisee.anisee.workers.dev-D63384?style=for-the-badge" alt="Sample Gallery">
-  </a>
-</div>
+![seesee_elf — aerial silk](images/v2/comparison/01-seesee-elf-aerial-silk.webp)
 
-|  |  |  |
-| :---: | :---: | :---: |
-| ![AniSee preview 1](https://huggingface.co/SeeSee21/AniSee/resolve/main/images/1.png)  | ![AniSee preview 2](https://huggingface.co/SeeSee21/AniSee/resolve/main/images/2.png)  | ![AniSee preview 3](https://huggingface.co/SeeSee21/AniSee/resolve/main/images/3.png)  |
-| ![AniSee preview 4](https://huggingface.co/SeeSee21/AniSee/resolve/main/images/4.png)  | ![AniSee preview 5](https://huggingface.co/SeeSee21/AniSee/resolve/main/images/5.png)  | ![AniSee preview 6](https://huggingface.co/SeeSee21/AniSee/resolve/main/images/6.webp) |
-| ![AniSee preview 7](https://huggingface.co/SeeSee21/AniSee/resolve/main/images/7.webp) | ![AniSee preview 8](https://huggingface.co/SeeSee21/AniSee/resolve/main/images/8.png)  | ![AniSee preview 9](https://huggingface.co/SeeSee21/AniSee/resolve/main/images/9.png)  |
-| ![AniSee preview 10](https://huggingface.co/SeeSee21/AniSee/resolve/main/images/10.png) | ![AniSee preview 11](https://huggingface.co/SeeSee21/AniSee/resolve/main/images/11.webp) | ![AniSee preview 12](https://huggingface.co/SeeSee21/AniSee/resolve/main/images/12.webp) |
-
----
+![seesee_kitsune — shrine](images/v2/comparison/03-seesee-kitsune-shrine.webp)
 
-## ✨ What is AniSee?
+![Dragon woman — beach](images/v2/comparison/05-dragon-woman-beach.webp)
 
-**AniSee** is a personal full fine-tune of CircleStone Labs' **Anima Preview3 Base**, retrained on my own curated dataset to push the model further into a cleaner, more focused anime aesthetic.
+![Pure tag prompt — fox girl](images/v2/comparison/09-tags-fox-girl.webp)
 
-It is **not a LoRA merge** — AniSee is a full fine-tune (~20K steps) with the **LLM adapter only very lightly co-trained**, following the official Anima fine-tuning guidelines. The goal is to keep everything that makes Anima a strong illustration base — Danbooru tags, natural language prompts, mixed prompts, full Qwen text encoder, Qwen-Image VAE — while shifting the default style toward a stronger anime look in line with my other checkpoints.
+![Pixel art — jungle](images/v2/comparison/04-pixel-art-jungle.webp)
 
-AniSee is mainly intended for:
+The remaining comparisons are in [`images/v2/comparison/`](images/v2/comparison), and each
+variant has its own 20-image gallery in [`images/v2/turbo/`](images/v2/turbo),
+[`images/v2/aesthetic/`](images/v2/aesthetic) and [`images/v2/base/`](images/v2/base).
 
-- Anime-style illustrations
-- Character-focused images
-- Cleaner anime aesthetics
-- Style experiments
-- Testing Anima-based fine-tunes inside ComfyUI
+---
 
-This is the **first release** — only the Diffusion Model variant for now. If testing goes well, an **AIO** version and a **4-Step Turbo** version (based on the new **CDM — Continuous-Time Distribution Matching** distillation method) will follow.
+## 🆕 What changed in V2
 
----
+V2 moves from Anima Preview3 onto the final Anima v1.0 / v1.1 releases, and it changes how the
+model was made.
 
-## 🎯 Key Features
+**V1 was a full fine-tune. V2 is not.** V2 is a **LoKr merge**: a LoKr network — linear 64 /
+alpha 64, conv 16 / alpha 16, full-rank, factor 4 — was trained for roughly 24,000 steps on
+`anima-base-v1.0`, then baked into three different Anima variants at strength 1.0. That is a
+lighter touch than a full fine-tune, and it is the honest description of what these files are.
 
