Add multilingual documentation (EN/KO/ZH/JA) and custom model quantization request channels

README.md

@@ -1,5 +1,10 @@
 ---
 license: other
+language:
+  - en
+  - ko
+  - zh
+  - ja
 base_model:
   - Comfy-Org/MiniMax-H3
 pipeline_tag: image-to-video
@@ -16,17 +21,30 @@ tags:
   - rung-omega
   - theoremflow-omega
   - comfyui
+  - multilingual
 ---
 
 # MiniMax-H3-Rung-Omega (`5.976 BPW`, `14.98 GiB` Joint-tANS / `16.22 GiB` Raw Base-243) + TheoremFlow-Ω (`8.92x` Speedup)
 
+<p align="left">
+  <strong>🌐 Language / 언어 / 语言 / 言語:</strong>&nbsp;&nbsp;
+  <a href="./README.md"><strong>🇺🇸 English</strong></a> &nbsp;|&nbsp;
+  <a href="./README_ko.md"><strong>🇰🇷 한국어</strong></a> &nbsp;|&nbsp;
+  <a href="./README_zh.md"><strong>🇨🇳 中文 (简体)</strong></a> &nbsp;|&nbsp;
+  <a href="./README_ja.md"><strong>🇯🇵 日本語</strong></a>
+</p>
+
 <p align="left">
   <a href="https://buymeacoffee.com/tripleyoung" target="_blank">
     <img src="https://cdn.buymeacoffee.com/buttons/v2/default-yellow.png" alt="Buy Me A Coffee" height="48" />
   </a>
 </p>
 
-> ☕ **Support Open Quantization & Inference Acceleration Research**: If **`MiniMax-H3-Rung-Omega`** or **`TheoremFlow-Ω`** saves you GPU memory and generation time, consider supporting our independent AI research on [**Buy Me a Coffee (`buymeacoffee.com/tripleyoung`)**](https://buymeacoffee.com/tripleyoung)!
+> ☕ **Support Open Quantization & Inference Acceleration Research & Request Next Models**:
+> * 🇺🇸 **Want another model quantized with `Rung Ω` (`4-Trit Base-243`) + `TheoremFlow-Ω` acceleration?** Leave your model request in a [**Buy Me a Coffee support note (`buymeacoffee.com/tripleyoung`)**](https://buymeacoffee.com/tripleyoung) (Priority Queue ⚡) or open a thread in our [**Hugging Face Community Discussions**](https://huggingface.co/yeowon/MiniMax-H3-Rung-Omega/discussions)!
+> * 🇰🇷 **원하시는 다른 모델(비디오·오디오 DiT, LLM 등)의 `Rung Ω` 양자화 및 `TheoremFlow-Ω` 추론 가속이 필요하신가요?** [**Buy Me a Coffee 후원 메시지(`buymeacoffee.com/tripleyoung`)**](https://buymeacoffee.com/tripleyoung)에 모델명을 남겨주시면 **최우선 순위(Priority Queue ⚡)**로 양자화해 드립니다! ([**HF 커뮤니티 탭**](https://huggingface.co/yeowon/MiniMax-H3-Rung-Omega/discussions)에서도 요청 및 소통 가능 · [한국어 전체 문서](./README_ko.md))
+> * 🇨🇳 **希望我们将其他模型（视频/音频 DiT、LLM 等）进行 `Rung Ω` 量化与 `TheoremFlow-Ω` 加速吗？** 欢迎在 [**Buy Me a Coffee 赞助留言 (`buymeacoffee.com/tripleyoung`)**](https://buymeacoffee.com/tripleyoung) 中写下您需要的模型名称（**赞助请求享最优先队列 ⚡**），或在 [**Hugging Face 社区讨论区 (Discussions)**](https://huggingface.co/yeowon/MiniMax-H3-Rung-Omega/discussions) 留言交流！([查看中文完整文档](./README_zh.md))
+> * 🇯🇵 **他のモデル（動画・音声 DiT、LLM など）の `Rung Ω` 量子化や `TheoremFlow-Ω` 高速化をご希望ですか？** [**Buy Me a Coffee のサポートメッセージ (`buymeacoffee.com/tripleyoung`)**](https://buymeacoffee.com/tripleyoung) にご希望のモデル名をご記入いただければ、**最優先キュー (Priority Queue ⚡)** で量子化いたします！（[**HF コミュニティ Discussions**](https://huggingface.co/yeowon/MiniMax-H3-Rung-Omega/discussions) でもリクエスト受付中・[日本語ドキュメント全文](./README_ja.md)）
 
 **`yeowon/MiniMax-H3-Rung-Omega`** provides the **50-Block Pure `Rung Ω` Balanced Base-3 (`4-Trit`, Base-243 5-in-4-byte)** quantized checkpoint and the **`TheoremFlow-Ω` Mathematical Floor Inference Engine** for the **MiniMax-H3 FL2VA (First/Last-Frame-to-Video+Audio, 33.1B Total / 20.11B Active Parameters)** multimodal Diffusion Transformer.
 
