# [Krea 2 DiT] Clementine (`zclclementine`) — Fictional AI Influencer & Consistent Character

> **Collector's Issue #014 (Mainstream Editorial Edition — `civitai.com`)** | Curated by *Maison Latentia*

> **Model Architecture:** `Krea 2 DiT` (22GB bf16 Cloud Trained) + `Dual-Stage Micro-Texture Wash`

>

> ✦ **100% Original Virtual AI Persona:**

> *Clementine is a wholly fictional digital character and original creation (OC) curated by Maison Latentia. Engineered as a virtual AI influencer archetype, this model contains zero real-person dataset likenesses and does not depict any living individual.*

## ⚡ 3-Second Copy-Paste Quickstart (Krea 2 Edition)

| Setting | Stage 1: Krea 2 RAW DiT (Likeness & Pose) | Stage 2: Micro-Texture Wash (Skin Polish) |

| :--- | :--- | :--- |

| **Model / Text Encoder** | `krea2_raw_bf16.safetensors` + `qwen3vl_4b_fp8_scaled` | Same |

| **VAE** | `qwen_image_vae.safetensors` | `qwen_image_vae.safetensors` |

| **LoRA Weight** | **`0.90`** (`zclclementine_krea2_best.safetensors`) | Active in chain |

| **Sampler / Scheduler** | `Euler` / `Normal` / `Steps: 28` / `CFG: 3.0` / `Shift: 5.0` | `ER-SDE` / `Beta` / `Steps: 8` / `CFG: 1.8` / **`Denoise: 0.22`** |

| **Native Canvas** | `896 x 1152` (3:4) or `832 x 1216` (9:16) | Native resolution |

### 📋 1-Click Copy-Paste Prompts

**1. Krea 2 — High-Fashion Editorial Close-Up (`896x1152`):**

```text

<lora:zclclementine_krea2_step_0800:0.9> Describe the image by detailing the color, shape, size, texture, quantity, text, spatial relationships of the objects and background: A candid 85mm editorial portrait of zclclementinewoman, (zclclementine:1.0), warm golden hour rim lighting catching natural skin pores and subtle subsurface scattering, expressive direct eye contact, shot on 85mm f/1.4 prime lens, shallow depth of field, raw analog film texture.

```

> 💡 **Drag & Drop Workflow Included:** Every gallery image on this page has **embedded ComfyUI + A1111 generation metadata** (click the `(i)` icon to remix!), and `zclclementine_krea2_workflow.json` is attached directly in the files tab.

---

## 👤 Meet Clementine (Virtual AI Influencer Lore & Profile)

Meet **Clementine** (`zclclementine`) — The quintessential coltish Manic Pixie Dream Girl fantasy infused with authentic 2000s scene-queen and pop-punk energy—endless willowy legs in signature extremely low-rise skinny jeans with high-riding visible thong straps or pleated micro skirts with torn fishnets and arm warmers, luminous cornflower-blue doe eyes, razor-cut blonde skater/scene hair with visible dark roots and bold neon multicolor bang highlights, dainty symmetric pink heart face tattoos, a cheeky silver tongue ring, and effortless vintage-indie charm. Cast through MuseForge's Phase 0 Craniofacial Landmark Pipeline and trained on **MOSS-Image Base 5B** using Fizgig v8.8.1's **Slide clean-end schedule** and **LoRA Isolation (DOP)**, she locks 10/10 multi-angle facial identity and raw 2K pore fidelity without bleeding onto bystanders.