-- ✅ Full fine-tune on Anima Preview3 Base — **not a LoRA merge**
-- ✅ ~20K training steps on a curated anime dataset
-- ✅ Clean, focused anime aesthetics
-- ✅ Supports **Danbooru-style tags**, **natural language**, and **mixed prompts**
-- ✅ Compatible with the standard Anima ComfyUI workflow
-- ✅ Drop-in replacement for `anima-preview3-base.safetensors`
-- ✅ Uses the existing **Qwen text encoder** + **Qwen-Image VAE** — included in the repo
-- ✅ LoRA training friendly — same base architecture as Anima
+The upside of working that way: one training run gives three flavours. Same AniSee character,
+three foundations with genuinely different behaviour.
 
 ---
 
-## 🗺️ AniSee Roadmap
+## 🔬 How the three variants actually differ
 
-### ✅ Released
+The three were run side by side across ten prompts — seven natural language, three pure
+Danbooru tags — with the seed locked identical across all three models. What came out of it:
 
-#### 🎨 AniSee Base
+- **Turbo diverges the most.** On the same seed it lands on a different pose and framing far
+  more often than Base and Aesthetic differ from each other. That is the distillation of the
+  Turbo foundation, not a flaw in the merge — you trade diversity for speed and stability.
+- **Base and Aesthetic are near-twins in composition.** Same seed, almost the same layout. The
+  difference is in the finish: Aesthetic has richer colour and softer light, Base is flatter
+  and more neutral.
+- **Tag prompts still work.** The V2 training data is natural language only, with no Danbooru
+  tags at all — yet the base model's tag understanding survives the merge completely intact.
+- **Text rendering stays limited.** Anima places a single word or short phrase reasonably well;
+  longer lines come out garbled. Inherited from the base model; V2 does not improve it.
 
-Full fine-tune of Anima Preview3 Base — Diffusion Model variant. This is the foundation of the AniSee family.
+---
 
-### 🔜 Planned
+## 🔧 Installation (ComfyUI)
 
-#### 📦 AniSee AIO
+Download the checkpoint you want plus the text encoder and the VAE, and drop them here:
 
-All-in-one checkpoint with **Diffusion Model + Qwen Text Encoder + Qwen-Image VAE** integrated into a single file. Single-file convenience, no extra loaders needed.
+```
+ComfyUI/models/diffusion_models/   AniSee-V2-Turbo.safetensors
+ComfyUI/models/text_encoders/      qwen_3_06b_base.safetensors
+ComfyUI/models/vae/                qwen_image_vae.safetensors
+```
 
-#### 🚀 AniSee Turbo (4-Step, CDM)
+Then wire up the standard Anima workflow:
 
-If testing of the Base goes well, a 4-Step Turbo variant distilled with the brand-new **Continuous-Time Distribution Matching (CDM)** method (Liu et al., 2026). CDM migrates the DMD framework from discrete anchoring to continuous optimization, achieving state-of-the-art few-step generation **without GAN or reward-model auxiliary objectives**. Should give clean 4-step anime generations with strong fine details.
+- **Load Diffusion Model** → `AniSee-V2-Turbo.safetensors`
+- **Load CLIP** → `qwen_3_06b_base.safetensors` (type: `stable_diffusion`)
+- **Load VAE** → `qwen_image_vae.safetensors`
 
-📄 Paper: <https://byliutao.github.io/cdm_page/>
+⚠️ These are diffusion model files, so use **Load Diffusion Model**, *not* the Checkpoint
+Loader. If you already run Anima you have the text encoder and the VAE already.
 
-#### 🔧 Official AniSee ComfyUI Workflow
+### With `huggingface_hub`
 
-A dedicated workflow with the auto-prefix, optional Qwen3-VL prompt enhancer, LoRA support, and Ultimate SD Upscale is already included in this repo under `workflows/AniSee.json`.
+```python
+from huggingface_hub import hf_hub_download
 
-More updates coming as testing progresses! 🎨
+model = hf_hub_download("SeeSee21/AniSee", "AniSee-V2-Turbo.safetensors")
+te    = hf_hub_download("SeeSee21/AniSee", "text_encoders/qwen_3_06b_base.safetensors")
+vae   = hf_hub_download("SeeSee21/AniSee", "vae/qwen_image_vae.safetensors")
+```
 
 ---
 
-## ⚙️ AniSee Base — Recommended Settings
+## ⚙️ Recommended settings
 
-The settings I personally use and recommend as a starting point:
+Sampler `er_sde` with scheduler `simple` for every variant — neutral style, flat colours,
+sharp lines.
 