@@ -169,11 +187,14 @@ tf_state = install_theoremflow_omega_schedule(
 
 ---
 
-## ☕ Support Our Research
+## ☕ Support Our Research & Request Custom Model Quantization
 
 We release **`MiniMax-H3-Rung-Omega`** and **`TheoremFlow-Ω`** openly with zero gated access barriers so the community can run high-fidelity multimodal Video+Audio generation faster and in smaller VRAM footprints.
 
-If this model or inference engine helps your workflow or research, please consider buying us a coffee!
+### 🎯 Want Another Model Quantized? (`Rung Ω` + `TheoremFlow-Ω`)
+Have a large Video/Audio Diffusion Transformer (Wan, HunyuanVideo, LTX, Sora-class DiT) or LLM you want compressed to **`Rung Ω` (`3-Trit` / `4-Trit` / `5-Trit` Base-243)** with **`TheoremFlow-Ω` acceleration**?
+1. **⚡ Priority Supporter Queue**: Leave the target model name in your [**Buy Me a Coffee support note (`buymeacoffee.com/tripleyoung`)**](https://buymeacoffee.com/tripleyoung) — supporter requests are prioritized first!
+2. **💬 Community Channel & Model Requests**: Open a discussion thread anytime in the [**Hugging Face Community Tab (`Discussions`)**](https://huggingface.co/yeowon/MiniMax-H3-Rung-Omega/discussions) to request models, ask questions, or talk directly with us.
 
 <p align="left">
   <a href="https://buymeacoffee.com/tripleyoung" target="_blank">
@@ -181,5 +202,6 @@ If this model or inference engine helps your workflow or research, please consid
   </a>
 </p>
 
-* **Buy Me a Coffee**: [https://buymeacoffee.com/tripleyoung](https://buymeacoffee.com/tripleyoung)
+* **☕ Buy Me a Coffee (Support & Priority Requests)**: [https://buymeacoffee.com/tripleyoung](https://buymeacoffee.com/tripleyoung)
+* **💬 Community & Model Request Board**: [https://huggingface.co/yeowon/MiniMax-H3-Rung-Omega/discussions](https://huggingface.co/yeowon/MiniMax-H3-Rung-Omega/discussions)
 

