Blame for README.md

---
license: other
license_name: krea-2-community-license
license_link: LICENSE.md
base_model: krea/Krea-2-Turbo
base_model_relation: finetune
pipeline_tag: text-to-image
inference: false
extra_gated_prompt: "By accessing this model you agree to be bound by the Krea 2 Community License Agreement (LICENSE.md) and the Krea 2 Acceptable Use Policy (ACCEPTABLE_USE_POLICY.md)."
extra_gated_fields:
  I agree to the Krea 2 Community License Agreement and the Acceptable Use Policy: checkbox
tags:
  - text-to-image
  - anime
  - krea2
  - krea-2
  - pixel-art
  - comfyui
  - safetensors
  - checkpoint
language:
  - en
---

<div align="center">

<img src="https://huggingface.co/SeeSee21/See-Krea2-Genesis/resolve/main/images/cover.jpg" alt="See-Krea2 cover" width="100%">

# 🎨 See-Krea2 · Genesis V2

### More detail. Richer scenes. A personal pixel-art look — an anime fine-tune of Krea 2 Turbo

<a href="https://huggingface.co/SeeSee21"><img src="https://img.shields.io/badge/Hugging%20Face-SeeSee21-FFD21E?style=for-the-badge&logo=huggingface&logoColor=black" alt="Hugging Face"></a>
<a href="https://civitai.red/models/2846618"><img src="https://img.shields.io/badge/Civitai-This%20model-1971C2?style=for-the-badge" alt="Civitai"></a>
<a href="https://www.youtube.com/@SeeSee-AI"><img src="https://img.shields.io/badge/YouTube-SeeSee--AI-FF0000?style=for-the-badge&logo=youtube&logoColor=white" alt="YouTube"></a>
<a href="https://github.com/seesee75-commits"><img src="https://img.shields.io/badge/GitHub-seesee75--commits-181717?style=for-the-badge&logo=github&logoColor=white" alt="GitHub"></a>

<p align="center">
<a href="https://huggingface.co/SeeSee21"><img src="https://img.shields.io/badge/Type-Checkpoint-7C3AED?style=flat-square" alt="Type"></a> <a href="https://huggingface.co/krea/Krea-2-Turbo"><img src="https://img.shields.io/badge/Base-Krea%202%20Turbo-0EA5E9?style=flat-square" alt="Base"></a> <a href="https://huggingface.co/SeeSee21"><img src="https://img.shields.io/badge/Format-BF16%20·%20INT8%20·%20FP8%20·%20NVFP4%20·%20INT4-16A34A?style=flat-square" alt="Format"></a> <a href="https://huggingface.co/SeeSee21"><img src="https://img.shields.io/badge/Steps-8-EF4444?style=flat-square" alt="Steps"></a>
</p>

</div>

<br>

> **Krea 2 See-Krea2-Genesis** is a modified version (fine-tune) of **Krea 2 Turbo** by Krea.ai, Inc., made by SeeSee21. It is **not an official Krea product and is not endorsed by Krea.** Krea 2 is licensed under the Krea 2 Community License Agreement. For more information, visit https://krea.ai/krea-2-licensing. See [LICENSE.md](LICENSE.md), [NOTICE](NOTICE) and [ACCEPTABLE_USE_POLICY.md](ACCEPTABLE_USE_POLICY.md). On Civitai this model is published as *See-Krea2*.

<div style="background:#7C3AED22;border-left:4px solid #7C3AED;padding:12px 16px;border-radius:8px">
<b>🚀 Quick start:</b> download the <b>Genesis-V2</b> file you want (INT8 ConvRot is the workflow default) plus the <b>Qwen3-VL text encoder</b> and the <b>Qwen-Image VAE</b>, put the diffusion model in <code>ComfyUI/models/diffusion_models/</code>, and use the included workflow with <b>8 steps</b>, <b>CFG 1.0</b>, euler / simple. Trigger words: <code>seesee_elf</code>, <code>seesee_kitsune</code>.
</div>

## ✨ At a glance

| Property | Details |
|:--|:--|
| 🧩 **Type** | Anime fine-tune of Krea 2 Turbo (diffusion model files; text encoder and VAE are separate) |
| 🏗️ **Base model** | [krea/Krea-2-Turbo](https://huggingface.co/krea/Krea-2-Turbo) (Krea 2 Community License) |
| 📦 **Versions** | ⭐ Genesis-V2 (current) · Genesis-V1 (BF16 and compressed builds) |
| 📦 **Formats** | BF16 master (26.28 GB) · INT8 ConvRot (13.15 GB) · FP8 scaled (13.14 GB) · MXFP8 (14.51 GB) · NVFP4 (7.67 GB) · INT4 ConvRot (6.90 GB) |
| 🎛️ **Settings** | 8 steps · CFG 1.0 · euler / simple · natural-language prompts |
| 🔑 **Triggers** | `seesee_elf` · `seesee_kitsune` · pixel-art and Game Boy look |
| 🧰 **Workflows** | In `workflows/` ([Civitai page](https://civitai.red/models/2936142)) |

**Jump to:** [🚀 Start here](#🚀-start-here) · [🖼️ Gallery](#🖼️-gallery) · [⬇️ Downloads](#⬇️-downloads) · [⚙️ Settings and prompting](#⚙️-settings-and-prompting) · [🧩 Workflow](#🧩-workflow) · [📚 More details](#📚-more-details) · [📜 License](#📜-license)

## 🚀 Start here

**Three steps to your first image**

1. **Download one diffusion-model file** from [⬇️ Downloads](#⬇️-downloads). Not sure which? Take **INT8 ConvRot** (13.15 GB): best overall quality, the workflow default, works on RTX 30-series and newer.
2. **Download the two shared files:** the Qwen3-VL text encoder and the Qwen-Image VAE (same section). If you already run Krea 2 Turbo, reuse yours.
3. **Load the workflow** `workflows/See-Krea2-Ultimate-v1.1.json` (see [🧩 Workflow](#🧩-workflow)) and generate with **8 steps · CFG 1.0 · euler / simple**.

**Which build fits my GPU?**

| Build | Size | Best for |
|:--|:-:|:--|
| ⭐ **INT8 ConvRot** | 13.15 GB | Best overall quality. Pick this if unsure. Any GPU with INT8 support (RTX 30-series and newer) |
| **FP8 scaled** | 13.14 GB | RTX 40-series and newer (native FP8 path) |
| **MXFP8** | 14.51 GB | Needs a loader with MXFP8 support (newer NVIDIA GPUs) |
| **NVFP4** | 7.67 GB | RTX 50-series (Blackwell): much smaller download |
| **INT4 ConvRot** | 6.90 GB | Low VRAM or smallest download; needs ConvRot W4A4 support |
| **BF16 master** | 26.28 GB | Unquantized reference |

All builds look good side by side. A higher "weight error" does **not** mean worse images (see [📚 More details](#📚-more-details)).

**Prompt cues:** `seesee_elf` · `seesee_kitsune` · for the pixel look add `pixel` or `16Bit`. Use natural-language prompts. The negative prompt has no effect at CFG 1.0.

**Where the files go**

| File | Folder |
|:--|:--|
| Diffusion model (one Genesis V2 file) | `ComfyUI/models/diffusion_models/` |
| Text encoder `qwen3vl_4b_bf16.safetensors` (CLIP type `krea2`) | `ComfyUI/models/text_encoders/` |
| VAE `qwen_image_vae.safetensors` | `ComfyUI/models/vae/` |


## 🖼️ Gallery

*All images were posted to the Genesis-V2 version on Civitai.*

<table>
<tr>
<td><img src="https://huggingface.co/SeeSee21/See-Krea2-Genesis/resolve/main/images/preview_1.jpg" alt="Genesis-V2 preview 1"></td>
<td><img src="https://huggingface.co/SeeSee21/See-Krea2-Genesis/resolve/main/images/preview_2.jpg" alt="Genesis-V2 preview 2"></td>
<td><img src="https://huggingface.co/SeeSee21/See-Krea2-Genesis/resolve/main/images/preview_3.jpg" alt="Genesis-V2 preview 3"></td>
</tr>
<tr>
<td><img src="https://huggingface.co/SeeSee21/See-Krea2-Genesis/resolve/main/images/preview_4.jpg" alt="Genesis-V2 preview 4"></td>
<td><img src="https://huggingface.co/SeeSee21/See-Krea2-Genesis/resolve/main/images/preview_5.jpg" alt="Genesis-V2 preview 5"></td>
<td></td>
</tr>
</table>

## ⬇️ Downloads

**Genesis-V2** is the current version (longer training at higher resolution). Civitai: [Genesis-V2](https://civitai.red/models/2846618?modelVersionId=3320733) · Workflow: [🎨 See-Krea2 Ultimate](#🧩-workflow)