-```yaml
-Steps:       40
-CFG:         4.5
-Sampler:     er_sde
-Scheduler:   simple
-Resolution:  ~1 MP   # e.g. 1024×1024, 896×1152, 1152×896
-```
+| Checkpoint | Steps | CFG | Note |
+|---|---|---|---|
+| AniSee-V2-Turbo | 12 | 1 | Negative prompt has no effect at CFG 1 |
+| AniSee-V2-Aesthetic | 40 | 4–5 | Leave out `score_*` tags |
+| AniSee-V2-Base | 40 | 4–5 | Most neutral, most flexible |
+| anisee (v1) | 40 | 4.5 | The original recommendation |
 
-**CFG Guide:** 4.0–5.0 is the sweet spot for balanced quality and creativity. Going above 5.0 starts to risk burning the image, especially with heavy quality tags. If results feel too harsh, drop CFG slightly or reduce quality tag count.
+**CFG guide.** 4.0–5.0 is the sweet spot for the 40-step variants. Above 5.0 you start risking
+burnt images, especially with heavy quality tags. If results feel harsh, drop CFG slightly or
+cut down on quality tags.
 
-**Sampler alternatives** (all work well, just different character):
+**On Aesthetic:** per the Anima documentation, skip `score_*` tags entirely — in both the
+positive and the negative prompt. The base is already high quality and score tags push it into
+slop territory. `masterpiece, best quality` is fine to keep.
 
-| Sampler / Scheduler | Character |
-| --- | --- |
-| `er_sde` + `simple` *(default)* | Neutral style, flat colors, sharp lines |
-| `euler_a` | Softer, thinner lines, slightly more 2.5D feel, tolerates higher CFG |
-| `dpmpp_2m_sde_gpu` | Similar to er_sde but more "creative", can get wild on short prompts |
+### Sampler alternatives
 
-Feel free to experiment — these are just starting points, not hard rules.
+| Sampler | Character |
+|---|---|
+| `er_sde` + `simple` | The default. Neutral, flat colours, sharp lines |
+| `euler_a` | Softer, thinner lines, slightly 2.5D, tolerates higher CFG |
+| `dpmpp_2m_sde_gpu` | Similar to er_sde but more creative; can get wild on short prompts |
+| `euler` | A bit more creative. Good on Turbo and Aesthetic, which are naturally stable |
 
 ---
 
-## 📐 Resolution Guide
+## 📐 Resolution
 
-| Use Case | Resolution |
-| --- | --- |
-| ⭐ Square / General purpose | **1024 × 1024** |
-| Portrait / Character art | **896 × 1152** |
-| Landscape / Scenes | **1152 × 896** |
-| Wider cinematic | **1254 × 836** |
-| Widescreen | **1365 × 768** |
+| Use case | Resolution |
+|---|---|
+| ⭐ Square / general purpose | 1024 × 1024 |
+| Portrait / character art | 896 × 1152 or 832 × 1216 |
+| Landscape / scenes | 1152 × 896 |
+| Wider cinematic | 1254 × 836 |
+| Widescreen | 1365 × 768 |
 
-Stay around **1 MP** for the cleanest results. The Anima base starts breaking down somewhere around 2 MP, so if you want bigger images, generate at 1 MP first and upscale afterwards.
+Stay around 1 MP for the cleanest results. Anima works between 512² and 1536², but starts
+breaking down somewhere around 2 MP — generate at 1 MP and upscale afterwards for bigger images.
 
 ---
 
-## 💡 Prompting Guide
+## 💡 Prompting
 
-AniSee inherits Anima's prompting system. It accepts:
-
-- Danbooru / anime-style tags
-- Natural language prompts
-- Mixed prompts (tags + sentences)
-
-A good prompt structure:
+AniSee inherits Anima's prompting system and accepts Danbooru-style tags, natural language, and
+any mix of the two.
 
 ```
-[quality tags] [meta tags] [safety tag] [subject (1girl/1boy/etc)]
-[character] [appearance] [pose] [clothing] [background] [lighting] [style]
+[quality tags] [meta tags] [safety tag] [subject] [character]
+[appearance] [pose] [clothing] [background] [lighting] [style]
 ```
 
