Update README.md

README.md

@@ -1,156 +1,153 @@
+---
+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)  
+**Base Model:** SDXL 1.0 (`juggernautXL_ragnarokBy.safetensors`)  
 **Token:** `jck1`  
 **Author:** BonusLockSMith  
 
 ---
 
-## Overview
+## 🧠 Overview
 
-`jackieV1` is a **Textual Inversion embedding** trained on SDXL 1.0 using the
+`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.
+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
 
-Use case examples:
+- 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  
 
-- Portrait shots with a consistent face
-- Full-body character renders with different outfits
-- 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
+## 🔧 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`  
-
-**Dataset:**
-
-- ~27 curated images
-- Captioned with `.txt` files (per-image 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
-      ...
-Repeats: 20
-
-Effective train images with repeats: 540
-
-Hyperparameters (key ones):
-
-Resolution: 512x512, buckets enabled, bucket_no_upscale = true
-
-Epochs: 3
-
-Max train steps: 800
-
-Batch size: 2
-
-Optimizer: AdamW8bit
 
-Learning rate: 5e-4
+- **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)  
 
-LR scheduler: cosine
+---
 
-Loss type: l2
+## 📂 Dataset
 
-Mixed precision: fp16
+**Count:** ~27 curated images  
+**Captions:** Per-image `.txt` files (manual captions)
 
-Xformers: enabled
+**Directory layout (kohya):**
 
-No regularization images used
+```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).
 
-Checkpoints:
+🚀 Usage
+1. Load as Textual Inversion (TI)
+In your SDXL pipeline (A1111, ComfyUI, or another frontend):
 
-jackieV1-000001.safetensors (epoch 1)
+Place jackieV1.safetensors in your embeddings / TI folder:
 
-jackieV1-000002.safetensors (epoch 2)
+Example (A1111): …/embeddings/jackieV1.safetensors
 
-jackieV1.safetensors (final embedding)
+Example (ComfyUI): ComfyUI/models/embeddings/jackieV1.safetensors
 
-How to Use
-Automatic1111
-Download jackieV1.safetensors.
+Restart / refresh your UI so the new embedding is detected.
 
-Place it into:
+2. Prompting
+Use the token jck1 together with a suitable class word:
 
 text
 Copy code
-stable-diffusion-webui/embeddings/
-Restart the UI.
+masterpiece, best quality, portrait of woman jck1, soft studio lighting, 85mm lens, sdxl style
+Examples:
 
-Use the trigger token jck1 in your prompt.
+cinematic close-up of woman jck1, dramatic rim light, bokeh background
 
-Example prompt (portrait):
+full-body shot of woman jck1 in leather jacket, nighttime city street, neon lights
 
-text
-Copy code
-jck1, woman, portrait, upper body, neutral expression,
-soft studio lighting, detailed skin, sharp focus, 8k
-Example negative prompt:
+woman jck1, casual outfit, sitting on couch, natural daylight, living room
 
-text
-Copy code
-low quality, bad anatomy, extra limbs, distorted face, blurry, grainy
-Example prompt (full body, scene):
+To change style, stack it with style prompts or LoRAs without removing the jck1 token.
 
-text
-Copy code
-jck1, woman, full body, casual clothing, standing on a city sidewalk at sunset,
-cinematic lighting, shallow depth of field, realistic, detailed skin, 8k
-ComfyUI
-Copy jackieV1.safetensors to:
+⚠️ Limitations & Notes
+Non-photorealistic extremes (very stylized anime, heavy abstraction) may weaken identity lock.
 
-text
-Copy code
-ComfyUI/models/embeddings/jackieV1.safetensors
-Restart ComfyUI so it rescans models.
-
-In your CLIP text node, include jck1 in the positive prompt.
+Very wide shots where the face is tiny in frame can lose detail.
 
-Example ComfyUI positive prompt:
-
-text
-Copy code
-jck1, woman, portrait, upper body, soft studio lighting,
-cinematic, ultra detailed, SDXL, juggernaut style
-Pair with your usual negative prompt / sampler / CFG settings.
+Trained specifically on one subject; not intended for general “woman” or crowd scenes.
 
-Suggested Settings
-These are starting points only; tweak to your taste.
+License and allowed use must match the underlying SDXL / JuggernautXL checkpoint.
 
-Base model: juggernautXL_ragnarokBy.safetensors (SDXL)
+Treat this embedding as subject-specific research / creative tool, not a generic commercial model, unless you’ve verified the base model’s terms.
 
-Sampler: Euler a / DPM++ 2M Karras
+🔄 Versioning
+v1.0 (jackieV1) — initial SDXL textual inversion
 
-Steps: 25–35
+27 images, 20 repeats, 800 steps
 
-CFG scale: 5–8
+Target: strong identity lock with flexible styling
 
-Resolution: 768×1024 or 832×1216 (portrait) / similar area for landscape
+Planned future versions:
 
-For strong identity lock, keep jck1 toward the front of the prompt and
-avoid stacking too many style tokens that overpower it.
+Possible v1.x updates with more varied poses / lighting
 
-Notes, Rights, and Usage
-This embedding was trained on a private Jackie dataset.
+Separate LoRA for pose & outfit control if needed
 
-You are responsible for ensuring that your use of this embedding respects the
-rights and privacy of any real person whose likeness it may represent.
+📣 Contact & Attribution
+Created by BonusLockSMith as part of the GritAI / WTFE local-generation stack.
 
-Do not use this model for harassment, defamation, deepfake abuse, or any
-malicious activity.
+If you use this embedding in a public project, a simple credit is appreciated:
 
-If you create cool results or derivative work from this embedding, feel free to
-tag @BonusLockSmith.
+“Jackie SDXL TI by BonusLockSMith (HuggingFace: BonusLockSMith/jackie-ti)”