Rewrite model card with showcase images and real energy

README.md

@@ -17,59 +17,53 @@ pipeline_tag: text-to-image
 
 # 4CGT — 四彫画匠 (Shichō Gashō)
 
-**Master Artisan of Sculpted Art** — quality-engineered anime image generation with cross-platform determinism, DoRA anatomical accuracy, and integrated 2D→3D conversion.
+Anime image generation that nails anatomy, follows your prompt, and gives you the same pixel-perfect result on any machine. Every time.
 
-**GitHub**: [teenu/4cgt](https://github.com/teenu/4cgt)
-**Website**: [4cgt.com](https://4cgt.com)
+<p align="center">
+<img src="https://huggingface.co/epigene/4cgt/resolve/main/showcase/showcase_1.png" width="24%" />
+<img src="https://huggingface.co/epigene/4cgt/resolve/main/showcase/showcase_2.png" width="24%" />
+<img src="https://huggingface.co/epigene/4cgt/resolve/main/showcase/showcase_3.png" width="24%" />
+<img src="https://huggingface.co/epigene/4cgt/resolve/main/showcase/showcase_4.png" width="24%" />
+</p>
+<p align="center"><sub>Seed 7777, 4242, 1337, 3939. DoRA v0.271. ~25s each on RTX 5090.</sub></p>
 
-## Assets
+**GitHub**: [teenu/4cgt](https://github.com/teenu/4cgt) | **Website**: [4cgt.com](https://4cgt.com) | **Sponsor**: [GitHub Sponsors](https://github.com/sponsors/teenu)
 
-This repo contains all assets needed to run 4CGT:
+## Assets
 
-| File | Size | Required? |
-|------|------|-----------|
+| File | Size | Required |
+|------|------|----------|
 | `NoobAI-XL-Vpred-v1.0.safetensors` | 6.7 GB | Yes |
 | `dora/noobai_vp10_stabilizer_v0.271_fp16.safetensors` | 44 MB | Yes (recommended) |
 | `dora/noobai_vp10_stabilizer_v0.280a_fp16.safetensors` | 64 MB | Yes |
 | `style/*.csv` | 109 MB | Yes |
 | `controlnet/openpose_fp32.safetensors` | 4.7 GB | Optional (pose control) |
-| `poses/*.png` | 140 KB | Optional (example poses) |
+| `poses/*.png` | 140 KB | Optional (examples) |
 
-## Quick Start
+## Quick start
 
 ```bash
-# 1. Clone the tool
-git clone https://github.com/teenu/4cgt.git
-cd 4cgt
-
-# 2. Install PyTorch (pick your GPU)
+git clone https://github.com/teenu/4cgt.git && cd 4cgt
 pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu124
-
-# 3. Install dependencies
 pip install -r requirements.txt
 
-# 4. Download assets from this repo
+# Download everything from this repo
 pip install huggingface-hub
 huggingface-cli download epigene/4cgt --local-dir assets
+mv assets/NoobAI-XL-Vpred-v1.0.safetensors . && mv assets/dora/* dora/ && mv assets/style/* style/
+rm -rf assets
 
-# 5. Move assets into place
-mv assets/NoobAI-XL-Vpred-v1.0.safetensors .
-mv assets/dora/* dora/
-mv assets/controlnet/* controlnet/
-mv assets/style/* style/
-mv assets/poses/* poses/
-
-# 6. Run
 python main.py
 ```
 
-## Four Pillars
+## What makes this different
 
-- **Anatomical Accuracy** — DoRA weight-decomposed adapters optimized for v-prediction
-- **Prompt Fidelity** — Long prompt support up to 600 tokens with A1111-compatible syntax
-- **Image Quality** — Lossless FP32 VAE, BF16/FP32 only, FP16 rejected by design
-- **Deterministic Output** — Same seed = identical image hash across macOS, Windows, Linux
+- **DoRA anatomical stabilizers** eliminate the melted-face lottery. Characters have correct proportions, every time.
+- **600-token prompts** with A1111 weight syntax. Standard CLIP cuts off at 77. 4CGT doesn't.
+- **Same seed = same hash.** Byte-identical output across macOS, Windows, Linux. Not "similar" — identical.
+- **2D to 3D in <1 second.** Integrated Apple Sharp pipeline. Generate, convert to Gaussian Splat, render flythrough video.
+- **Lossless pipeline.** FP16 is rejected on load. BF16/FP32 only. FP32 VAE always.
 
 ## License
 
-AGPL-3.0. See [GitHub](https://github.com/teenu/4cgt) for full license.
+AGPL-3.0. See [GitHub](https://github.com/teenu/4cgt) for details.