-**Important tag rules** (inherited from Anima):
+### Tag rules inherited from Anima
+
+- Lowercase tags, spaces instead of underscores
+- Score tags are the only ones using underscores, e.g. `score_7`
+- Artist tags need an `@` prefix, e.g. `@artistname` — without it the effect is very weak
+- Where a tag differs between Danbooru and Gelbooru, prefer the Gelbooru spelling
+- Prompt weighting works but needs heavier weights than SDXL, e.g. `(chibi:2)`
 
-- Use **lowercase** for tags, spaces instead of underscores
-- **Score tags** are the only tags that use underscores (`score_7`, etc.)
-- **Artist tags** must be prefixed with `@` — e.g. `@artistname`
+### 🔑 V2 trigger words
 
-### ✅ Good (mixed prompt)
+The V2 dataset was captioned in **natural language only**, so these want to be used in
+sentences rather than dropped in as bare tags:
+
+| Trigger | Concept |
+|---|---|
+| `seesee_elf` | White-haired elf |
+| `seesee_kitsune` | Fox woman |
+
+Give them a described setting, two sentences minimum. Very short prompts produce unpredictable
+results on Anima — the model fills the gaps with its own biases, and you may not like what it
+picks.
+
+### ✅ Good — mixed prompt
 
 ```
 masterpiece, best quality, score_7, highres, illustration, safe, 1girl,
@@ -200,7 +225,7 @@ at night, neon lights reflecting on wet asphalt, cinematic lighting,
 detailed anime illustration
 ```
 
-### ✅ Good (natural language)
+### ✅ Good — natural language
 
 ```
 masterpiece, best quality, score_7, highres, illustration.
@@ -213,33 +238,21 @@ detailed fabric shading, calm serene expression.
 
 ### ❌ Avoid
 
-Very short tag dumps like `anime girl, silver hair, hoodie` — the model can produce unexpected results when the prompt is too sparse. Aim for at least a few descriptive tags or 2+ sentences.
-
----
+```
+anime girl, silver hair, hoodie
+```
 
-## ⭐ Recommended Positive Prefix
+Too sparse. Aim for a decent set of descriptive tags, or two or more sentences.
 
-Start every prompt with:
+### ⭐ Recommended positive prefix
 
 ```
 masterpiece, best quality, score_7, highres, illustration,
 ```
 
-Then add your subject, character, scene, and style tags after that.
+On **Aesthetic**, drop the `score_7` and use `masterpiece, best quality,` on its own.
 
-You can also experiment with other quality tag combinations:
-
-- `masterpiece, best quality, score_7, safe` *(Anima default)*
-- `masterpiece, best quality, score_8, highres, official art`
-- `score_9, masterpiece, absurdres, anime screenshot`
-
-But the prefix above is what I personally use and recommend as a starting point.
-
----
-
-## ⭐ Recommended Negative Prompt
-
-This is the negative I run with — it cleans up most common issues without being so aggressive that it kills the style:
+### ⭐ Recommended negative prompt
 
 ```
 worst quality, low quality, score_1, score_2, score_3, artist name,
@@ -249,167 +262,102 @@ patreon username, web address, signature, watermark, artist name,
 censored, mosaic censoring
 ```
 
-If your images come out too flat or lose style, reduce the weights on the heavier terms (e.g. drop `(low quality:1.4)` back to `low quality`).
-
----
-
-## 🛡️ Safety Tags
+If images come out flat or lose style, ease off the heavy weights — drop `(low quality:1.4)`
+back to plain `low quality`. On Aesthetic, remove the `score_*` terms. On Turbo at CFG 1 it is
+ignored entirely.
 
-Inherited from Anima. Use one of these in the positive prompt:
+### 🛡️ Safety tags
 
-- `safe` — for normal generations *(recommended default)*
-- `sensitive`
-- `nsfw`
-- `explicit`
+Inherited from Anima. Use one in the positive prompt: `safe` (recommended default),
+`sensitive`, `nsfw`, or `explicit`.
 
 ---
 
-## 🔧 Installation
-
-### Step 1 — Download the files
-
-You need three files (all included in this repo):
-
-- `anisee.safetensors` — the model
-- `text_encoders/qwen_3_06b_base.safetensors` — text encoder
-- `vae/qwen_image_vae.safetensors` — VAE
-
-### Step 2 — Place the files
+## 📁 Repository structure
 