### Diffusion model (pick one)

| File | Format | ComfyUI folder |
|:--|:--|:--|
| [`See-Krea2-Genesisv2.safetensors`](https://huggingface.co/SeeSee21/See-Krea2-Genesis/resolve/main/Genesis-V2/See-Krea2-Genesisv2.safetensors) | BF16 master (26.28 GB) | `models/diffusion_models/` |
| [`See-Krea2-Genesisv2-int8_convrot.safetensors`](https://huggingface.co/SeeSee21/See-Krea2-Genesis/resolve/main/Genesis-V2/See-Krea2-Genesisv2-int8_convrot.safetensors) | INT8 ConvRot (13.15 GB) · workflow default | `models/diffusion_models/` |
| [`See-Krea2-Genesisv2-fp8_scaled.safetensors`](https://huggingface.co/SeeSee21/See-Krea2-Genesis/resolve/main/Genesis-V2/See-Krea2-Genesisv2-fp8_scaled.safetensors) | FP8 scaled (13.14 GB) | `models/diffusion_models/` |
| [`See-Krea2-Genesisv2-mxfp8.safetensors`](https://huggingface.co/SeeSee21/See-Krea2-Genesis/resolve/main/Genesis-V2/See-Krea2-Genesisv2-mxfp8.safetensors) | MXFP8 (14.51 GB) | `models/diffusion_models/` |
| [`See-Krea2-Genesisv2-nvfp4.safetensors`](https://huggingface.co/SeeSee21/See-Krea2-Genesis/resolve/main/Genesis-V2/See-Krea2-Genesisv2-nvfp4.safetensors) | NVFP4 (7.67 GB) | `models/diffusion_models/` |
| [`See-Krea2-Genesisv2-int4_convrot.safetensors`](https://huggingface.co/SeeSee21/See-Krea2-Genesis/resolve/main/Genesis-V2/See-Krea2-Genesisv2-int4_convrot.safetensors) | INT4 ConvRot (6.90 GB) | `models/diffusion_models/` |

### Text encoder and VAE (required for every build)

Qwen components, Apache-2.0.

**Text encoder (pick one)**

Civitai: [Text Encoders](https://civitai.red/models/2846618?modelVersionId=3216537) · *Standard Text Encoders from Original Model* · no SFW gallery image on Civitai.

| File | Format | ComfyUI folder |
|:--|:--|:--|
| [`qwen3vl_4b_fp8_scaled.safetensors`](https://huggingface.co/SeeSee21/See-Krea2-Genesis/resolve/main/Text-Encoder/qwen3vl_4b_fp8_scaled.safetensors) | Qwen3-VL-4B FP8 scaled | `models/text_encoders/` |
| [`qwen3vl_4b_bf16.safetensors`](https://huggingface.co/SeeSee21/See-Krea2-Genesis/resolve/main/Text-Encoder/qwen3vl_4b_bf16.safetensors) | Qwen3-VL-4B BF16 | `models/text_encoders/` |

**VAE**

Civitai: [VAE](https://civitai.red/models/2846618?modelVersionId=3216519) · *Standard VAE from the original model*

| File | Format | ComfyUI folder |
|:--|:--|:--|
| [`qwen_image_vae.safetensors`](https://huggingface.co/SeeSee21/See-Krea2-Genesis/resolve/main/VAE/qwen_image_vae.safetensors) | Qwen-Image VAE | `models/vae/` |


<details>
<summary><b>📄 Genesis-V2 · full Civitai version description (all six builds explained)</b></summary>

🎨 Genesis V2 · BF16 Master

**26.28 GB · Krea 2 Turbo · Anime checkpoint**

The original Genesis V2 checkpoint and the source for all compressed builds. It contains BF16 weights with selected F32 tensors retained. Choose this file if you want the unquantized reference and can accommodate the larger model.

⭐ Genesis V2 · INT8 ConvRot

**13.15 GB · Krea 2 Turbo · Anime checkpoint**

An 8-bit Genesis V2 build with ConvRot rotation, at roughly half the master’s file size. It quantizes 256 layers, including the text-fusion linears, using the original Genesis quantization profile. A starting option for workflows with native INT8 ConvRot support.

⚡ Genesis V2 · FP8 Scaled

**13.14 GB · Krea 2 Turbo · Anime checkpoint**

Genesis V2 in scaled FP8 E4M3, at roughly half the master’s file size. This build quantizes 256 layers and keeps full-precision matrix multiplication enabled on 96 selected layers. Choose it for a workflow that supports scaled FP8.

💾 Genesis V2 · NVFP4 · Full-MM

**7.67 GB · Krea 2 Turbo · Anime checkpoint**

A compact 4-bit floating-point Genesis V2 build. All 256 quantized layers have **full\_precision\_matrix\_mult enabled**. This is the Full-MM variant; it should not be assumed to have the same speed or memory use as an NVFP4 build with that setting disabled. Use a compatible NVFP4 workflow.

🪶 Genesis V2 · INT4 ConvRot

**6.90 GB · Krea 2 Turbo · Anime checkpoint**

The smallest Genesis V2 file in this lineup. It uses ConvRot W4A4 on 256 layers for a substantially smaller download. Choose it when compact storage matters, and compare the results in your own workflow. Requires support for the ConvRot W4A4 format.

🧪 Genesis V2 · MXFP8

**14.51 GB · Krea 2 Turbo · Anime checkpoint**

Genesis V2 in microscaling FP8. This build quantizes **224 main-block layers** and leaves the text-fusion path in higher precision. Its different layer selection makes it larger than the standard FP8 build. MXFP8 is a separate format and needs matching loader and runtime support.

**✨ What’s new in V2**  
Longer training at higher resolution, more image detail and better prompt following. Includes my personal pixel-art look and Game Boy-inspired training material.

**⚙️ Start here**  
8 steps · CFG 1.0 · euler / simple · around 1024 × 1024.

**🗝️ Prompt cues**  
**seesee\_elf** · **seesee\_kitsune**  
For the pixel look, include **pixel** or **16Bit**.

**Use with:** the standard Krea 2 workflow, Qwen3-VL text encoder and Qwen-Image VAE. File size is not a VRAM requirement.

</details>

<details>
<summary><b>🖼️ VAE gallery</b></summary>

All images below were posted to the **VAE** version on Civitai.

<table>
<tr>
<td><img src="https://huggingface.co/SeeSee21/See-Krea2-Genesis/resolve/main/images/preview_11.jpg" alt="VAE preview 1"></td>
<td></td>
<td></td>
</tr>
</table>

</details>


## ⚙️ Settings and prompting

The settings I personally use and recommend as a starting point:

```
Steps:       8
CFG:         1.0
Sampler:     euler
Scheduler:   simple
Denoise:     1.0
Resolution:  ~1 MP, for example 1024 × 1024
```

**Sampler alternatives**

* **euler + simple** — my default. Neutral, clean, predictable.
* **er\_sde + simple** — slightly more texture and character, works well at 10–12 steps.

**⚠️ CFG must stay at 1.0**

This is a turbo model. Raising CFG above 1 will burn the image — that is not a style choice, it breaks.

Two consequences people trip over:

* **Your negative prompt does nothing.** At CFG 1.0 the unconditional branch is never computed, so negative text has no mathematical effect on the result. Whatever you put there is ignored.
* **Exclusions must be phrased positively.** If you do not want gloves, write `bare hands` in the positive prompt. Writing `gloves` in the negative field changes nothing.

### 💡 Prompting guide

Natural-language descriptions are my preferred way to prompt this model. Start with what the image should show, then add the details that matter most.

**A useful order:**  
Medium and style → camera and framing → pose and action → face and hair → clothing → background → lighting.

* **Write clear sentences.** You do not need to hit a particular word count; add detail when it gives the model useful information.
* **Name the framing.** For example: close-up portrait, low-angle shot or wide full-body composition.
* **Be specific about hands and actions.** Describe each hand separately when the pose matters.
* **Describe materials and clothing layers.** This gives the model more to work with than a clothing label alone.
* **Give the background its own description.** V2’s training included detailed environments as well as detailed characters.
* **Keep the subject clear.** Consistent pronouns and an explicit character count help avoid ambiguity.
* **For the pixel look, include pixel or 16Bit.** For the trained character cues, use seesee\_elf or seesee\_kitsune exactly.

If a detail is missing, first make the wording more specific instead of relying on long tag lists or weighting syntax.

