README.md

---
model_name: jackieV1
model_type: textual_inversion
pipeline_tag: text-to-image
base_model: "unknown"
library_name: kohya-ss
license: other  # update later to match the base model license
tags:
  - textual-inversion
  - sdxl
  - embedding
  - identity
  - kohya-ss
language:
  - en
inference: false
pretty_name: "jackieV1 — SDXL Textual Inversion Embedding"
---

# jackieV1 — SDXL Textual Inversion Embedding

**Type:** Textual Inversion (TI)  
**Base Model:** SDXL 1.0 (`juggernautXL_ragnarokBy.safetensors`)  
**Token:** `jck1`  
**Author:** BonusLockSMith  

---

## 🧠 Overview

`jackieV1` is a **Textual Inversion identity embedding** trained on SDXL 1.0 using the  
`juggernautXL_ragnarokBy.safetensors` checkpoint.

The goal of this embedding is to **lock in a consistent Jackie identity** across different outfits, scenes, and lighting setups while letting background, pose, and styling change freely.

### Intended Use

- Portrait shots with a consistent face  
- Full-body character renders with different outfits and environments  
- Style / setting changes (studio, outdoor, cinematic, etc.) that keep the same person  

This is an **identity / character lock** tool, not a general style LoRA.

---

## 🔧 Training Details

**Pipeline:** `sdxl_train_textual_inversion.py` (kohya_ss / sd-scripts)  
**Method:** DreamBooth-style TI on SDXL  

- **Token string:** `jck1`  
- **Init word:** `woman`  
- **Vectors per token:** `8`  
- **Resolution:** `512 × 512`  
- **Steps:** `800` (3 epochs, 540 effective images with repeats)  
- **Batch size:** `2`  
- **Optimizer:** 8-bit AdamW (`AdamW8bit`)  
- **LR scheduler:** cosine  
- **Learning rate:** `5e-4`  
- **Mixed precision:** `fp16`  
- **Latent caching:** enabled  
- **Buckets:** enabled, `min_bucket_reso = 256`, `max_bucket_reso = 1024`  
- **VAE:** SDXL VAE loaded from the JuggernautXL checkpoint  
- **Regularization:** none (no reg images used for this v1)  

**Kohya output files:**

- `jackieV1-000001.safetensors` — checkpoint after epoch 1  
- `jackieV1-000002.safetensors` — checkpoint after epoch 2  
- `jackieV1.safetensors` — final embedding (epoch 3, 800 steps)  

---

## 📂 Dataset

**Count:** ~27 curated images  
**Captions:** Per-image `.txt` files (manual captions)

**Directory layout (kohya):**

```text
D:/AI-Hub/kohya_ss/datasets/jackie-ti/
├── img/
│   └── 20_woman jck1/
│       ├── jackie_0010_00001_.png
│       ├── jackie_0010_00001_.txt
│       ├── ...
├── log/
└── model/
    ├── jackieV1.safetensors
    ├── jackieV1-000001.safetensors
    ├── jackieV1-000002.safetensors
    └── config_textual_inversion-*.toml
Each caption contains both the class word and token (e.g., woman jck1 plus scene details).

🚀 Usage
1. Load as Textual Inversion (TI)
In your SDXL pipeline (A1111, ComfyUI, or another frontend):

Place jackieV1.safetensors in your embeddings / TI folder:

Example (A1111): …/embeddings/jackieV1.safetensors

Example (ComfyUI): ComfyUI/models/embeddings/jackieV1.safetensors

Restart / refresh your UI so the new embedding is detected.

2. Prompting
Use the token jck1 together with a suitable class word:

text
Copy code
masterpiece, best quality, portrait of woman jck1, soft studio lighting, 85mm lens, sdxl style
Examples:

cinematic close-up of woman jck1, dramatic rim light, bokeh background

full-body shot of woman jck1 in leather jacket, nighttime city street, neon lights

woman jck1, casual outfit, sitting on couch, natural daylight, living room

To change style, stack it with style prompts or LoRAs without removing the jck1 token.

⚠️ Limitations & Notes
Non-photorealistic extremes (very stylized anime, heavy abstraction) may weaken identity lock.

Very wide shots where the face is tiny in frame can lose detail.

Trained specifically on one subject; not intended for general “woman” or crowd scenes.

License and allowed use must match the underlying SDXL / JuggernautXL checkpoint.

Treat this embedding as subject-specific research / creative tool, not a generic commercial model, unless you’ve verified the base model’s terms.

🔄 Versioning
v1.0 (jackieV1) — initial SDXL textual inversion

27 images, 20 repeats, 800 steps

Target: strong identity lock with flexible styling

Planned future versions:

Possible v1.x updates with more varied poses / lighting

Separate LoRA for pose & outfit control if needed

📣 Contact & Attribution
Created by BonusLockSMith as part of the GritAI / WTFE local-generation stack.

If you use this embedding in a public project, a simple credit is appreciated:

“Jackie SDXL TI by BonusLockSMith (HuggingFace: BonusLockSMith/jackie-ti)”