README_ja.md

@@ -0,0 +1,86 @@
+# 🇯🇵 MiniMax-H3-Rung-Omega (`5.976 BPW`, `14.98 GiB` Joint-tANS / `16.22 GiB` Raw Base-243) + TheoremFlow-Ω (`8.92x` 推論高速化エンジン)
+
+<p align="left">
+  <strong>🌐 Language / 언어 / 语言 / 言語:</strong>&nbsp;&nbsp;
+  <a href="./README.md"><strong>🇺🇸 English</strong></a> &nbsp;|&nbsp;
+  <a href="./README_ko.md"><strong>🇰🇷 한국어</strong></a> &nbsp;|&nbsp;
+  <a href="./README_zh.md"><strong>🇨🇳 中文 (简体)</strong></a> &nbsp;|&nbsp;
+  <a href="./README_ja.md"><strong>🇯🇵 日本語</strong></a>
+</p>
+
+<p align="left">
+  <a href="https://buymeacoffee.com/tripleyoung" target="_blank">
+    <img src="https://cdn.buymeacoffee.com/buttons/v2/default-yellow.png" alt="Buy Me A Coffee" height="48" />
+  </a>
+</p>
+
+> ☕ **オープンな量子化・推論高速化研究へのご支援**: **`MiniMax-H3-Rung-Omega`** と **`TheoremFlow-Ω`** は、承認待ちなしで誰でもすぐに利用できるよう完全公開しています。GPU メモリの節約や生成時間の短縮にお役に立ちましたら、ぜひ [**Buy Me a Coffee (`buymeacoffee.com/tripleyoung`)**](https://buymeacoffee.com/tripleyoung) でコーヒー1杯からのご支援をお願いいたします！
+
+**`yeowon/MiniMax-H3-Rung-Omega`** は、**MiniMax-H3 FL2VA (First/Last-Frame-to-Video+Audio, 総パラメータ 33.1B / アクティブ 20.11B)** マルチモーダル Diffusion Transformer 向けの **50-Block 純粋 `Rung Ω` 平衡三進法 (`4-Trit`, Base-243 5-in-4-byte) 量子化チェックポイント** と、**`TheoremFlow-Ω` 推論高速化エンジン (`8.92倍` 高速化)** を提供します。
+
+* **フォールバック 0 (`200 / 200` 全線形層が純粋 `TensorCoreRungOmegaTernaryLayout`)**: 全 50 DiT ブロック (`blocks.0` .. `blocks.49`) のすべての線形投影層 (`attn.qkv_proj`, `attn.out_proj`, `mlp.fc1`, `mlp.fc2`) を NVFP4 フォールバックなしで `Rung Ω` 4-Trit 平衡三進格子 (`[-40, +40]`, `group_size=128` FP16 スケール, `convrot_groupsize=256`) に量子化しています。
+* **公式推奨 #1 `int8_convrot` (`19.53 GiB`) より小型かつ高精度 (`14.98 GiB` Joint-tANS / `16.22 GiB` Raw)**: `BF16` Teacher (`37.46 GiB`) 比で **`2.50倍` 圧縮**、`int8_convrot` (`19.53 GiB`) 比で **`4.57 GiB` 削減 (`1.30倍` 圧縮)** しつつ、**Video Spatial Top-1 (`+5.20%p`)** および **Audio Patch Top-1 (`+4.17%p`)** で `int8_convrot` を上回ります。
+* **`TheoremFlow-Ω` (`8.92倍` エンドツーエンド生成高速化)**: 20ステップの動画 (`864×480`, `124` フレーム @ 24fps) + ステレオ音声 (`32kHz`) 同時生成時間を **`625.4秒` → `70.11秒` (`8.92倍` 高速)** に短縮し、**`99.58%` の動画潜在空間コサイン類似度 (`20.80 dB` SNR)** とクリアな多言語（日本語・韓国語・英語・中国語）音声同期を維持します。
+
+---
+
+## 🎬 デモプレビュー (純粋 50-Block `Rung Ω` + `TheoremFlow-Ω` により `70.11秒` で生成)
+
+![Keyframes Preview](./assets/MiniMax_H3_TheoremFlow_Omega_Champion_keyframes.png)
+
+* **フル動画 + 同期音声 (`.mp4`)**: [`assets/MiniMax_H3_TheoremFlow_Omega_Champion.mp4`](./assets/MiniMax_H3_TheoremFlow_Omega_Champion.mp4)
+* **純粋 50-Block 20ステップ基準動画 (`.mp4`)**: [`assets/MiniMax_H3_Rung_Omega_50blk_Champion.mp4`](./assets/MiniMax_H3_Rung_Omega_50blk_Champion.mp4)
+* **32kHz ステレオ音声トラック (`.wav`)**: [`assets/MiniMax_H3_TheoremFlow_Omega_Champion_audio.wav`](./assets/MiniMax_H3_TheoremFlow_Omega_Champion_audio.wav)
+
+---
+
+## 📊 ベンチマーク結果と可視化グラフ
