Upload 3 files

README.md

@@ -1,27 +1,63 @@
----
-tags:
-- text-to-image
-- lora
-- diffusers
-- template:diffusion-lora
-widget:
-- output:
-    url: images/ComfyUI_00193_.png
-  text: '-'
-base_model: black-forest-labs/FLUX.2-klein-base-9B
-instance_prompt: null
-license: other
-license_name: flux-non-commercial-license
-license_link: >-
-  https://huggingface.co/black-forest-labs/FLUX.2-klein-base-9B/blob/main/LICENSE.md
----
 # FLUX.2 Klein 9B 360 ERP Outpaint LoRA
 
-<Gallery />
+[![ComfyUI Node](https://img.shields.io/badge/ComfyUI-Panorama--Stickers-blue)](https://github.com/nomadoor/ComfyUI-Panorama-Stickers)
+[![4B Model](https://img.shields.io/badge/Hugging%20Face-4B%20Model-green)](https://huggingface.co/nomadoor/flux-2-klein-4B-360-erp-outpaint-lora)
+[![README JA](https://img.shields.io/badge/README-日本語-lightgrey)](./README_ja.md)
 
+This is a LoRA for FLUX.2 Klein 9B that generates 360-degree panoramas.
 
+It uses an approach where reference images are placed on an empty ERP canvas, and the surrounding area is filled by outpainting. This makes it possible to generate panorama images from a flexible layout.
 
-## Download model
+## Examples
 
+| Input | Output |
+| --- | --- |
+| ![example input](https://i.gyazo.com/d2f8b13debee5e8ef733110f72ffe87a.png) | ![example output](https://i.gyazo.com/0ad707176d21bf6d6d2d02fbb141aa81.png) |
 
-[Download](/nomadoor/flux-2-klein-9B-360-erp-outpaint-lora/tree/main) them in the Files & versions tab.
+## Usage
+
+1. Prepare a `2:1` ERP control image with projected reference images placed on a green canvas.
+2. Enter the trigger prompt below at the beginning of your prompt.
+3. Add any extra scene description after the trigger prompt.
+
+Notes:
+
+- The reference image is expected to be projected into ERP from a **pinhole view** based on its intended panorama placement.
+- A UI for this workflow is available as a Gradio app / ComfyUI custom node.
+- [nomadoor/ComfyUI-Panorama-Stickers](https://github.com/nomadoor/ComfyUI-Panorama-Stickers)
+
+## Trigger Prompt
+
+```text
+Fill the green spaces according to the image. Outpaint as a seamless 360 equirectangular panorama (2:1). Keep the horizon level. Match left and right edges.
+```
+
+## Recommended Resolution
+
+- Generation: `2048x1024` (2K ERP)
+- For panorama use, 2K may still look coarse in some cases, so use upscaling if needed
+
+## Notes
+
+- 🚨Distilled models do not produce good results in most cases. I am still looking for a solution, but for now you should use it with the base model.
+- Most of the training data is realistic scenery, so non-photoreal domains may be less stable.
+- Perfect seam matching is still difficult, and some outputs may benefit from post-processing such as inpainting.
+
+## Training Setup
+
+- [TRAINING_NOTES.md](./TRAINING_NOTES.md)
+
+## Workflow
+
+- `simple.json`
+- `advanced.json (WIP)`
+
+## License
+
+Please follow the license and usage terms of the base model [`black-forest-labs/FLUX.2-klein-base-9B`](https://huggingface.co/black-forest-labs/FLUX.2-klein-base-9B).
+See `LICENSE` included in this folder for additional notes.
+
+## Support
+
+I ended up spending a bit too much on GPU rentals 🫠
+If you would like to support the project, I would really, really appreciate it if you use [☕Ko-fi](https://ko-fi.com/nomadoor) or [❤GitHub Sponsors](https://github.com/sponsors/nomadoor/).