 ```
-ComfyUI/models/diffusion_models/
-└── anisee.safetensors
-
-ComfyUI/models/text_encoders/
-└── qwen_3_06b_base.safetensors
-
-ComfyUI/models/vae/
-└── qwen_image_vae.safetensors
+AniSee-V2-Turbo.safetensors          3.9 GB   V2, distilled foundation, 12 steps
+AniSee-V2-Aesthetic.safetensors      3.9 GB   V2, aesthetic foundation, 40 steps
+AniSee-V2-Base.safetensors           3.9 GB   V2, neutral foundation, 40 steps
+anisee.safetensors                   3.9 GB   V1, full fine-tune of Preview3
+
+text_encoders/qwen_3_06b_base.safetensors     Qwen3-0.6B text encoder
+vae/qwen_image_vae.safetensors                Qwen-Image VAE
+
+images/v2/comparison/                 10 three-way comparisons
+images/v2/turbo/                      20 samples
+images/v2/aesthetic/                  20 samples
+images/v2/base/                       20 samples
+images/                               V1 gallery
+
+workflows/                            ComfyUI workflows
+config.json                           Model metadata
 ```
 
-If you already run **Anima Preview3 Base**, you already have the text encoder and VAE — AniSee is a direct drop-in.
-
-### Step 3 — Load in ComfyUI
-
-Use the standard Anima workflow, or the official AniSee workflow from `workflows/anisee-workflow-SDUltimate.json`:
-
-- **Load Diffusion Model** → `anisee.safetensors`
-- **Load Text Encoder** → `qwen_3_06b_base.safetensors`
-- **Load VAE** → `qwen_image_vae.safetensors`
-
-Then your usual sampler, encode, decode, save chain.
-
 ---
 
-## 🧩 Official Workflow
-
-<div align="center">
-  <img src="https://huggingface.co/SeeSee21/AniSee/resolve/main/images/anisee-workflow-cover.png" alt="AniSee Workflow" width="100%">
-</div>
+## 📈 Version history
 
-A ready-to-use ComfyUI workflow is included at [`workflows/anisee-workflow-SDUltimate.json`](./workflows/anisee-workflow-SDUltimate.json).
+### V2 — LoKr merge onto Anima v1.0 / v1.1
 
-It features:
+- Three variants released together: Turbo, Aesthetic, Base
+- LoKr network — linear 64 / alpha 64, conv 16 / alpha 16, full-rank, factor 4
+- ~24,000 training steps on `anima-base-v1.0`, merged at strength 1.0
+- Natural-language-only dataset; trigger words `seesee_elf` and `seesee_kitsune`
+- Diffusion model files — text encoder and VAE included in this repo
+- Moves off Anima Preview3 onto the final Anima v1.0 / v1.1 releases
 
-- 📦 Model + Text Encoder + VAE loaders pre-configured
-- 🔗 **Auto Quality Prefix** — no need to type `masterpiece, best quality, score_7, ...` yourself
-- 🎲 **Optional Qwen3-VL Prompt Enhancer** — converts short one-liners into full Danbooru tag lists
-- 📖 Optional **LoRA** stack via Lora Manager (one-click toggle)
-- 🔼 Optional **UltimateSDUpscale** 2× with side-by-side compare
-- 🎨 Pre-configured with `er_sde` / `simple` / 40 steps / CFG 4.5
-- ➖ Pre-loaded recommended negative prompt
-- 📝 Built-in MarkdownNote with all settings + quick reference
+### v1.0 — initial release
 
-**Required custom nodes** (all installable via ComfyUI Manager):
-
-- [ComfyUI-Easy-Use](https://github.com/yolain/ComfyUI-Easy-Use)
-- [ComfyUI_UltimateSDUpscale](https://github.com/ssitu/ComfyUI_UltimateSDUpscale)
-- [ComfyUI-Lora-Manager](https://github.com/willmiao/ComfyUI-Lora-Manager)
-- [ComfyUI-QwenVL](https://github.com/1038lab/ComfyUI-QwenVL)
-- [rgthree-comfy](https://github.com/rgthree/rgthree-comfy)
-
-For the optional 2× upscaler, also place `4x-UltraSharp.pth` in `ComfyUI/models/upscale_models/`:
-
-- [OpenModelDB — 4x-UltraSharp](https://openmodeldb.info/models/4x-UltraSharp)
-- [HuggingFace — Kim2091/UltraSharp](https://huggingface.co/Kim2091/UltraSharp)
+- Full fine-tune of Anima Preview3 Base, ~20,000 steps on a curated anime dataset