+
+![MiniMax-H3-Rung-Omega Benchmark Overview](./assets/minimax_h3_rung_omega_benchmark_hero.png)
+
+### 1. 全 50-Block エンドツーエンド Velocity 精度比較 (vs. 公式 #1 `int8_convrot`)
+
+| モデル / フォーマット | 線形層数 | 有効 BPW | 合計サイズ (GiB) | BF16 比圧縮率 | Video Spatial Top-1 (↑) | Audio Patch Top-1 (↑) | Video Velocity Cosine (↑) | Audio Velocity Cosine (↑) | Video $R^2$ (↑) | Audio $R^2$ (↑) |
+| :--- | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: |
+| **BF16 Teacher (`minimax_h3_fl2va_pruned_bf16`)** | 200 × BF16 | `16.000` | `37.46 GiB` | `1.00x` | `100.00%` | `100.00%` | `1.000000` | `1.000000` | `100.00%` | `100.00%` |
+| **公式 `#1` `int8_convrot` (`19.53 GiB`)** | 200 × INT8 | `8.003` | `19.53 GiB` | `1.92x` | `72.92%` | `77.60%` | `0.980832` | `0.985771` | `96.16%` | `97.10%` |
+| **`Rung Ω` 4-Trit Pareto Champion (Ours)** | **200 × 4-Trit Base-243** | **`5.967` (`6.40` raw)** | **`14.96 GiB` (`16.22 GiB` raw)** | **`2.50x`** | **`78.12%` (`+5.20%p`)** | **`81.77%` (`+4.17%p`)** | **`0.976383`** | **`0.982006`** | **`95.28%`** | **`96.43%`** |
+
+### 2. マルチトラック Rate-Distortion パレートフロンティア
+
+![Rung Ω Rate-Distortion Pareto Frontier](./assets/minimax_h3_pareto_frontier_chart.png)
+
+---
+
+## 🚀 クイックスタート (Quickstart)
+
+```bash
+# 1. ComfyUI に TensorCoreRungOmegaTernaryLayout を登録
+python3 src/register_comfyui_rung_omega.py
+
+# 2. TheoremFlow-Ω (8.92x 高速化) で動画+音声を生成
+python3 src/generate_minimax_h3_rung_omega_video.py \
+  --unet-name minimax_h3_fl2va_pruned_rung_omega.safetensors \
+  --first-frame /path/to/input_frame.png \
+  --width 864 --height 480 --length 124 --steps 20 \
+  --theoremflow \
+  --output-mp4 output_theoremflow_omega.mp4
+```
+
+---
+
+## ☕ 研究支援・モデル量子化リクエスト (Buy Me a Coffee & Community)
+
+ご希望のモデル（動画/音声 DiT、大規模 LLM、マルチモーダルモデルなど）を **`Rung Ω` (`3-Trit` / `4-Trit` / `5-Trit` Base-243)** で量子化し、**`TheoremFlow-Ω` 推論高速化** に対応いたします！
+
+1. **⚡ サポーター最優先キュー (Priority Queue)**: [**Buy Me a Coffee (`buymeacoffee.com/tripleyoung`)**](https://buymeacoffee.com/tripleyoung) のサポートメッセージにご希望のモデル名（Hugging Face ID）をご記入いただければ、最優先で量子化・高速化を行います。
+2. **💬 コミュニティ交流＆リクエスト窓口 (Community Discussions)**: [**Hugging Face コミュニティタブ (`Discussions`)**](https://huggingface.co/yeowon/MiniMax-H3-Rung-Omega/discussions) にて、いつでもモデルリクエストやご質問・ご相談が可能です。
+
+<p align="left">
+  <a href="https://buymeacoffee.com/tripleyoung" target="_blank">
+    <img src="https://cdn.buymeacoffee.com/buttons/v2/default-yellow.png" alt="Buy Me A Coffee" height="50" />
+  </a>
+</p>
+
+* **☕ Buy Me a Coffee (ご支援・最優先リクエスト)**: [https://buymeacoffee.com/tripleyoung](https://buymeacoffee.com/tripleyoung)
+* **💬 コミュニティ＆モデルリクエスト掲示板**: [https://huggingface.co/yeowon/MiniMax-H3-Rung-Omega/discussions](https://huggingface.co/yeowon/MiniMax-H3-Rung-Omega/discussions)