## 🧩 Workflow

Civitai: [See-Krea2-Workflows – Ultimate T2I v1.1](https://civitai.red/models/2936142?modelVersionId=3323616) · Download: [`See-Krea2-Ultimate-v1.1.json`](https://huggingface.co/SeeSee21/See-Krea2-Genesis/resolve/main/workflows/See-Krea2-Ultimate-v1.1.json)

![See-Krea2 Ultimate v1.1 – node graph overview](https://huggingface.co/SeeSee21/See-Krea2-Genesis/resolve/main/images/workflows/ultimate/overview.jpg)

**What it does:** 🧩 load models → 🧬 optional LoRA stack → ✍️ prompt → ✨ optional Qwen3-VL prompt enhancer → 🎨 base image 768 × 1368 → 🔬 Detail Refiner (AnimeSharp) and/or 🔼 Tiled Upscale, each with a before/after comparer; PNGs are saved with Civitai-readable metadata.

**⚙️ Settings (workflow page):** ships with Detail Refiner ON · Prompt Enhancer, LoRA Stack and Tiled Upscale OFF · 10 steps · CFG 1.0 · euler_ancestral / beta (model page starting point: 8 steps · euler / simple) · triggers `seesee_elf`, `seesee_kitsune`, pixel look `pixel art` / `16Bit`.

<details>
<summary><b>📦 Required models and custom nodes</b></summary>

**📦 Models used**

| File | Required? | Download | ComfyUI folder |
|:--|:--|:--|:--|
| See-Krea2 Genesis V2 checkpoint (default INT8 ConvRot) | required | [`See-Krea2-Genesisv2-int8_convrot.safetensors`](https://huggingface.co/SeeSee21/See-Krea2-Genesis/resolve/main/Genesis-V2/See-Krea2-Genesisv2-int8_convrot.safetensors) or any file from [⭐ Genesis-V2](#⬇️-downloads) | `models/diffusion_models/` |
| VAE (Qwen-Image VAE) | required | [`qwen_image_vae.safetensors`](https://huggingface.co/SeeSee21/See-Krea2-Genesis/resolve/main/VAE/qwen_image_vae.safetensors) | `models/vae/` |
| Text encoder Qwen3-VL 4B (CLIP type `krea2`) | required | [`qwen3vl_4b_fp8_scaled.safetensors`](https://huggingface.co/SeeSee21/See-Krea2-Genesis/resolve/main/Text-Encoder/qwen3vl_4b_fp8_scaled.safetensors) · [BF16](https://huggingface.co/SeeSee21/See-Krea2-Genesis/resolve/main/Text-Encoder/qwen3vl_4b_bf16.safetensors) | `models/text_encoders/` |
| 2x-AnimeSharpV4_RCAN (Kim2091) | required (Detail Refiner) | [`2x-AnimeSharpV4_RCAN.safetensors`](https://github.com/Kim2091/Kim2091-Models/releases/download/2x-AnimeSharpV4/2x-AnimeSharpV4_RCAN.safetensors) | `models/upscale_models/` |
| Qwen3VL-4B-Instruct-Q4_K_M.gguf (Prompt Enhancer) | optional · downloads automatically on first use | [Qwen/Qwen3-VL-4B-Instruct-GGUF](https://huggingface.co/Qwen/Qwen3-VL-4B-Instruct-GGUF) | `models/LLM/GGUF/` |
| Krea 2 LoRAs (LoRA Stack) | optional | your own | `models/loras/` |

**🔌 Custom nodes**

| Custom node pack (external) | Used for | Required? |
|:--|:--|:--|
| [ComfyUI-SeeSee-Civitai-Metadata](https://github.com/seesee75-commits/ComfyUI-SeeSee-Civitai-Metadata) | save nodes with Civitai metadata | required |
| [rgthree-comfy](https://github.com/rgthree/rgthree-comfy) | group switches, sampling settings, image comparers | required |
| [ComfyUI-Lora-Manager](https://github.com/willmiao/ComfyUI-Lora-Manager) | LoRA stack + trigger words | required |
| [ComfyUI-QwenVL](https://github.com/1038lab/ComfyUI-QwenVL) | GGUF prompt enhancer | required |
| [ComfyUI-Easy-Use](https://github.com/yolain/ComfyUI-Easy-Use) | prompt-rules text node | required |
| [ComfyUI_UltimateSDUpscale](https://github.com/ssitu/ComfyUI_UltimateSDUpscale) | Tiled Upscale | required |

Install with [ComfyUI-Manager](https://github.com/Comfy-Org/ComfyUI-Manager) (search the name) or `git clone` into `custom_nodes/`, then restart ComfyUI. Needs a recent ComfyUI with Krea 2 support and subgraphs (tested on ComfyUI 0.33.0, frontend 1.49.6).

*Further results: the workflow page has only videos and the graph image; no additional SFW result images.*

</details>

<details>
<summary>📄 See-Krea2 Ultimate · full Civitai workflow description</summary>

### 🎨 See-Krea2 Ultimate — Anime Text-to-Image Workflow for Krea 2

ComfyUI workflow for **See-Krea2 Genesis V2**, my anime checkpoint on Krea 2 Turbo. Write an idea, get a detailed anime illustration — with an optional Qwen3-VL prompt enhancer, an optional LoRA stack and two finishing branches. **Every saved PNG carries Civitai-readable generation data:** upload it and Civitai fills in prompt, seed, sampler and the model link by itself.

**How it works:** 🧩 load models → 🧬 optional LoRAs → ✍️ prompt → ✨ optional enhancer (turns tags or German input into one English Krea-style paragraph) → 🎨 base image at 768 × 1368 (9:16, 1 MP) → 🔬 Detail Refiner (AnimeSharp 2× → 1.5 MP → 8 steps at denoise 0.20) and/or 🔼 Tiled Upscale (1.5×, tile by tile) → 💾 every stage saves its own PNG with Civitai metadata. Everything is grouped, color-coded, switchable with one click and documented with notes inside the workflow.

* **Ships with:** Detail Refiner ON · Prompt Enhancer, LoRA Stack and Tiled Upscale OFF
* **Base settings:** 10 steps, CFG 1.0, euler\_ancestral / beta — the model page's starting point is 8 steps, euler / simple. Everything is set centrally in ⚙️ Generation + Hi-Res Settings
* **Triggers:** `seesee_elf`, `seesee_kitsune` · pixel look: `pixel art` or `16Bit`
* **Performance reference:** RTX 5060 Ti 16 GB with the INT8 ConvRot file ≈ 40–60 s for base + Detail Refiner, ≈ 140 s with Tiled Upscale on top

---

#### 📥 Required Models

* **See-Krea2 Genesis V2** checkpoint → `ComfyUI/models/diffusion_models/`  
  Download (version **Genesis-V2**): <https://civitai.red/models/2846618/see-krea2>
* **VAE** (Qwen-Image VAE) → `ComfyUI/models/vae/`  
  Download (version **VAE**): <https://civitai.red/models/2846618/see-krea2>
* **Text encoder** Qwen3-VL 4B, FP8 (BF16 also available) → `ComfyUI/models/text_encoders/`  
  Download (version **Text Encoders**): <https://civitai.red/models/2846618/see-krea2>
* **2x-AnimeSharpV4\_RCAN.safetensors** by Kim2091 → `ComfyUI/models/upscale_models/`  
  Download: <https://github.com/Kim2091/Kim2091-Models/releases/download/2x-AnimeSharpV4/2x-AnimeSharpV4_RCAN.safetensors>

**Checkpoint precision:** INT8 ConvRot (workflow default, 13.15 GB), FP8 scaled (13.14 GB), MXFP8 (14.51 GB), NVFP4 (7.67 GB), INT4 ConvRot (6.90 GB) or the BF16 master (26.28 GB). The compressed formats need matching support in your ComfyUI; file size is not a VRAM requirement.

> ⚠️ Civitai renames files on download (e.g. `seeKrea2_genesisV2_….safetensors`). Open the workflow and pick your file in each loader.  
> The CLIP loader type stays on `krea2`.

**Optional:**

* **Qwen3VL-4B-Instruct-Q4\_K\_M.gguf** for the ✨ Prompt Enhancer — downloads automatically on first use into `ComfyUI/models/LLM/GGUF/`  
  Source: <https://huggingface.co/Qwen/Qwen3-VL-4B-Instruct-GGUF>
* Compatible Krea 2 LoRAs for the 🧬 LoRA Stack — add their trigger words to your prompt yourself

#### 🧩 Required Custom Nodes

Install via **ComfyUI-Manager** (search the name) or git clone into `custom_nodes/`:

* **SeeSee Civitai Metadata** — the save nodes: Civitai-readable metadata in every PNG — <https://github.com/seesee75-commits/ComfyUI-SeeSee-Civitai-Metadata>
* **rgthree-comfy** — group switches, sampling settings, image comparers — <https://github.com/rgthree/rgthree-comfy>
* **ComfyUI-Lora-Manager** — LoRA stack + trigger words — <https://github.com/willmiao/ComfyUI-Lora-Manager>
* **ComfyUI-QwenVL** — GGUF prompt enhancer — <https://github.com/1038lab/ComfyUI-QwenVL>
* **ComfyUI-Easy-Use** — prompt-rules text node — <https://github.com/yolain/ComfyUI-Easy-Use>
* **ComfyUI\_UltimateSDUpscale** — Tiled Upscale — <https://github.com/ssitu/ComfyUI_UltimateSDUpscale>

Needs a recent ComfyUI with Krea 2 support and subgraphs (tested on ComfyUI 0.33.0, frontend 1.49.6).

