Honest card: 4-bit NF4 Z-Image build + usage + Studio link

README.md

@@ -5,54 +5,52 @@ base_model:
 pipeline_tag: text-to-image
 library_name: diffusers
 language:
-  - ko
   - en
-  - zh
-  - ja
-  - ar
-  - th
+  - ko
 tags:
   - text-to-image
   - image-generation
-  - korean
-  - multilingual
-  - hangul
-  - korean-text-rendering
-  - typography
-  - poster
+  - quantized
+  - bitsandbytes
+  - nf4
+  - 4-bit
   - on-device
+  - korean
   - pocket
   - vidraft
 ---
 
-# 🖼️ POCKET-Image-Zimage — Character-Perfect Text in Any Image
+# 🖼️ POCKET-Image-Zimage — 4-bit (NF4) Z-Image for on-device
 
-**POCKET-Image** is VIDRAFT's on-device text-in-image system. It places **character-perfect text in
-any language** — Korean · 中文 · 日本語 · العربية · ไทย · Latin and more — onto any scene you describe.
-Built on the open, commercial-friendly [Z-Image](https://huggingface.co/Tongyi-MAI/Z-Image) foundation
-and VIDRAFT's **POCKET-Core** engine.
+A **4-bit (NF4) quantized build of [Z-Image](https://huggingface.co/Tongyi-MAI/Z-Image)** (Apache-2.0),
+packaged by VIDRAFT for **low-VRAM, on-device** image generation — part of the POCKET line.
 
-> Where global models smear non-Latin scripts ("안녕하세요" → "안ㅐ기"), POCKET-Image renders every glyph exactly.
+- 📦 **~6 GB on disk** (transformer + text encoder in NF4, VAE in fp16)
+- ⚡ **Runs from ~8.6 GB VRAM** (≈**4.5 GB** with CPU offload) — vs **23.3 GB** for bf16
+- 🎯 ~2.7–5× smaller footprint, quality on par with the bf16 base
 
-## ✨ Highlights
-- ✅ **100% accurate text**, any language — RTL (Arabic) and complex shaping (Thai, Indic) handled
-- 🖼️ **Any background** from a prompt — text is optional (empty ⇒ pure image)
-- ⚡ **No GPU, no NPU** — runs on plain CPU + RAM
-- 🪶 **From 4.5 GB** footprint · 💻 Windows · macOS · Linux (RTX or Apple Silicon)
-- 🔒 **Fully local** — no cloud, no account
+## Usage
+```python
+import torch
+from diffusers import ZImagePipeline   # or ZImageImg2ImgPipeline / ZImageInpaintPipeline
 
-## 🎨 Try it
-- **Studio (generate here):** https://huggingface.co/spaces/FINAL-Bench/POCKET-Image-Studio
+pipe = ZImagePipeline.from_pretrained(
+    "FINAL-Bench/POCKET-Image-Zimage", torch_dtype=torch.bfloat16
+).to("cuda")
+img = pipe("a serene mountain lake at sunrise, photorealistic", num_inference_steps=20).images[0]
+img.save("out.png")
+```
+Requires **`bitsandbytes`** (CUDA). Measured reload + generate peak: ~10.9 GB VRAM.
+For Apple Silicon / CPU, an `optimum-quanto` int8 build (~13.4 GB) is the portable option.
 
-## 💻 On-device footprint (measured)
-| Machine | Peak memory |
-|---|---|
-| RTX 3050 / 4060 | **8.6 GB** |
-| RTX 2060 (6 GB, offloaded) | **4.5 GB** |
-| MacBook (M1 Pro–M4, 16 GB+) | **13.4 GB** |
+## 🎨 The full POCKET-Image system
+This repo hosts the **quantized base model** only. The headline **character-perfect Korean &
+multilingual text** feature is delivered by the POCKET-Image *pipeline*, not by these weights alone.
+Try the full system here:
 
-Base model: [Tongyi-MAI/Z-Image](https://huggingface.co/Tongyi-MAI/Z-Image) (Apache-2.0). Engine: VIDRAFT POCKET-Core.
+- **🎨 Studio (generate here):** https://huggingface.co/spaces/FINAL-Bench/POCKET-Image-Studio
 
+Base model: [Tongyi-MAI/Z-Image](https://huggingface.co/Tongyi-MAI/Z-Image) (Apache-2.0) · Quantization: bitsandbytes NF4 · By VIDRAFT.
 
 <!-- POCKET-FAMILY -->
 
@@ -68,7 +66,7 @@ Base model: [Tongyi-MAI/Z-Image](https://huggingface.co/Tongyi-MAI/Z-Image) (Apa
 - 🇰🇷 [POCKET-KR-GGUF](https://huggingface.co/FINAL-Bench/POCKET-KR-GGUF) — Korean, Android
 - 🍎 [POCKET-KR-MLX](https://huggingface.co/FINAL-Bench/POCKET-KR-MLX) — Korean, iPhone / Mac
 - 🌍 [POCKET-EN-GGUF](https://huggingface.co/FINAL-Bench/POCKET-EN-GGUF) — English, phone / PC
-- 🖼️ [POCKET-Image-Zimage](https://huggingface.co/FINAL-Bench/POCKET-Image-Zimage) — character-perfect text in any image
+- 🖼️ [POCKET-Image-Zimage](https://huggingface.co/FINAL-Bench/POCKET-Image-Zimage) — 4-bit Z-Image (this repo)
 
 **Demos & tools (Spaces)**
 - 🎨 [POCKET-Image Studio](https://huggingface.co/spaces/FINAL-Bench/POCKET-Image-Studio) — text-in-image, generate in-page