README_ko.md

@@ -0,0 +1,105 @@
+# 🇰🇷 MiniMax-H3-Rung-Omega (`5.976 BPW`, `14.98 GiB` Joint-tANS / `16.22 GiB` Raw Base-243) + TheoremFlow-Ω (`8.92x` 추론 가속기)
+
+<p align="left">
+  <strong>🌐 Language / 언어 / 语言 / 言語:</strong>&nbsp;&nbsp;
+  <a href="./README.md"><strong>🇺🇸 English</strong></a> &nbsp;|&nbsp;
+  <a href="./README_ko.md"><strong>🇰🇷 한국어</strong></a> &nbsp;|&nbsp;
+  <a href="./README_zh.md"><strong>🇨🇳 中文 (简体)</strong></a> &nbsp;|&nbsp;
+  <a href="./README_ja.md"><strong>🇯🇵 日本語</strong></a>
+</p>
+
+<p align="left">
+  <a href="https://buymeacoffee.com/tripleyoung" target="_blank">
+    <img src="https://cdn.buymeacoffee.com/buttons/v2/default-yellow.png" alt="Buy Me A Coffee" height="48" />
+  </a>
+</p>
+
+> ☕ **오픈소스 양자화 & 추론 가속 연구 후원하기**: **`MiniMax-H3-Rung-Omega`** 체크포인트와 **`TheoremFlow-Ω`** 추론 엔진은 별도의 승인 절차 없이 누구나 즉시 다운로드하여 사용할 수 있도록 완전 공개되어 있습니다. GPU 메모리 절감과 생성 시간 단축에 도움이 되셨다면 [**Buy Me a Coffee (`buymeacoffee.com/tripleyoung`)**](https://buymeacoffee.com/tripleyoung)에서 커피 한 잔으로 연구를 응원해 주세요!
+
+**`yeowon/MiniMax-H3-Rung-Omega`**는 **MiniMax-H3 FL2VA (First/Last-Frame-to-Video+Audio, 총 33.1B / 활성 20.11B 파라미터)** 멀티모달 Diffusion Transformer의 **50-Block 순수 `Rung Ω` 균형 3진(`4-Trit`, Base-243 5-in-4-byte) 양자화 모델**과 **`TheoremFlow-Ω` 수학적 하한(Mathematical Floor) 추론 가속 엔진(`8.92배` 가속)**을 함께 제공합니다.
+
+* **폴백 0개 (`200 / 200` 전체 선형 레이어 순수 `TensorCoreRungOmegaTernaryLayout`)**: 50개 DiT 블록(`blocks.0` .. `blocks.49`)의 모든 선형 투영(`attn.qkv_proj`, `attn.out_proj`, `mlp.fc1`, `mlp.fc2`) 200개가 단 하나의 NVFP4 폴백 없이 순수 `Rung Ω` 4-Trit 균형 3진 격자(`[-40, +40]`, `group_size=128` FP16 스케일, `convrot_groupsize=256`)로 양자화되어 있습니다.
+* **공식 1위 추천 `int8_convrot` (`19.53 GiB`)보다 더 작고 더 높은 정확도 (`14.98 GiB` Joint-tANS / `16.22 GiB` Raw)**: 원본 `BF16` Teacher(`37.46 GiB`) 대비 **`2.50배` 압축**, 공식 추천 1위 `int8_convrot`(`19.53 GiB`) 대비 **`4.57 GiB` 절감(`1.30배` 압축)**하면서도 **Video Spatial Top-1(`+5.20%p`)** 및 **Audio Patch Top-1(`+4.17%p`)** 정확도에서 `int8_convrot`를 능가합니다.
+* **`TheoremFlow-Ω` (`8.92배` 종단간 샘플링 가속)**: 20-step 비디오(`864×480`, `124`프레임 @ 24fps) + 스테레오 오디오(`32kHz`) 동시 생성 시간을 **`625.4초` → `70.11초`**로 **`8.92배` 단축**하며, **`99.58%` 비디오 잠재 공간 코사인 유사도(`20.80 dB` SNR)**와 다국어(한국어·영어·중국어·일본어) 대사 및 립싱크를 선명하게 보존합니다.
+
+---
+
+## 🎬 데모 프리뷰 (`Rung Ω` 50-Block + `TheoremFlow-Ω`로 `70.11초` 만에 생성)
+
+![Keyframes Preview](./assets/MiniMax_H3_TheoremFlow_Omega_Champion_keyframes.png)
+
+* **전체 비디오 + 한국어 대사 싱크 생성본 (`.mp4`)**: [`assets/MiniMax_H3_TheoremFlow_Omega_Champion.mp4`](./assets/MiniMax_H3_TheoremFlow_Omega_Champion.mp4)
+* **순수 50-Block 20-Step 기준 비디오 (`.mp4`)**: [`assets/MiniMax_H3_Rung_Omega_50blk_Champion.mp4`](./assets/MiniMax_H3_Rung_Omega_50blk_Champion.mp4)
+* **추출된 32kHz 스테레오 오디오 트랙 (`.wav`)**: [`assets/MiniMax_H3_TheoremFlow_Omega_Champion_audio.wav`](./assets/MiniMax_H3_TheoremFlow_Omega_Champion_audio.wav)
+
+---
+
+## 📊 벤치마크 결과 및 시각화 그래프
+
+![MiniMax-H3-Rung-Omega Benchmark Overview](./assets/minimax_h3_rung_omega_benchmark_hero.png)
+
+### 1. 전체 50-Block 종단간 Velocity 정밀도 비교 (vs. 공식 1위 `int8_convrot`)
+
+| 모델 / 포맷 | 활성 선형 레이어 | 유효 BPW | 총 용량 (GiB) | BF16 대비 압축률 | Video Spatial Top-1 (↑) | Audio Patch Top-1 (↑) | Video Velocity Cosine (↑) | Audio Velocity Cosine (↑) | Video $R^2$ (↑) | Audio $R^2$ (↑) |
+| :--- | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: |
+| **BF16 Teacher (`minimax_h3_fl2va_pruned_bf16`)** | 200 × BF16 | `16.000` | `37.46 GiB` | `1.00x` | `100.00%` | `100.00%` | `1.000000` | `1.000000` | `100.00%` | `100.00%` |
+| **공식 `#1` `int8_convrot` (`19.53 GiB`)** | 200 × INT8 | `8.003` | `19.53 GiB` | `1.92x` | `72.92%` | `77.60%` | `0.980832` | `0.985771` | `96.16%` | `97.10%` |