#### 🛠️ Installation

1. Install the 6 node packs with ComfyUI-Manager → restart ComfyUI.
2. Download checkpoint, VAE, text encoder and AnimeSharp into the folders above.
3. Open the workflow and select your files in the loaders.
4. Write your idea in ✍️ **Positive Prompt**, switch groups on or off with the rgthree toggles, hit **Run**. Images land in `output/` as `See-Krea2_…`, `See-Krea2-Refined_…` and `See-Krea2-Tiled-Upscale_…`.

#### 🏷️ Sharing your images on Civitai

The save nodes write the generation data Civitai reads. Upload the PNG as it is and Civitai shows prompt, seed, steps, sampler, CFG and size, and links See-Krea2 and your LoRAs automatically — no picking resources by hand. The ComfyUI workflow stays inside the PNG, so anyone can drag it back into ComfyUI. Editing or converting the image before upload can strip that data.

#### 📝 Notes

* ✨ **Prompt Enhancer:** check 👁️ **Final Prompt Preview** before important renders — the 4B model can misread details. It has its own seed; keep it to reproduce an expansion.
* **CFG 1.0:** the negative prompt has no effect. Describe the visible state you want instead ("bare hands" rather than "no gloves").
* 🔬 **Detail Refiner** and 🔼 **Tiled Upscale** both start from the base image. Hi-Res Amount 1.5 means a 1.5 MP target in the refiner and 1.5× width and height in the tiled upscale.
* **Tiled Upscale seams:** if you see lines, lower denoise to 0.15; with little VRAM try 768 px tiles or Tiled Decode.
* Adult content: consenting-adult content only, where platform and law allow.

**Credits:** Krea (Krea 2 Turbo) · Qwen (Qwen3-VL, Qwen-Image VAE) · Kim2091 (AnimeSharpV4) · rgthree · willmiao (LoRA Manager) · 1038lab (QwenVL) · yolain (Easy-Use) · ssitu (Ultimate SD Upscale) · checkpoint + workflow: **SeeSee21**

</details>

## 📚 More details

*Everything below is collapsed. Click a title to open it.*

<details>
<summary><b>✨ About the model and what changed in V2</b></summary>

### What is See-Krea2-Genesis?

See-Krea2-Genesis is my personal anime checkpoint built on **Krea 2 Turbo**. The aim is to make anime illustration the model’s natural visual direction while keeping the fast, straightforward Turbo workflow.

**V2 builds on the original Genesis with longer training at higher resolution.** The focus is on more image detail, better prompt following, detailed anime women and richer backgrounds. I also trained in my own pixel look, two character triggers, and a retro handheld-game direction using my own Game Boy captures.

### What changed in V2?

* **Longer training at higher resolution.** V2 received more training with higher-resolution material than V1.
* **More detail in the image.** The training included detailed anime women and detailed backgrounds, with attention to both the character and the surrounding scene.
* **Better prompt following.** V2 follows the scene and character details in your description more closely.
* **Two trained triggers:** **seesee\_elf** and **seesee\_kitsune**.
* **My own pixel-art look.** It can be requested with different phrases, including a simple **pixel art** cue.
* **Game Boy-inspired imagery.** My own Game Boy captures were included in the training data.

**📝 About English, Japanese and manga text:**  
The training included English and Japanese text as well as manga material. **I have not confirmed an improvement in readable text generation.** Please treat spelling, lettering and longer text as something to test, rather than a promised V2 feature.

</details>

<details>
<summary><b>🧝 Trained triggers, pixel-art look and Game Boy scenes</b></summary>

#### Trained character triggers

These two trigger words were included during training. Use the exact spelling in your positive prompt, then describe the scene, pose, clothing and lighting you want.

#### seesee\_elf

**Trigger: seesee\_elf**

**EXAMPLE 01 · SEESEE\_ELF**

**Example prompt**

> eesee\_elf, A single female character stands leaning against a glowing pink vending machine in a neon-lit urban alleyway. The image is rendered in a high-quality, semi-realistic digital anime style with sharp focus and vibrant color grading. The framing is a full-body vertical shot, capturing the subject from head to toe with a slight low-angle perspective. She leans her left hip against the machine, her left hand tucked into her skirt pocket, while her right hand holds a pink beverage can at chest level. Her gaze is directed forward, looking slightly past the camera. She has long, straight platinum blonde hair that falls past her waist, pointed elf ears, and large amber eyes. Her face is pale with a neutral, calm expression. Her body is slender with long legs covered in sheer black pantyhose. She wears a cropped pink blazer over a black bandeau top, a matching high-waisted pink mini skirt, and black platform high-heeled shoes with ankle straps. A black lace choker and a silver necklace with a cross pendant adorn her neck, and long black earrings dangle from her ears. The background features a wet, reflective cobblestone street and blurred neon signs, including a vertical pink sign reading "OPEN" and a purple sign reading "BAR". The vending machine displays Japanese text "ドリンク" at the top and contains rows of colorful bottles. The lighting is dominated by the intense pink glow from the vending machine, which casts a strong magenta hue over the scene and creates sharp reflections on the wet ground.

#### seesee\_kitsune

**Trigger: seesee\_kitsune**

*The original example prompt and image for this trigger are not included on this page.*

#### My pixel-art look

V2 includes my own pixel look. You can ask for it in different ways; the prompt should contain **pixel** or **16Bit**. A simple **pixel art** cue is a useful starting point. Combine the style cue with a normal description of the character or scene.

**Style cues to try:** pixel art · a detailed pixel illustration · 16Bit pixel art

### Pixel look · Example 1

**EXAMPLE 03 · PIXEL-ART**

**Example prompt**

> A solitary figure sits in a command chair facing a vast spaceship viewport, rendered in pixel art with a visible grid of square pixels, a limited palette of neon blues, magentas, and reds, and hard-edged blocky shapes. The camera angle is a symmetrical, eye-level rear view from behind the pilot, looking forward into the cockpit. The figure is a dark silhouette with a rounded head, seated centrally and facing away from the viewer toward the controls. The cockpit interior is dense with angular consoles on the left and right, featuring small glowing blue screens and clusters of red, orange, and green square buttons. The large, multi-paned window reveals outer space, dominated by a large blue planet with bright rings on the left and a swirling nebula of pink, purple, and orange on the right. A small, cyan wireframe sphere floats in the nebula. The floor is dark blue with red rectangular lights, and thick, segmented pipes run along the bottom corners of the frame.

![example03_pixel_cockpit](https://huggingface.co/SeeSee21/See-Krea2-Genesis/resolve/main/images/inline/example03_pixel_cockpit.jpg)

#### Game Boy-inspired scenes

I also included my own Game Boy captures in the dataset. To ask for a retro handheld-game scene, try a description like this:

> **a retro 8-bit video game level, rendered in Game Boy Color era pixel art with a visible coarse pixel grid**

Use that as a style direction, then describe the level, setting and objects you want to see.

**EXAMPLE 05 · GAME BOY-STYLE**

**Example prompt**

> [A blue cat character stands on a grassy ledge in a retro 8-bit video game level, rendered in Game Boy Color era pixel art with a visible coarse pixel grid, small resolution, and limited color palette. The framing is a horizontal side-scrolling view showing a light blue sky with white clouds, a distant green mountain range, and a tan sandy ground area. The blue cat faces forward with a simple closed-eye expression, positioned near small orange flowers and green bushes on the right. Floating above are two yellow, round creatures with green accents, one resting on a small floating platform and another hovering in the upper right corner. A brown wooden crate sits on the far left edge, and a blue water hazard spans the bottom left. The bottom of the screen displays a black HUD bar with white pixelated text reading "WORLD 1", a counter showing "x4", and a timer reading "T:79".

![example05_gameboy](https://huggingface.co/SeeSee21/See-Krea2-Genesis/resolve/main/images/inline/example05_gameboy.jpg)

</details>

<details>
<summary><b>🖼️ Example prompts</b></summary>

Five prompts covering very different territory. Each one is written the way the model likes to be talked to — use them as templates and swap the content.