+- LLM adapter only very lightly co-trained, per Anima's fine-tuning guidelines
+- Drop-in replacement for `anima-preview3-base.safetensors`
 
 ---
 
-## 📁 Repository Structure
-
-```
-AniSee/
-├── README.md
-├── config.json
-│
-├── anisee.safetensors                    # the model (~4.18 GB)
-│
-├── text_encoders/
-│   └── qwen_3_06b_base.safetensors       # text encoder (same as Anima)
-│
-├── vae/
-│   └── qwen_image_vae.safetensors        # VAE (same as Anima)
-│
-├── images/
-│   ├── cover.png                         # social preview / model cover
-│   ├── anisee-workflow-cover.png         # workflow preview image
-│   ├── 1.png   2.png   3.png   4.png
-│   ├── 5.png   6.webp  7.webp  8.png
-│   ├── 9.png   10.png  11.webp 12.webp
-│   └── ...
-│
-└── workflows/
-    └── anisee-workflow-SDUltimate.json
-```
-
----
+## 🗺️ Roadmap
 
-## 📈 Version History
+**✅ Released** — AniSee v1, AniSee V2 (Turbo, Aesthetic, Base)
 
-### v1.0 — Initial Release
+**🔜 Planned**
 
-- **AniSee Base** — full fine-tune of Anima Preview3 Base
-- ~20K training steps on a curated anime dataset
-- LLM adapter only very lightly co-trained *(following Anima's fine-tuning guidelines)*
-- Diffusion Model variant *(single `.safetensors` file)*
-- Compatible with the standard Anima ComfyUI workflow
-- Drop-in replacement for `anima-preview3-base.safetensors`
-- Includes the official ComfyUI workflow with auto quality prefix + Qwen3-VL prompt enhancer
+- **AniSee V2 AIO** — an all-in-one build of V2, so V2 gets the same one-file convenience the
+  V1 AIO has on CivitAI
+- **Official AniSee ComfyUI workflow** — a dedicated workflow, though the standard Anima
+  workflow already covers both generations
 
 ---
 
 ## 🔗 Links
 
-- **CivitAI Page:** [civitai.red/models/2628747/anisee](https://civitai.red/models/2628747/anisee)
-- **Example Gallery:** [anisee.anisee.workers.dev](https://anisee.anisee.workers.dev/)
-- **Base Model:** [circlestone-labs/Anima](https://huggingface.co/circlestone-labs/Anima)
-- **CDM Paper** *(planned Turbo variant)*: [byliutao.github.io/cdm_page](https://byliutao.github.io/cdm_page/)
-- **Author:** [SeeSee21 on Hugging Face](https://huggingface.co/SeeSee21)
+- [AniSee on CivitAI](https://civitai.com/models/2628747/anisee) — including the V1 all-in-one build
+- [AniSee sample gallery](https://anisee.anisee.workers.dev/)
+- [Anima by CircleStone Labs](https://huggingface.co/circlestone-labs/Anima) — the base model
 
 ---
 
 ## 🙏 Credits
 
-- **Base Model:** [Anima Preview3 Base](https://huggingface.co/circlestone-labs/Anima) by **CircleStone Labs** and **Comfy Org**
-- **Underlying Architecture:** Built on NVIDIA Cosmos-Predict2-2B *(Anima is a "Derivative Model")*
-- **Fine-Tune:** SeeSee21
-- **CDM Distillation Method** *(planned Turbo variant)*: [Continuous-Time Distribution Matching for Few-Step Diffusion Distillation](https://byliutao.github.io/cdm_page/) — Liu et al., 2026
-- **Workflow Custom Nodes:** yolain, ssitu, Will Miao, AILab (1038lab), rgthree
+- **Base model:** Anima by CircleStone Labs and Comfy Org — Preview3 Base for V1, v1.0 / v1.1 for V2
+- **Architecture:** built on NVIDIA Cosmos-Predict2-2B; Anima is a derivative model
+- **Fine-tune and merges:** SeeSee21
 
 ---
 
 ## 📜 License
 
-AniSee inherits the **CircleStone Labs Non-Commercial License** from Anima. The model and derivatives are usable **only for non-commercial purposes**. As a derivative of Cosmos-Predict2-2B-Text2Image, the [NVIDIA Open Model License Agreement](https://www.nvidia.com/en-us/agreements/enterprise-software/nvidia-open-model-license/) also applies insofar as it covers Derivative Models.
+AniSee inherits the **CircleStone Labs Non-Commercial License** from Anima. The model and its