README_ja.md

@@ -0,0 +1,61 @@
+# FLUX.2 Klein 9B 360 ERP Outpaint LoRA
+
+[![ComfyUI Node](https://img.shields.io/badge/ComfyUI-Panorama--Stickers-blue)](https://github.com/nomadoor/ComfyUI-Panorama-Stickers)
+[![4B Model](https://img.shields.io/badge/Hugging%20Face-4B%20Model-green)](https://huggingface.co/nomadoor/flux-2-klein-4B-360-erp-outpaint-lora)
+[![README EN](https://img.shields.io/badge/README-English-lightgrey)](./README.md)
+
+任意の画像から 360 度パノラマを生成する、FLUX.2 Klein 9B 向けの LoRA です。
+
+空の ERP キャンバス上に参照画像を配置し、周囲を outpainting で埋めるアプローチを採用しています。これにより、自由度の高いレイアウトからパノラマ画像を生成できます。
+
+## Examples
+
+| Input | Output |
+| --- | --- |
+| ![example input](https://i.gyazo.com/d2f8b13debee5e8ef733110f72ffe87a.png) | ![example output](https://i.gyazo.com/0ad707176d21bf6d6d2d02fbb141aa81.png) |
+
+## 使い方
+
+1. `2:1` の ERP グリーンキャンバス上に、投影済みの参照画像を配置した `control` 画像を作成します。
+ 
+   * 参照画像は、パノラマ内での配置結果を **pinhole 視点から ERP に展開** したものを使用します。
+   * そのための UI（Gradio / ComfyUI カスタムノード）を用意しているので、ぜひ活用してください。
+   * [nomadoor/ComfyUI-Panorama-Stickers](https://github.com/nomadoor/ComfyUI-Panorama-Stickers)
+2. 下記のトリガープロンプトを **先頭** に入力します。
+3. 必要に応じて、その後ろに任意のプロンプトを追加してください。
+
+### トリガープロンプト
+
+```text
+Fill the green spaces according to the image. Outpaint as a seamless 360 equirectangular panorama (2:1). Keep the horizon level. Match left and right edges.
+```
+
+## 推奨解像度
+
+* 生成: `2048x1024`（2K ERP）
+* パノラマ用途では 2K でも粗く見える場合があるため、必要に応じてアップスケールを併用してください。
+
+## 注意点
+
+* 🚨Distilledモデルではほとんど良い結果が得られません。解決方法を探していますが、基本的にはbase modelと共に使用してください。
+* 学習データのほとんどが実写風景なため、非実写ドメインでは安定しない可能性があります。
+* 左右端（シーム）の完全一致は難しく、ケースによっては後処理（inpaint など）が必要です。
+
+## 学習設定
+
+* [TRAINING_NOTES.md](./TRAINING_NOTES.md)
+
+## Workflow
+
+* `simple.json`
+* `advanced.json(WIP)`
+
+## ライセンス
+
+この LoRA の利用条件は、ベースモデル [`black-forest-labs/FLUX.2-klein-base-9B`](https://huggingface.co/black-forest-labs/FLUX.2-klein-base-9B) のライセンス / 利用規約に従ってください。
+補足は同梱の `LICENSE` を参照してください。
+
+## Support
+
+GPU のレンタルで、少し予算を使いすぎました🫠  
+もし応援していただけるなら、[☕Ko-fi](https://ko-fi.com/nomadoor) または [❤GitHub Sponsors](https://github.com/sponsors/nomadoor/) からサポートいただけるととてもとても嬉しいです。

TRAINING_NOTES.md

@@ -0,0 +1,24 @@
+# Training Notes (9B)
+
+## Summary
+- Base model: `FLUX.2-klein-base-9B`
+- LoRA rank: `16`
+- Optimizer: `AdamW8bit`
+- Learning rate: `3e-5`
+- Training steps (tentative best): `600`
+
+## Dataset
+- Source images were collected from open-license panorama/HDRI archives on the web.
+- License policy was centered on permissive sources (`CC0` / Public Domain), with part of the set including `CC BY` assets.
+- After preprocessing and filtering, training was run on a curated ERP pair set of about `1000` samples.
+
+## Preprocess / Pair Generation
+- Target format: ERP (equirectangular) `2:1`
+- Source panoramas were normalized to ERP and cleaned (invalid aspect/failed reads removed).
+- Control images were generated by:
+  - building a green ERP canvas,
+  - sampling `1-3` reference patches,
+  - projecting patches with a pinhole-based spherical mapping,
+  - handling seam continuity with wrap-aware placement.
+- The model is trained to fill green regions while preserving visible reference context.
+