### 1 · Dark sci-fi portrait

![example_1_dark_scifi_portrait](https://huggingface.co/SeeSee21/See-Krea2-Genesis/resolve/main/images/inline/example_1_dark_scifi_portrait.jpg)

```
A striking dark sci-fi anime portrait shows a cybernetic girl against a fractured
digital void. The illustration uses high-contrast monochrome tones with glowing cyan
glitch accents, precise line art and a cold surreal aesthetic. A tight close-up frames
her head and shoulders from a slightly low angle. She holds perfectly still, chin
lifted, staring past the viewer. Her short white hair is cut in a sharp asymmetric bob
with one side shaved, and thin luminous seams trace her cheekbone and temple. Her eyes
are solid cyan with no visible pupil. She wears a matte black high-collared bodysuit
with exposed cabling along the neck. Fragments of broken screen geometry drift behind
her. Rim lighting picks out the edge of her jaw against the darkness.
```

### 2 · Neon graphic full-body

![example_2_neon_graphic](https://huggingface.co/SeeSee21/See-Krea2-Genesis/resolve/main/images/inline/example_2_neon_graphic.jpg)

```
A stylised graphic anime artwork depicts a mischievous spirit girl rendered entirely
in liquid chrome, outlined in electric magenta. The full-body composition places her
mid-leap above a glowing hexagonal sigil, one arm thrown back and the other reaching
forward with fingers splayed, grinning with her eyes squeezed shut. Her form is a
glossy silver silhouette with flowing hair that breaks apart into floating droplets.
High-contrast flat color, thick neon outlines and hard shadow shapes define the style.
The background is deep violet with radiating light streaks. Magenta reflections pool
beneath her on a dark mirrored floor.
```

### 3 · Fantasy scene with heavy background

![example_3_fantasy_temple](https://huggingface.co/SeeSee21/See-Krea2-Genesis/resolve/main/images/inline/example_3_fantasy_temple.jpg)

```
An anime illustration shows a sorceress standing at the centre of a flooded jungle
temple, framed by enormous carved stone faces half-swallowed by roots. A wide full-body
shot from slightly below places her small against the architecture. She stands ankle
deep in still water, one arm outstretched with palm turned upward, looking calmly
toward the viewer. Her long moss-green hair falls past her waist and her eyes glow
pale gold. She wears layered ivory robes bound with braided cord, a heavy bronze
pectoral collar and bare feet. Vines and hanging orchids drape the ruins behind her.
Shafts of warm daylight break through the canopy and scatter across the water surface.
```

### 4 · Slice of life

![example_4_slice_of_life](https://huggingface.co/SeeSee21/See-Krea2-Genesis/resolve/main/images/inline/example_4_slice_of_life.jpg)

```
A soft anime illustration of a young florist arranging a bouquet on a wooden workbench
in a small sunlit shop. A relaxed medium shot from across the counter catches her
mid-motion, holding a stem of white ranunculus in one hand and secateurs in the other,
head tilted in concentration. Her wavy honey-blonde hair is pinned up loosely with
a few strands escaping, and she has warm grey eyes and a light scattering of freckles.
She wears a soft blue linen shirt with the sleeves rolled to the elbow under a canvas
apron marked with plant stains. Buckets of flowers, brown paper and twine crowd the
bench around her. Late afternoon light comes through the shop window and warms the
whole scene.
```

### 5 · Bright character portrait

![example_5_bright_portrait](https://huggingface.co/SeeSee21/See-Krea2-Genesis/resolve/main/images/inline/example_5_bright_portrait.jpg)

```
A cheerful anime adventurer poses in a crisp modern digital portrait with vibrant
colors and soft skin shading. A waist-up framing angled slightly from the side catches
her turning toward the camera, one hand raised in a small wave, the other steadying
a satchel strap on her shoulder. Her round face shows flushed cheeks, bright teal eyes
and a wide open smile. Her dark auburn hair is tied into two short braids with orange
ribbons. She wears a rust-colored travelling cloak over a cream tunic, leather bracers
and a compass on a cord around her neck. Autumn woodland blurs softly behind her.
Golden late-day light catches the loose strands of her hair.
```

</details>

<details>
<summary><b>📦 Which build should I download? (details and weight error)</b></summary>

### ⭐ Genesis-int8\_convrot

**Size:** 13.2 GB  
**Weight error:** 0.99%

**Recommended for:** Best overall quality.

Works on any GPU that supports INT8, including **RTX 30-series and newer**.

➡️ **Choose this version if you are unsure which one to download.**

---

### Genesis-fp8\_scaled

**Size:** 13.1 GB  
**Weight error:** 2.66%

**Recommended for:** RTX 40-series (Ada) and newer GPUs.

Uses the native **FP8 path** and is a good choice for GPUs with strong FP8 support.

---

### Genesis-int8

**Size:** 13.2 GB  
**Weight error:** 1.55%

**Recommended for:** Users who want INT8 with slightly faster runtime performance.

Similar size and quality to `Genesis-int8_convrot`, but optimized more toward inference speed.

---

### Genesis-nvfp4

**Size:** 7.7 GB  
**Weight error:** 9.41%

**Recommended for:** RTX 50-series / Blackwell GPUs.

Uses native **NVFP4** and requires significantly less storage and VRAM than the INT8/FP8 versions.

➡️ A great option if you have a Blackwell GPU and want a much smaller model.

---

### Genesis-int4\_convrot

**Size:** 6.9 GB  
**Weight error:** 16.43%

**Recommended for:** Low-VRAM systems or users who want the smallest possible download.

This is the **smallest Genesis build**, but it also has the highest quantization error.

➡️ Choose this version when VRAM or disk space matters more than maximum quality.

#### How to read the "weight error"

**Lower is better.**

The weight error indicates how much the quantized model weights differ from the original model.

* **0.99%** → Very close to the original weights
* **1–3%** → Excellent quality / very small difference
* **~9%** → More aggressive compression
* **~16%** → Significant compression, mainly intended to save VRAM and disk space

A higher weight error does **not automatically mean that generated images will be worse by the same percentage**. It only describes the difference between the original and quantized model weights.

### Quick recommendation

⭐ **Best quality:** `Genesis-int8_convrot`  
⚡ **RTX 40-series:** `Genesis-fp8_scaled`  
🚀 **Faster INT8:** `Genesis-int8`  
💾 **RTX 50-series / smaller model:** `Genesis-nvfp4`  
🪶 **Smallest possible version:** `Genesis-int4_convrot`

### ⚠️ Read the "weight error" column correctly

That number is the measured deviation of the **weights** from the BF16 master, not of the images. The two are only loosely related.

In side-by-side testing all five builds look good, and the differences are small enough that you generally have to put two images next to each other to spot them. So **do not rule out the small builds because of the number** — if 6.9 GB is what fits your card, try it before assuming it is a downgrade.

What the numbers are actually good for is ranking the formats against each other, and explaining why `int8_convrot` beats `fp8_scaled` at the same file size: FP8 E4M3 has three mantissa bits, INT8 has 256 evenly spaced steps, and the ConvRot rotation spreads outliers before quantising instead of letting one large value eat the scale.

</details>

<details>
<summary><b>🔧 Installation (full steps)</b></summary>

Use the standard Krea 2 three-file setup. Download one diffusion-model variant and place the files in these folders:

**Diffusion model**  
ComfyUI/models/diffusion\_models/See-Krea2-Genesisv2-[format].safetensors

**Text encoder**  
ComfyUI/models/text\_encoders/qwen3vl\_4b\_bf16.safetensors

**VAE**  
ComfyUI/models/vae/qwen\_image\_vae.safetensors

For the BF16 master, the filename is **See-Krea2-Genesisv2.safetensors**, without a format suffix.

1. **Load Diffusion Model:** choose your downloaded Genesis V2 file.
2. **Load CLIP:** choose qwen3vl\_4b\_bf16.safetensors, type krea2.
3. **Load VAE:** choose qwen\_image\_vae.safetensors.

If you already run Krea 2 Turbo, reuse your existing text encoder and VAE. For quantized files, make sure your installed loader supports the selected format. Existing Krea 2 LoRAs can be tested with this checkpoint; their visual effect and compatibility still depend on the individual LoRA and workflow.

</details>

<details>
<summary><b>📦 Genesis-V1 (BF16)</b></summary>

Initial Genesis release. Civitai: [Genesis-BF16](https://civitai.red/models/2846618?modelVersionId=3214105)

| File | Format | ComfyUI folder |
|:--|:--|:--|
| [`See-Krea2-Genesis.safetensors`](https://huggingface.co/SeeSee21/See-Krea2-Genesis/resolve/main/Genesis-V1/See-Krea2-Genesis.safetensors) | BF16 (26.28 GB) | `models/diffusion_models/` |

**⚙️ Settings (from the original description):** see [⚙️ Recommended Settings](#⚙️-settings-and-prompting) — starting point 8 steps · CFG 1.0 · euler / simple. Use with the Qwen3-VL text encoder and the Qwen-Image VAE below.