+derivatives may be used for non-commercial purposes only. As a derivative of
+Cosmos-Predict2-2B-Text2Image, the **NVIDIA Open Model License Agreement** also applies insofar
+as it covers derivative models.
 
-For commercial licensing of the base model, please contact CircleStone Labs at `tdrussell@circlestone.ai`.
+**Generated images are not covered by that restriction.** You may use the images you make
+commercially — selling images, paid commissions, concept art or assets for a paid product are
+all fine. What needs a separate license is hosting the model behind a paid API, embedding the
+weights in a monetized product, or running it on a paid generation platform.
 
----
+For commercial licensing of the base model, contact CircleStone Labs at `tdrussell@circlestone.ai`.
 
-## ❤️ Notes
-
-AniSee is a personal anime-focused fine-tune of Anima Preview3 Base, built to bring a stronger anime look and visual direction in line with my other checkpoints.
-
-It is still in active testing — the AIO and 4-Step Turbo (CDM) variants will follow once the Base has been validated in the wild.
+---
 
-**AniSee — clean anime, built on Anima. 🎨**
+*AniSee — a personal anime model built on Anima.* 🎨

config.json

@@ -7,26 +7,80 @@
   "base_model_relation": "finetune",
   "author": "SeeSee21",
   "pipeline_tag": "text-to-image",
-  "training": {
-    "steps": "~20000",
-    "dataset": "curated anime dataset",
-    "adapter_co_training": "lightly co-trained (per Anima guidelines)"
-  },
+  "recommended_variant": "v2_turbo",
   "prompting": {
     "style": "tags + natural-language (mixed prompts supported)",
     "negative_prompt_support": "full",
     "recommended_prefix": "masterpiece, best quality, score_7, highres, illustration,",
+    "aesthetic_prefix_note": "On AniSee-V2-Aesthetic, omit score_* tags in both positive and negative.",
     "tag_rules": {
       "case": "lowercase",
       "separator": "spaces (not underscores)",
       "score_tags_exception": "score_7, score_8, etc. use underscores",
-      "artist_tags": "must be prefixed with @ (e.g. @artistname)"
-    }
+      "artist_tags": "must be prefixed with @ (e.g. @artistname)",
+      "weighting": "works, but needs heavier weights than SDXL (e.g. (chibi:2))"
+    },
+    "trigger_words": {
+      "seesee_elf": "white-haired elf (V2 only)",
+      "seesee_kitsune": "fox woman (V2 only)"
+    },
+    "trigger_note": "The V2 dataset was captioned in natural language only. Use trigger words inside descriptive sentences, not as bare tags."
   },
   "variants": {
-    "base": {
+    "v2_turbo": {
+      "file": "AniSee-V2-Turbo.safetensors",
+      "size": "~3.90 GB",
+      "generation": "V2",
+      "method": "LoKr merge",
+      "merged_into": "anima-turbo-v1.1",
+      "comfyui_path": "ComfyUI/models/diffusion_models/",
+      "recommended_settings": {
+        "steps": 12,
+        "cfg": 1,
+        "sampler": "er_sde",
+        "scheduler": "simple",
+        "resolution": "832x1216 (portrait, ~1 MP)"
+      },
+      "notes": "Fastest variant, ~6 s per image. The negative prompt has no effect at CFG 1."
+    },
+    "v2_aesthetic": {
+      "file": "AniSee-V2-Aesthetic.safetensors",
+      "size": "~3.90 GB",
+      "generation": "V2",
+      "method": "LoKr merge",
+      "merged_into": "anima-aesthetic-v1.1",
+      "comfyui_path": "ComfyUI/models/diffusion_models/",
+      "recommended_settings": {
+        "steps": 40,
+        "cfg": 4.5,
+        "sampler": "er_sde",
+        "scheduler": "simple",
+        "resolution": "832x1216 (portrait, ~1 MP)"
+      },
+      "notes": "Richest colour and softest light. Do not use score_* tags with this variant."
+    },
+    "v2_base": {