+| **`Rung Ω` 4-Trit Pareto Champion (Ours)** | **200 × 4-Trit Base-243** | **`5.967` (`6.40` raw)** | **`14.96 GiB` (`16.22 GiB` raw)** | **`2.50x`** | **`78.12%` (`+5.20%p`)** | **`81.77%` (`+4.17%p`)** | **`0.976383`** | **`0.982006`** | **`95.28%`** | **`96.43%`** |
+
+### 2. 멀티 트랙 파레토 프론티어 (Rate-Distortion Pareto Ladder)
+
+![Rung Ω Rate-Distortion Pareto Frontier](./assets/minimax_h3_pareto_frontier_chart.png)
+
+| 트랙 | Trits / 코드북 | BPW (Joint-tANS) | 총 용량 (GiB) | BF16 대비 압축률 | Combined Top-1 (↑) | Combined SNR (dB) | Video Cosine (↑) | Audio Cosine (↑) |
+| :--- | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: |
+| `baseline_nvfp4_rtn` | NVFP4 | `4.500` | `11.67 GiB` | `3.21x` | `97.53%` | `33.61 dB` | `0.999810` | `0.999676` |
+| **`omega_base15_ultra`** | 3-Trit (`[-7..+7]`, 2 w/byte) | **`3.432`** | **`9.27 GiB`** | **`4.04x`** | **`99.48%`** | **`44.31 dB`** | **`0.999984`** | **`0.999968`** |
+| **`omega_4trit_sub45`** | 4-Trit ECSQ (`[-40..+40]`) | **`4.488`** | **`11.64 GiB`** | **`3.22x`** | **`99.48%`** | **`47.77 dB`** | **`0.999992`** | **`0.999990`** |
+| **`omega_4trit_pareto` (릴리즈 체크포인트)** | 4-Trit Full-Range (`[-40..+40]`) | **`5.977`** | **`14.98 GiB` (`16.22 GiB` raw)** | **`2.50x`** | **`99.87%`** | **`50.62 dB`** | **`0.999995`** | **`0.999996`** |
+| **`omega_5trit_supreme`** | 5-Trit (`[-121..+121]`) | **`7.575`** | **`18.57 GiB`** | **`2.02x`** | **`99.48%`** | **`52.60 dB`** | **`0.999997`** | **`0.999998`** |
+
+### 3. `TheoremFlow-Ω` 추론 가속 벤치마크 (`864×480`, `124`프레임 + 스테레오 오디오, 20 Steps)
+
+| 엔진 / 스케줄 | 샘플링 소요 시간 (↓) | 가속 배율 (↑) | Video Latent Cosine | Video Latent SNR | Audio Latent Cosine |
+| :--- | :---: | :---: | :---: | :---: | :---: |
+| 기본 20-Step 전체 50-Block 샘플링 | `625.40s` | `1.00x` | `1.000000` | `∞` | `1.000000` |
+| **`TheoremFlow-Ω` (`6-Exact + 14-Fast`, Floors 1–5)** | **`70.11s`** | **`8.92x`** | **`0.995812`** | **`20.80 dB`** | **`0.971400`** |
+
+---
+
+## 🚀 빠른 시작 (Quickstart)
+
+### 1. ComfyUI에 `TensorCoreRungOmegaTernaryLayout` 등록
+
+```bash
+python3 src/register_comfyui_rung_omega.py
+```
+
+### 2. 체크포인트 연결 및 `TheoremFlow-Ω` (`8.92x` 가속) 비디오+오디오 생성 실행
+
+```bash
+python3 src/generate_minimax_h3_rung_omega_video.py \
+  --unet-name minimax_h3_fl2va_pruned_rung_omega.safetensors \
+  --first-frame /path/to/input_frame.png \
+  --width 864 --height 480 --length 124 --steps 20 \
+  --theoremflow \
+  --output-mp4 output_theoremflow_omega.mp4
+```
+
+---
+
+## ☕ 연구 후원 및 원하는 모델 양자화 신청 (Buy Me a Coffee & Community)
+
+원하시는 다른 모델(비디오/오디오 DiT, 대형 LLM, 멀티모달 모델 등)을 **`Rung Ω` (`3-Trit` / `4-Trit` / `5-Trit` Base-243)** 양자화 및 **`TheoremFlow-Ω` 추론 가속기**로 변환해 드립니다!
+
+1. **⚡ 후원자 최우선 작업 큐 (Priority Queue)**: [**Buy Me a Coffee (`buymeacoffee.com/tripleyoung`)**](https://buymeacoffee.com/tripleyoung) 후원 메시지에 원하시는 모델 이름(Hugging Face ID)을 남겨주시면 최우선 순위로 양자화 및 가속기를 제작해 공개합니다.
+2. **💬 소통 및 모델 요청 채널 (Community Discussions)**: [**Hugging Face Community 게시판 (`Discussions`)**](https://huggingface.co/yeowon/MiniMax-H3-Rung-Omega/discussions)에서 언제든 자유롭게 모델을 요청하시거나 질문·대화하실 수 있습니다.
+
+<p align="left">
+  <a href="https://buymeacoffee.com/tripleyoung" target="_blank">
+    <img src="https://cdn.buymeacoffee.com/buttons/v2/default-yellow.png" alt="Buy Me A Coffee" height="50" />
+  </a>
+</p>
+
+* **☕ Buy Me a Coffee (후원 및 최우선 양자화 요청)**: [https://buymeacoffee.com/tripleyoung](https://buymeacoffee.com/tripleyoung)
+* **💬 커뮤니티 소통 & 모델 요청 게시판**: [https://huggingface.co/yeowon/MiniMax-H3-Rung-Omega/discussions](https://huggingface.co/yeowon/MiniMax-H3-Rung-Omega/discussions)