<details>
<summary>📄 Genesis-BF16 (V1) · full Civitai version description</summary>

🎨 See-Krea2-Genesis | Personal Anime Checkpoint on Krea 2 Turbo

**8-Step Turbo • Clean Anime Aesthetics • Natural Language Prompting • Qwen3-VL Text Encoder**

*Diffusion Model • 1–2 MP Native • LoRA-friendly • ComfyUI-ready*

✨ What is See-Krea2-Genesis?

See-Krea2-Genesis is my personal anime checkpoint built on **Krea 2 Turbo**. The goal was simple: keep everything that makes Krea 2 Turbo good — fast 8-step generation, strong prompt comprehension, clean rendering — and push the default look firmly toward anime illustration.

What you get is a model that produces anime art as its natural output instead of needing to be argued into it, while the 8-step turbo behavior stays fully intact.

See-Krea2-Genesis is mainly intended for:

* Anime-style illustrations
* Character-focused images
* Detailed outfits, poses and scene composition
* Fast iteration — 8 steps means you can explore ideas quickly
* Style experiments inside ComfyUI

🎯 Key Features

* ✅ Anime-focused checkpoint on the **Krea 2 Turbo** base
* ✅ **8 steps at CFG 1.0** — full images in seconds
* ✅ Turbo distillation fully preserved, no step-count regression
* ✅ Strong natural-language prompt understanding through the Qwen3-VL text encoder
* ✅ Handles long, detailed scene descriptions — pose, camera angle and lighting all respond
* ✅ Clean line art and painted texture without heavy post-processing
* ✅ Works with the standard Krea 2 ComfyUI workflow
* ✅ Drop-in replacement for `krea2_turbo_bf16.safetensors`
* ✅ LoRA-friendly — same architecture, existing Krea 2 LoRAs stack normally
* ✅ BF16, no quantisation artefacts — quantised variants can be derived from this file

⚙️ Recommended Settings

The settings I personally use and recommend as a starting point:

```
Steps:       8
CFG:         1.0
Sampler:     euler
Scheduler:   simple
Denoise:     1.0
Resolution:  ~1 MP, for example 1024 × 1024
```

Sampler alternatives

* **euler + simple** — my default. Neutral, clean, predictable.
* **er\_sde + simple** — slightly more texture and character, works well at 10–12 steps.

⚠️ CFG must stay at 1.0

This is a turbo model. Raising CFG above 1 will burn the image — that is not a style choice, it breaks.

Two consequences people trip over:

* **Your negative prompt does nothing.** At CFG 1.0 the unconditional branch is never computed, so negative text has no mathematical effect on the result. Whatever you put there is ignored.
* **Exclusions must be phrased positively.** If you do not want gloves, write `bare hands` in the positive prompt. Writing `gloves` in the negative field changes nothing.

📐 Resolution Guide

| Use case | Resolution |
|:--|:--|
| ⭐ Square / general purpose | 1024 × 1024 |
| Portrait / character art | 896 × 1152 |
| Landscape / scenes | 1152 × 896 |
| Tall portrait | 1024 × 1536 |
| Widescreen | 1344 × 768 |

Around 1 MP gives the cleanest and most reliable results. The model holds up well up to roughly 1.5 MP. Beyond that, generate at 1 MP and upscale afterwards rather than pushing the native resolution.

💡 Prompting Guide

This model rewards **written description, not tag dumps**. The text encoder is a language model, so it actually reads sentences — and that is where most of the control lives.

Prompt structure that works

```
[medium and style] → [camera and framing] → [pose and action] →
[face and hair] → [body] → [clothing in detail] → [background] → [lighting]
```

You do not need to label the sections. Just write them in that order as flowing sentences.

✅ Do

* Write full sentences. 80–200 words is the sweet spot.
* Name the camera angle explicitly — *low-angle shot*, *close-up portrait*, *full-body view from behind*.
* Describe both hands separately if the pose matters. *"one hand raised to her forehead, the other resting on her waist"* lands far more reliably than *"posing"*.
* Describe clothing in layers and materials, not just as a name.
* Watch your pronouns. This encoder reads context — a stray `his` in a single-character prompt can summon a second person.

❌ Avoid

* Short tag dumps like `anime girl, red hair, desert`. Tags still work, but you lose almost all control over pose and camera.
* Weight syntax such as `(red hair:1.3)`. Qwen3-VL encodes tokens in context, so scaling embeddings distorts the meaning rather than emphasising it. If a detail is not showing up, describe it more specifically instead of weighting it.
* Filling the negative prompt. See the CFG note above — it is ignored.

🖼️ Example Prompts

Five prompts covering very different territory. Each one is written the way the model likes to be talked to — use them as templates and swap the content.

1 · Dark sci-fi portrait

```
A striking dark sci-fi anime portrait shows a cybernetic girl against a fractured
digital void. The illustration uses high-contrast monochrome tones with glowing cyan
glitch accents, precise line art and a cold surreal aesthetic. A tight close-up frames
her head and shoulders from a slightly low angle. She holds perfectly still, chin
lifted, staring past the viewer. Her short white hair is cut in a sharp asymmetric bob
with one side shaved, and thin luminous seams trace her cheekbone and temple. Her eyes
are solid cyan with no visible pupil. She wears a matte black high-collared bodysuit
with exposed cabling along the neck. Fragments of broken screen geometry drift behind
her. Rim lighting picks out the edge of her jaw against the darkness.
```

2 · Neon graphic full-body

```
A stylised graphic anime artwork depicts a mischievous spirit girl rendered entirely
in liquid chrome, outlined in electric magenta. The full-body composition places her
mid-leap above a glowing hexagonal sigil, one arm thrown back and the other reaching
forward with fingers splayed, grinning with her eyes squeezed shut. Her form is a
glossy silver silhouette with flowing hair that breaks apart into floating droplets.
High-contrast flat color, thick neon outlines and hard shadow shapes define the style.
The background is deep violet with radiating light streaks. Magenta reflections pool
beneath her on a dark mirrored floor.
```

3 · Fantasy scene with heavy background

```
An anime illustration shows a sorceress standing at the centre of a flooded jungle
temple, framed by enormous carved stone faces half-swallowed by roots. A wide full-body
shot from slightly below places her small against the architecture. She stands ankle
deep in still water, one arm outstretched with palm turned upward, looking calmly
toward the viewer. Her long moss-green hair falls past her waist and her eyes glow
pale gold. She wears layered ivory robes bound with braided cord, a heavy bronze
pectoral collar and bare feet. Vines and hanging orchids drape the ruins behind her.
Shafts of warm daylight break through the canopy and scatter across the water surface.
```

4 · Slice of life

```
A soft anime illustration of a young florist arranging a bouquet on a wooden workbench
in a small sunlit shop. A relaxed medium shot from across the counter catches her
mid-motion, holding a stem of white ranunculus in one hand and secateurs in the other,
head tilted in concentration. Her wavy honey-blonde hair is pinned up loosely with
a few strands escaping, and she has warm grey eyes and a light scattering of freckles.
She wears a soft blue linen shirt with the sleeves rolled to the elbow under a canvas
apron marked with plant stains. Buckets of flowers, brown paper and twine crowd the
bench around her. Late afternoon light comes through the shop window and warms the
whole scene.
```

5 · Bright character portrait

```
A cheerful anime adventurer poses in a crisp modern digital portrait with vibrant
colors and soft skin shading. A waist-up framing angled slightly from the side catches
her turning toward the camera, one hand raised in a small wave, the other steadying
a satchel strap on her shoulder. Her round face shows flushed cheeks, bright teal eyes
and a wide open smile. Her dark auburn hair is tied into two short braids with orange
ribbons. She wears a rust-colored travelling cloak over a cream tunic, leather bracers
and a compass on a cord around her neck. Autumn woodland blurs softly behind her.
Golden late-day light catches the loose strands of her hair.
```

ℹ️ *A section of the original version description about adult content is not reproduced here.*

🔧 Installation

This is a diffusion-model file, so it uses the standard Krea 2 three-file setup.

Place the files here:

```
ComfyUI/models/diffusion_models/See-Krea2-Genesis.safetensors
ComfyUI/models/text_encoders/qwen3vl_4b_bf16.safetensors
ComfyUI/models/vae/qwen_image_vae.safetensors
```

Then load them with the standard Krea 2 workflow:

* **Load Diffusion Model** → `See-Krea2-Genesis.safetensors`
* **Load CLIP** → `qwen3vl_4b_bf16.safetensors`, type `krea2`
* **Load VAE** → `qwen_image_vae.safetensors`

If you already run Krea 2 Turbo, you have the text encoder and VAE already — just drop in the checkpoint and point the loader at it.