* **Trigger Word:** `zclclementine` (Fused Class Token: `zclclementinewoman`)

* **Archetype:** Coltish Manic Pixie Dream Girl (Vintage Alt Indie Darling / 2000s Pop-Punk & Scene Hair, Multicolor Bangs & Face Tattoos)

* **Age & Cohort:** 20-23 years old (Born `2003-2006` — SSA Era-Anchored Name)

* **Origin & Role:** Brooklyn, NY / Silver Lake, Los Angeles, CA — Vintage Vinyl Record Specialist & Indie Film Cinematographer

### 🔍 Civitai.com Discovery & SEO Tags

```text

character, woman, photorealistic, coltish, manic pixie, manic pixie dream girl, mpdg, scene, emo, pop punk, skater, skater girl, scene hair, emo hair, blonde, dark roots, blue eyes, side-swept bangs, bang highlights, multicolor hair, razor cut, early 2000s, y2k, low rise, low rise jeans, low rise pants, skinny jeans, tight skinny jeans, visible thong, whale tail, thong straps, studded belt, micro skirt, miniskirt, fishnets, torn fishnets, arm fishnets, fishnet arm warmers, slender, tall, willowy, leggy, gamine, vintage, indie, alt, tattoos, face tattoo, heart tattoo, piercings, tongue ring, retro, freckles, consistent character, realistic, ai influencer, virtual influencer, digital model, original character, portrait, moss image, minimax h3, krea 2, krea2, refmod, workflow

```

---

## 🔬 Dual-Stage Rendering Architecture: Why Krea 2 Excels at Skin Realism

This LoRA was trained specifically on **Krea 2 RAW DiT** using **Fizgig v8.8.1 Automagic v3 + Slide schedule + DOP LoRA Isolation** on an NVIDIA A100-80GB GPU.

* **Stage 1 (DiT Base Generation):** Generates 3D craniofacial structure, facial bone architecture, eye color, and wardrobe using `Euler` at `Shift 5.0` and `CFG 3.0`.

* **Stage 2 (0.22 Micro-Texture Wash):** Passes the un-quantized latent to `ER-SDE` with a `Beta` scheduler at low denoise (`0.22`) to etch true pore-level micro-texture, fine subcutaneous peach fuzz, and realistic specular highlights without distorting character identity.

* **Empirically Benchmarked Prior to Release (Measured ONNX `buffalo_l` + `DINOv2-base`):** Verified at Weight `0.90` (`zclclementine_krea2_best.safetensors`, AutoV2 `22dd2d08eb`) with **`67.5%` ArcFace Facial Likeness** (peak `70.2%`, calibrated gate `>= 42.0%`), **`76.8%` DINOv2 Proportional Fidelity**, and **`86.5%` Laplacian Pore Micro-Contrast** across 13 real ComfyUI Krea 2 inference renders (`eval_report.json`).

---

## 🎯 Truth-in-Advertising Standard: Genuine Krea 2 Inference Proofs

Every preview image in this showcase gallery was generated by running inference on **Krea 2 with the trained LoRA active** (`zclclementine_krea2_best.safetensors` @ 0.90 weight), testing novel out-of-distribution prompts (novel wardrobe, complex lighting, crowd bystander isolation) to demonstrate true real-world generalization.

We **never** recycle input training dataset images as showcase gallery images. (The 30-shot training dataset is included in `dataset_reference_plates/` for creator inspection and audit).

---

## 🛠️ ComfyUI File Placement Guide

| File | Destination Directory in ComfyUI | Usage |

| :--- | :--- | :--- |

| `krea2_raw_bf16.safetensors` | `ComfyUI/models/diffusion_models/` | Base Krea 2 DiT model |

| `qwen3vl_4b_fp8_scaled.safetensors` | `ComfyUI/models/clip/` | Krea 2 Text Encoder (Qwen3-VL) |

| `qwen_image_vae.safetensors` | `ComfyUI/models/vae/` | Qwen-Image 16-channel VAE |

| `zclclementine_krea2_best.safetensors` | `ComfyUI/models/loras/` | Primary Character LoRA |

| `zclclementine_krea2_step_0800.safetensors` | `ComfyUI/models/loras/` | Step 800 Checkpoint LoRA |

| `zclclementine_krea2_workflow.json` | Drag directly onto ComfyUI canvas | Auto-configures Stage 1 + Stage 2 wash |

---

## 📸 Post Your Renders & Earn Buzz!

If you enjoy **Clementine**, please **leave a 👍 Like / Review and post your generations in the gallery below!** We actively tip **Buzz ⚡** to creative community renders.