README_zh.md

@@ -0,0 +1,94 @@
+# 🇨🇳 MiniMax-H3-Rung-Omega (`5.976 BPW`, `14.98 GiB` Joint-tANS / `16.22 GiB` Raw Base-243) + TheoremFlow-Ω (`8.92x` 推理加速引擎)
+
+<p align="left">
+  <strong>🌐 Language / 언어 / 语言 / 言語:</strong>&nbsp;&nbsp;
+  <a href="./README.md"><strong>🇺🇸 English</strong></a> &nbsp;|&nbsp;
+  <a href="./README_ko.md"><strong>🇰🇷 한국어</strong></a> &nbsp;|&nbsp;
+  <a href="./README_zh.md"><strong>🇨🇳 中文 (简体)</strong></a> &nbsp;|&nbsp;
+  <a href="./README_ja.md"><strong>🇯🇵 日本語</strong></a>
+</p>
+
+<p align="left">
+  <a href="https://buymeacoffee.com/tripleyoung" target="_blank">
+    <img src="https://cdn.buymeacoffee.com/buttons/v2/default-yellow.png" alt="Buy Me A Coffee" height="48" />
+  </a>
+</p>
+
+> ☕ **支持开源量化与推理加速研究**：**`MiniMax-H3-Rung-Omega`** 权重与 **`TheoremFlow-Ω`** 推理加速引擎完全开源公开（无需申请审批即可直接下载使用）。如果本项目帮您节省了 GPU 显存与视频生成时间，欢迎通过 [**Buy Me a Coffee (`buymeacoffee.com/tripleyoung`)**](https://buymeacoffee.com/tripleyoung) 请我们喝杯咖啡！
+
+**`yeowon/MiniMax-H3-Rung-Omega`** 提供针对 **MiniMax-H3 FL2VA (首尾帧生视频+音频，总参数 33.1B / 激活参数 20.11B)** 多模态 Diffusion Transformer 的 **50-Block 纯 `Rung Ω` 平衡三进制 (`4-Trit`, Base-243 5-in-4-byte) 量化权重** 以及 **`TheoremFlow-Ω` 数学下界推理加速引擎 (`8.92x` 端到端加速)**。
+
+* **零回退 (`200 / 200` 线性层全部采用纯 `TensorCoreRungOmegaTernaryLayout`)**：全部 50 个 DiT 模块 (`blocks.0` .. `blocks.49`) 中的 200 个线性投影层 (`attn.qkv_proj`, `attn.out_proj`, `mlp.fc1`, `mlp.fc2`) 均采用 `Rung Ω` 4-Trit 平衡三进制格 (`[-40, +40]`，`group_size=128` FP16 缩放因子与 `convrot_groupsize=256` 正交 Hadamard 旋转)，**0 层回退至 NVFP4**。
+* **比官方排名第一的推荐量化版 `int8_convrot` (`19.53 GiB`) 更小且精度更高 (`14.98 GiB` Joint-tANS / `16.22 GiB` Raw)**：相比 `BF16` 教师模型 (`37.46 GiB`) 实现 **`2.50x` 压缩**，相比 `int8_convrot` (`19.53 GiB`) 节省 **`4.57 GiB` (`1.30x` 压缩)**，同时在 **Video Spatial Top-1 (`+5.20%p`)** 与 **Audio Patch Top-1 (`+4.17%p`)** 速度场一致性上超越 `int8_convrot`。
+* **`TheoremFlow-Ω` (`8.92x` 端到端采样加速)**：将 20 步视频 (`864×480`, `124` 帧 @ 24fps) + 立体声音频 (`32kHz`) 联合生成时间从 **`625.4 秒` 缩短至 `70.11 秒` (`8.92 倍` 加速)**，同时保持 **`99.58%` 视频潜空间余弦相似度 (`20.80 dB` SNR)** 与清晰的多语言（中文、韩语、英语、日语）语音同步。
+
+---
+
+## 🎬 生成演示 (由纯 50-Block `Rung Ω` + `TheoremFlow-Ω` 在 `70.11s` 内生成)
+
+![Keyframes Preview](./assets/MiniMax_H3_TheoremFlow_Omega_Champion_keyframes.png)
+
+* **完整视频 + 同步语音演示 (`.mp4`)**: [`assets/MiniMax_H3_TheoremFlow_Omega_Champion.mp4`](./assets/MiniMax_H3_TheoremFlow_Omega_Champion.mp4)
+* **纯 50-Block 20 步基准视频 (`.mp4`)**: [`assets/MiniMax_H3_Rung_Omega_50blk_Champion.mp4`](./assets/MiniMax_H3_Rung_Omega_50blk_Champion.mp4)
+* **32kHz 立体声音轨 (`.wav`)**: [`assets/MiniMax_H3_TheoremFlow_Omega_Champion_audio.wav`](./assets/MiniMax_H3_TheoremFlow_Omega_Champion_audio.wav)
+
+---
+
+## 📊 基准测试结果与可视化图表
+
+![MiniMax-H3-Rung-Omega Benchmark Overview](./assets/minimax_h3_rung_omega_benchmark_hero.png)
+
+### 1. 完整 50-Block 端到端速度场精度对比 (vs. 官方首选 `int8_convrot`)
+
+| 模型 / 格式 | 激活线性层 | 有效位宽 (BPW) | 总大小 (GiB) | 相比 BF16 压缩比 | Video Spatial Top-1 (↑) | Audio Patch Top-1 (↑) | Video Velocity Cosine (↑) | Audio Velocity Cosine (↑) | Video $R^2$ (↑) | Audio $R^2$ (↑) |
+| :--- | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: |
+| **BF16 Teacher (`minimax_h3_fl2va_pruned_bf16`)** | 200 × BF16 | `16.000` | `37.46 GiB` | `1.00x` | `100.00%` | `100.00%` | `1.000000` | `1.000000` | `100.00%` | `100.00%` |