📈 Version History

v1.0 — Initial Release

* Anime-focused checkpoint on Krea 2 Turbo
* 8-step generation at CFG 1.0, turbo behavior preserved
* BF16 diffusion model
* Compatible with the standard Krea 2 ComfyUI workflow

🙏 Credits

* **Base Model:** Krea 2 Turbo by **Krea**
* **Text Encoder:** Qwen3-VL-4B
* **VAE:** Qwen-Image VAE
* **Checkpoint:** SeeSee21

📜 License

* Commercial use — **yes**
* Merging and derivative models — **yes**
* Selling generated images — **yes**
* Credit — **appreciated**

Use it for whatever you like (within the Krea 2 Community License). If you merge it into something of your own, a mention of where the anime part came from is all I ask.

*See-Krea2-Genesis — a personal anime checkpoint on Krea 2 Turbo. 🎨*

</details>

### 🖼️ Genesis-BF16 gallery

All images below were posted to the **Genesis-BF16** version on Civitai.

<table>
<tr>
<td><img src="https://huggingface.co/SeeSee21/See-Krea2-Genesis/resolve/main/images/preview_6.jpg" alt="Genesis-BF16 preview 1"></td>
<td><img src="https://huggingface.co/SeeSee21/See-Krea2-Genesis/resolve/main/images/preview_7.jpg" alt="Genesis-BF16 preview 2"></td>
<td><img src="https://huggingface.co/SeeSee21/See-Krea2-Genesis/resolve/main/images/preview_8.jpg" alt="Genesis-BF16 preview 3"></td>
</tr>
<tr>
<td><img src="https://huggingface.co/SeeSee21/See-Krea2-Genesis/resolve/main/images/preview_9.jpg" alt="Genesis-BF16 preview 4"></td>
<td><img src="https://huggingface.co/SeeSee21/See-Krea2-Genesis/resolve/main/images/preview_10.jpg" alt="Genesis-BF16 preview 5"></td>
<td></td>
</tr>
</table>

</details>

<details>
<summary><b>🔧 Genesis-V1 compressed builds</b></summary>

Each compressed build of V1 is its own version on Civitai. *Their Civitai images are not used here (none passed the SFW selection), so these builds have no own gallery.*

### Genesis-int8_convrot

Civitai: [Genesis-int8_convrot](https://civitai.red/models/2846618?modelVersionId=3216118)

| File | Format | ComfyUI folder |
|:--|:--|:--|
| [`See-Krea2-Genesis_int8_convrot.safetensors`](https://huggingface.co/SeeSee21/See-Krea2-Genesis/resolve/main/Genesis-V1/See-Krea2-Genesis_int8_convrot.safetensors) | INT8 ConvRot | `models/diffusion_models/` |

<details>
<summary>📄 Genesis-int8_convrot · full Civitai version description</summary>

**INT8 with ConvRot rotation — 13.2 GB.**

The best quality-per-byte build, and the one I recommend by default.

Standard INT8 uses one scale per output channel, so a single outlier weight forces the scale up and everything else in that row loses resolution. ConvRot rotates the weights in groups of 256 with an orthogonal transform first, which spreads that outlier across the whole group. The rotation is undone on the activations at inference, so nothing is lost by it.

Measured against the BF16 master this lands at **0.99 % weight deviation** — 2.7× closer than the FP8 build at the same file size, and it stays that accurate across every layer instead of varying between 0.9 % and 3.1 % the way plain INT8 does.

**Runs on:** any GPU with a working INT8 path — RTX 30xx and newer.  
**Trade-off:** the activation rotation costs a little speed at runtime. If you want that back, take Genesis-int8.

Settings: 8 steps · CFG 1.0 · euler · simple · ~1 MP. Requires a ComfyUI recent enough to support `comfy_quant` INT8 with ConvRot.

</details>

### Genesis-fp8_scaled

Civitai: [Genesis-fp8_scaled](https://civitai.red/models/2846618?modelVersionId=3215973)

| File | Format | ComfyUI folder |
|:--|:--|:--|
| [`See-Krea2-Genesis_fp8_scaled.safetensors`](https://huggingface.co/SeeSee21/See-Krea2-Genesis/resolve/main/Genesis-V1/See-Krea2-Genesis_fp8_scaled.safetensors) | FP8 scaled | `models/diffusion_models/` |

<details>
<summary>📄 Genesis-fp8_scaled · full Civitai version description</summary>

**FP8 E4M3, per-tensor scaled — 13.1 GB.**

The build with the widest loader support, and the format Krea themselves ship.

Output-side projections (attention output, attention gate, MLP down) are flagged for full-precision matrix multiply, so error does not accumulate where it matters most; the input projections use the fast FP8 path. That split is not my invention — it is the policy read directly out of Krea's own official FP8 release.

Measured weight deviation: **2.66 %**. That is simply what FP8 E4M3 costs — three mantissa bits do not give more, and Krea's own FP8 build measures the same.

**Runs on:** RTX 40xx (Ada) and newer for the native FP8 path. Older cards can still load it with software dequantisation.  
**Pick this if:** you are on Ada, or if another build fails to load and you want the safest option.

Settings: 8 steps · CFG 1.0 · euler · simple · ~1 MP.

</details>

### Genesis-int8

Civitai: [Genesis-int8](https://civitai.red/models/2846618?modelVersionId=3216228)

| File | Format | ComfyUI folder |
|:--|:--|:--|
| [`See-Krea2-Genesis_int8.safetensors`](https://huggingface.co/SeeSee21/See-Krea2-Genesis/resolve/main/Genesis-V1/See-Krea2-Genesis_int8.safetensors) | INT8 | `models/diffusion_models/` |

<details>
<summary>📄 Genesis-int8 · full Civitai version description</summary>

**Plain INT8, per-channel scaled — 13.2 GB.** Same size as the ConvRot build, slightly faster, slightly less accurate.

This is Genesis-int8\_convrot without the rotation step. No activation rotation at inference means a little less work per matrix multiply.

The cost shows up unevenly: **1.55 % average weight deviation**, but it ranges from 0.9 % on well-behaved layers up to 3.1 % on the ones with strong outliers. ConvRot exists precisely to flatten that spread.

**Runs on:** any GPU with a working INT8 path.  
**Pick this if:** throughput matters more to you than the last bit of fidelity, or if your ComfyUI build does not handle ConvRot INT8. Otherwise take Genesis-int8\_convrot — same size, better numbers.

Settings: 8 steps · CFG 1.0 · euler · simple · ~1 MP.

</details>

### Genesis-nvfp4

Civitai: [Genesis-nvfp4](https://civitai.red/models/2846618?modelVersionId=3216329)

| File | Format | ComfyUI folder |
|:--|:--|:--|
| [`See-Krea2-Genesis_nvfp4.safetensors`](https://huggingface.co/SeeSee21/See-Krea2-Genesis/resolve/main/Genesis-V1/See-Krea2-Genesis_nvfp4.safetensors) | NVFP4 | `models/diffusion_models/` |

<details>
<summary>📄 Genesis-nvfp4 · full Civitai version description</summary>

**NVFP4 block microscaling, group size 16 — 7.7 GB.** Just over half the size of the INT8 and FP8 builds.

Four-bit weights with a per-block FP8 scale and a global scale on top. On Blackwell (RTX 50xx) this runs natively; on older cards it still loads, with the dequantisation done in software.

Measured weight deviation is **9.41 %** — much higher than the 8-bit builds, as you would expect from four bits. In practice the difference is far smaller than that number suggests: side by side it is visible, on its own it holds up well. If the size fits your card better, try it rather than dismissing it.

**Runs on:** RTX 50xx natively; older GPUs via software dequantisation.  
**Pick this if:** you are on Blackwell, or you want a meaningfully smaller file.

Settings: 8 steps · CFG 1.0 · euler · simple · ~1 MP.

</details>

### Genesis-int4_convrot

Civitai: [Genesis-int4_convrot](https://civitai.red/models/2846618?modelVersionId=3216389)

| File | Format | ComfyUI folder |
|:--|:--|:--|
| [`See-Krea2-Genesis_int4_convrot.safetensors`](https://huggingface.co/SeeSee21/See-Krea2-Genesis/resolve/main/Genesis-V1/See-Krea2-Genesis_int4_convrot.safetensors) | INT4 ConvRot | `models/diffusion_models/` |

<details>
<summary>📄 Genesis-int4_convrot · full Civitai version description</summary>

**INT4 W4A4 with ConvRot — 6.9 GB.** The smallest build, roughly a quarter of the BF16 master.

Four-bit weights and four-bit activations, with the same ConvRot rotation the INT8 build uses to keep outliers from dominating the scale.

Measured weight deviation is **16.43 %**, the highest of the five by a clear margin. Take that as a ranking, not as a verdict: the number describes the weights, not the pictures, and the two are only loosely related. This build still produces good anime output — the gap only really shows when you place it next to the 8-bit versions.

**Runs on:** broad support through ConvRot W4A4.  
**Pick this if:** VRAM or disk space is the binding constraint. If you have room for 13 GB, take Genesis-int8\_convrot instead.