+      "file": "AniSee-V2-Base.safetensors",
+      "size": "~3.90 GB",
+      "generation": "V2",
+      "method": "LoKr merge",
+      "merged_into": "anima-base-v1.0",
+      "comfyui_path": "ComfyUI/models/diffusion_models/",
+      "recommended_settings": {
+        "steps": 40,
+        "cfg": 4.5,
+        "sampler": "er_sde",
+        "scheduler": "simple",
+        "resolution": "832x1216 (portrait, ~1 MP)"
+      },
+      "notes": "Most neutral and most flexible. The right foundation for training your own LoRAs."
+    },
+    "v1": {
       "file": "anisee.safetensors",
-      "size": "~4.18 GB",
+      "size": "~3.89 GB",
+      "generation": "V1",
+      "method": "full fine-tune",
+      "merged_into": "anima-preview3-base",
       "comfyui_path": "ComfyUI/models/diffusion_models/",
       "recommended_settings": {
         "steps": 40,
@@ -35,21 +89,41 @@
         "scheduler": "simple",
         "resolution": "832x1216 (portrait, ~1 MP)"
       },
-      "alternative_samplers": [
-        "euler_a",
-        "dpmpp_2m_sde_gpu"
-      ]
+      "notes": "Drop-in replacement for anima-preview3-base.safetensors."
     }
   },
+  "training": {
+    "v2": {
+      "method": "LoKr, merged at strength 1.0",
+      "network": {
+        "linear": 64,
+        "linear_alpha": 64,
+        "conv": 16,
+        "conv_alpha": 16,
+        "full_rank": true,
+        "factor": 4
+      },
+      "steps": "~24000",
+      "trained_on": "anima-base-v1.0",
+      "dataset": "curated dataset, natural-language captions only"
+    },
+    "v1": {
+      "method": "full fine-tune",
+      "steps": "~20000",
+      "trained_on": "anima-preview3-base",
+      "dataset": "curated anime dataset",
+      "adapter_co_training": "lightly co-trained (per Anima guidelines)"
+    }
+  },
+  "alternative_samplers": [
+    "euler_a",
+    "dpmpp_2m_sde_gpu",
+    "euler"
+  ],
   "planned_variants": {
-    "aio": {
-      "description": "All-in-one checkpoint (Model + Text Encoder + VAE in one file)",
+    "v2_aio": {
+      "description": "All-in-one V2 checkpoint (model + text encoder + VAE in one file)",
       "status": "planned"
-    },
-    "turbo_4step": {
-      "description": "4-Step Turbo variant distilled with CDM (Continuous-Time Distribution Matching)",
-      "paper": "https://byliutao.github.io/cdm_page/",
-      "status": "planned (pending Base validation)"
     }
   },
   "components": {
@@ -88,17 +162,16 @@
   },
   "supported_vram": "8GB+",
   "links": {
-    "civitai": "https://civitai.red/models/2628747/anisee",
+    "civitai": "https://civitai.com/models/2628747/anisee",
     "gallery": "https://anisee.anisee.workers.dev/",
     "base_model": "https://huggingface.co/circlestone-labs/Anima",
-    "cdm_paper": "https://byliutao.github.io/cdm_page/",
     "author": "https://huggingface.co/SeeSee21"
   },
   "notes": [
-    "Drop-in replacement for anima-preview3-base.safetensors.",
-    "Uses the same Qwen text encoder and Qwen-Image VAE as Anima.",
+    "V1 is a full fine-tune of Anima Preview3 Base. V2 is a LoKr merge onto Anima v1.0 / v1.1 — a different method, not a fine-tune.",
+    "All checkpoints are diffusion model files: load with Load Diffusion Model, not the Checkpoint Loader.",
+    "The Qwen text encoder and the Qwen-Image VAE are included in this repo.",
     "Tag rules: lowercase, spaces (not underscores), except score_7 etc. Artist tags prefixed with @.",
-    "Official ComfyUI workflow included in workflows/ with auto Quality Prefix and Qwen3-VL Prompt Enhancer.",
-    "AIO and 4-Step Turbo (CDM) variants planned for future releases."
+    "An all-in-one V1 checkpoint is available on CivitAI; a V2 all-in-one is planned."
   ]
 }