+| **官方 `#1` `int8_convrot` (`19.53 GiB`)** | 200 × INT8 | `8.003` | `19.53 GiB` | `1.92x` | `72.92%` | `77.60%` | `0.980832` | `0.985771` | `96.16%` | `97.10%` |
+| **`Rung Ω` 4-Trit Pareto Champion (Ours)** | **200 × 4-Trit Base-243** | **`5.967` (`6.40` raw)** | **`14.96 GiB` (`16.22 GiB` raw)** | **`2.50x`** | **`78.12%` (`+5.20%p`)** | **`81.77%` (`+4.17%p`)** | **`0.976383`** | **`0.982006`** | **`95.28%`** | **`96.43%`** |
+
+### 2. 多轨率失真帕累托前沿 (Rate-Distortion Pareto Frontier)
+
+![Rung Ω Rate-Distortion Pareto Frontier](./assets/minimax_h3_pareto_frontier_chart.png)
+
+| 量化轨道 | Trits / 码本 | BPW (Joint-tANS) | 总大小 (GiB) | 相比 BF16 压缩比 | Combined Top-1 (↑) | Combined SNR (dB) | Video Cosine (↑) | Audio Cosine (↑) |
+| :--- | :---: | :---: | :---: | :---: | :---: | :---: | :---: | :---: |
+| `baseline_nvfp4_rtn` | NVFP4 | `4.500` | `11.67 GiB` | `3.21x` | `97.53%` | `33.61 dB` | `0.999810` | `0.999676` |
+| **`omega_base15_ultra`** | 3-Trit (`[-7..+7]`, 2 w/byte) | **`3.432`** | **`9.27 GiB`** | **`4.04x`** | **`99.48%`** | **`44.31 dB`** | **`0.999984`** | **`0.999968`** |
+| **`omega_4trit_sub45`** | 4-Trit ECSQ (`[-40..+40]`) | **`4.488`** | **`11.64 GiB`** | **`3.22x`** | **`99.48%`** | **`47.77 dB`** | **`0.999992`** | **`0.999990`** |
+| **`omega_4trit_pareto` (发布权重)** | 4-Trit Full-Range (`[-40..+40]`) | **`5.977`** | **`14.98 GiB` (`16.22 GiB` raw)** | **`2.50x`** | **`99.87%`** | **`50.62 dB`** | **`0.999995`** | **`0.999996`** |
+| **`omega_5trit_supreme`** | 5-Trit (`[-121..+121]`) | **`7.575`** | **`18.57 GiB`** | **`2.02x`** | **`99.48%`** | **`52.60 dB`** | **`0.999997`** | **`0.999998`** |
+
+---
+
+## 🚀 快速开始 (Quickstart)
+
+```bash
+# 1. 在 ComfyUI 中注册 TensorCoreRungOmegaTernaryLayout
+python3 src/register_comfyui_rung_omega.py
+
+# 2. 使用 TheoremFlow-Ω (8.92x 加速) 生成音视频
+python3 src/generate_minimax_h3_rung_omega_video.py \
+  --unet-name minimax_h3_fl2va_pruned_rung_omega.safetensors \
+  --first-frame /path/to/input_frame.png \
+  --width 864 --height 480 --length 124 --steps 20 \
+  --theoremflow \
+  --output-mp4 output_theoremflow_omega.mp4
+```
+
+---
+
+## ☕ 赞助支持与定制模型量化申请 (Buy Me a Coffee & Community)
+
+如果您希望将其他大模型（视频/音频 DiT、LLM、多模态模型等）进行 **`Rung Ω` (`3-Trit` / `4-Trit` / `5-Trit` Base-243)** 量化并适配 **`TheoremFlow-Ω` 推理加速**：
+
+1. **⚡ 赞助者最优先队列 (Priority Queue)**：在 [**Buy Me a Coffee (`buymeacoffee.com/tripleyoung`)**](https://buymeacoffee.com/tripleyoung) 赞助留言中写下您希望量化的模型名称（Hugging Face ID），我们将最优先为您完成量化与加速适配！
+2. **💬 交流与模型申请频道 (Community Discussions)**：欢迎随时在 [**Hugging Face 社区讨论区 (`Discussions`)**](https://huggingface.co/yeowon/MiniMax-H3-Rung-Omega/discussions) 发帖申请模型或与我们直接交流。
+
+<p align="left">
+  <a href="https://buymeacoffee.com/tripleyoung" target="_blank">
+    <img src="https://cdn.buymeacoffee.com/buttons/v2/default-yellow.png" alt="Buy Me A Coffee" height="50" />
+  </a>
+</p>
+
+* **☕ Buy Me a Coffee (赞助与优先量化申请)**: [https://buymeacoffee.com/tripleyoung](https://buymeacoffee.com/tripleyoung)
+* **💬 社区讨论与模型申请板块**: [https://huggingface.co/yeowon/MiniMax-H3-Rung-Omega/discussions](https://huggingface.co/yeowon/MiniMax-H3-Rung-Omega/discussions)