Settings: 8 steps · CFG 1.0 · euler · simple · ~1 MP.

</details>

</details>

<details>
<summary><b>📈 Version history</b></summary>

### V2 · More training, more detail, more styles

* Longer training at higher resolution.
* More image detail and improved prompt following.
* Detailed anime women and detailed backgrounds in the training data.
* English and Japanese text, plus manga material; improved text rendering remains unconfirmed.
* New trained triggers: seesee\_elf and seesee\_kitsune.
* Personal pixel-art look, prompted with pixel or 16Bit.
* Game Boy-inspired direction trained with my own captures.
* BF16 master and five compressed formats: INT8 ConvRot, FP8 Scaled, NVFP4, INT4 ConvRot and MXFP8.

### V1 · Initial Genesis release

Anime-focused checkpoint on Krea 2 Turbo with an 8-step workflow at CFG 1.0. The original compressed builds included INT8 ConvRot, plain INT8, FP8 Scaled, NVFP4 and INT4 ConvRot. **Plain INT8 belongs to the V1 lineup; it is not one of the V2 files listed above.**

</details>

<details>
<summary><b>📁 Files in this repository</b></summary>

| Folder | Civitai version | Files |
|:--|:--|:--|
| `Genesis-V2/` | Genesis-V2 | `See-Krea2-Genesisv2.safetensors` (BF16 master), `-int8_convrot`, `-fp8_scaled`, `-nvfp4`, `-int4_convrot`, `-mxfp8` |
| `Genesis-V1/` | Genesis-BF16, Genesis-int8_convrot, Genesis-fp8_scaled, Genesis-int8, Genesis-nvfp4, Genesis-int4_convrot | `See-Krea2-Genesis.safetensors` (BF16), `_int8_convrot`, `_fp8_scaled`, `_int8`, `_nvfp4`, `_int4_convrot` |
| `Text-Encoder/` | Text Encoders | `qwen3vl_4b_bf16.safetensors`, `qwen3vl_4b_fp8_scaled.safetensors` (Qwen3-VL-4B) |
| `VAE/` | VAE | `qwen_image_vae.safetensors` (Qwen-Image VAE) |
| `workflows/` | See-Krea2-Workflows · Ultimate T2I v1.1 | `See-Krea2-Ultimate-v1.1.json` |

</details>

<details>
<summary><b>ℹ️ Content notice</b></summary>

This Hugging Face page only shows safe-for-work examples. Some sections of the original Civitai description are not reproduced here. Please follow the Krea 2 Community License and Acceptable Use Policy.

</details>

## 🙏 Credits

**Base model:** Krea 2 Turbo by Krea  
**Text encoder:** Qwen3-VL-4B  
**VAE:** Qwen-Image VAE  
**Checkpoint and additional training:** SeeSee

## 📜 License

This is a modified version of the Krea 2 model. It is not an official Krea product and is not endorsed by Krea.

Use is subject to the **Krea 2 Community License Agreement**. Please read the applicable terms at [Krea’s licensing page](https://krea.ai/krea-2-licensing), including any conditions that apply to derivatives and commercial use.

**My additional permissions, subject to the base-model license:**  
Commercial use: yes · Merging and derivatives: yes · Selling generated images: yes · Credit: appreciated.

These permissions do not override the base-model license. If Genesis contributes to your own model, a mention of where the anime part came from is appreciated.

**Made by SeeSee · Built on Krea 2 Turbo · Expanded in V2 🎨**

### Krea 2 Community License

This repository distributes a **Derivative** of Krea 2 Turbo under the **Krea 2 Community License Agreement v.1 (June 22, 2026)**. A complete copy is included as [LICENSE.md](LICENSE.md), and the incorporated [Acceptable Use Policy](ACCEPTABLE_USE_POLICY.md) is included too. By downloading or using these files you agree to be bound by these terms. They also apply to anyone you redistribute the files to.

- **Modification notice:** these weights have been modified from the original Krea 2 Turbo model (anime fine-tune and quantized variants by SeeSee21).
- **Not official:** this is not an official Krea product and is not endorsed by Krea.
- **Commercial use** is only permitted as the Krea 2 Community License allows (Section 2.3, revenue threshold). Otherwise you need an Enterprise License from Krea.
- **Content filtering (Section 4.2):** if you deploy this model, you must implement reasonable and appropriate content-filter measures. This repository ships weights and a workflow only: there is no hosted inference, the inference widget is disabled, and all preview images were reviewed by hand and are safe for work.
- Any additional permissions given by the author on Civitai apply **only within** the Krea 2 Community License; they do not override it.
- Text encoder (Qwen3-VL-4B) and VAE (Qwen-Image VAE) are Qwen components released under Apache-2.0.

---

<details>
<summary><b>🧭 More from SeeSee — all models</b></summary>

### 🧩 Checkpoints & AIO

| Model | Hugging Face | Civitai |
|:--|:--|:--|
| AniSee | [SeeSee21/AniSee](https://huggingface.co/SeeSee21/AniSee) | [Model](https://civitai.red/models/2628747) · [Workflows](https://civitai.red/models/2630318) |
| Z-Anime | [SeeSee21/Z-Anime](https://huggingface.co/SeeSee21/Z-Anime) | [Model](https://civitai.red/models/2483351) |
| Z-Image-Turbo-AIO | [SeeSee21/Z-Image-Turbo-AIO](https://huggingface.co/SeeSee21/Z-Image-Turbo-AIO) | [Model](https://civitai.red/models/2173571) · [Workflow](https://civitai.red/models/2174008) |
| ERNIE-AIO | [SeeSee21/ERNIE-AIO](https://huggingface.co/SeeSee21/ERNIE-AIO) | [Model](https://civitai.red/models/2551548) · [Workflow](https://civitai.red/models/2553748) |
| Chroma-Anime-AIO | [SeeSee21/Chroma-Anime-AIO](https://huggingface.co/SeeSee21/Chroma-Anime-AIO) | [Model](https://civitai.red/models/2022057) · [Workflow](https://civitai.red/models/2027641) |
| FLUX.2-klein-4B-AIO | [SeeSee21/FLUX.2-klein-4B-AIO](https://huggingface.co/SeeSee21/FLUX.2-klein-4B-AIO) | [Model](https://civitai.red/models/2327389) · [Workflow](https://civitai.red/models/2327746) |
| QWEN-Anime | [SeeSee21/QWEN-Anime](https://huggingface.co/SeeSee21/QWEN-Anime) | [Model](https://civitai.red/models/2122738) · [Workflow](https://civitai.red/models/2135240) |
| SeeMoss-Genesis | [SeeSee21/SeeMoss-Genesis](https://huggingface.co/SeeSee21/SeeMoss-Genesis) | [Model](https://civitai.red/models/3031179) · [Workflow](https://civitai.red/models/3031981) |
| See-Krea2 | [SeeSee21/See-Krea2-Genesis](https://huggingface.co/SeeSee21/See-Krea2-Genesis) | [Model](https://civitai.red/models/2846618) · [Workflow](https://civitai.red/models/2936142) |
| LTX See Motion | [SeeSee21/LTX-See-Motion](https://huggingface.co/SeeSee21/LTX-See-Motion) | [Model](https://civitai.red/models/2768763) · [Workflow](https://civitai.red/models/2784290) |

### 🎀 LoRAs

| Model | Hugging Face | Civitai |
|:--|:--|:--|
| Illustrious Anime Collection | [SeeSee21/Illustrious-Anime-Collection](https://huggingface.co/SeeSee21/Illustrious-Anime-Collection) | [Model](https://civitai.red/models/1994924) |
| Chroma Semi-Anime | [SeeSee21/Chroma-Semi-Anime](https://huggingface.co/SeeSee21/Chroma-Semi-Anime) | [Model](https://civitai.red/models/1992501) |
| Kitsune Furry Girl | [SeeSee21/Kitsune-Furry-Girl-Z-Image](https://huggingface.co/SeeSee21/Kitsune-Furry-Girl-Z-Image) | [Model](https://civitai.red/models/2218232) |
| Ayase Momo (Dandadan) | [SeeSee21/Ayase-Momo-Z-Image](https://huggingface.co/SeeSee21/Ayase-Momo-Z-Image) | [Model](https://civitai.red/models/2189797) |

</details>

<div align="center">

[🧩 Checkpoints & AIO](https://huggingface.co/collections/SeeSee21/checkpoints-and-aio) &nbsp;·&nbsp; [🎀 LoRAs](https://huggingface.co/collections/SeeSee21/loras) &nbsp;·&nbsp; [🌐 Civitai](https://civitai.red/user/SeeSeeLP) &nbsp;·&nbsp; [▶️ YouTube](https://www.youtube.com/@SeeSee-AI)

<sub>Made with ❤️ by SeeSee</sub>

</div>