Skip to content
20 changes: 16 additions & 4 deletions docs.json
Original file line number Diff line number Diff line change
Expand Up @@ -285,6 +285,8 @@
"tutorials/3d/hunyuan3D-2"
]
},
"tutorials/3d/pixal3d",
"tutorials/3d/trellis2",
{
"group": "LLM",
"pages": [
Expand Down Expand Up @@ -419,7 +421,8 @@
]
}
]
}
},
"tutorials/utility/sam3d-body"
]
},
{
Expand Down Expand Up @@ -3239,6 +3242,8 @@
"zh/tutorials/3d/hunyuan3D-2"
]
},
"zh/tutorials/3d/pixal3d",
"zh/tutorials/3d/trellis2",
{
"group": "LLM",
"pages": [
Expand Down Expand Up @@ -3373,7 +3378,8 @@
]
}
]
}
},
"zh/tutorials/utility/sam3d-body"
]
},
{
Expand Down Expand Up @@ -6213,6 +6219,8 @@
"ja/tutorials/3d/hunyuan3D-2"
]
},
"ja/tutorials/3d/pixal3d",
"ja/tutorials/3d/trellis2",
{
"group": "LLM",
"pages": [
Expand Down Expand Up @@ -6348,7 +6356,8 @@
]
}
]
}
},
"ja/tutorials/utility/sam3d-body"
]
},
{
Expand Down Expand Up @@ -9265,6 +9274,8 @@
"ko/tutorials/3d/hunyuan3D-2"
]
},
"ko/tutorials/3d/pixal3d",
"ko/tutorials/3d/trellis2",
{
"group": "LLM",
"pages": [
Expand Down Expand Up @@ -9398,7 +9409,8 @@
]
}
]
}
},
"ko/tutorials/utility/sam3d-body"
]
},
{
Expand Down
114 changes: 114 additions & 0 deletions ja/tutorials/3d/pixal3d.mdx
Original file line number Diff line number Diff line change
@@ -0,0 +1,114 @@
---
title: "Pixal3D 画像からモデルへの ComfyUI ワークフロー例"
description: "Pixal3D を使用して、1枚の画像からフル PBR テクスチャ付きの高忠実度 3Dモデルを生成します。Pixal3D は、Tencent ARC によるピクセル位置合わせ型の画像から 3Dモデルへの変換モデルです。"
sidebarTitle: "Pixal3D"
translationSourceHash: b46ae7b5
translationFrom: tutorials/3d/pixal3d.mdx
translationBlockHashes:
"_intro": ebe506d7
"How it works": 30b7a245
"Steps to run": 1a7d044c
"Model downloads": 890d9666
---
import UpdateReminder from "/snippets/ja/tutorials/update-reminder.mdx"

**Pixal3D** は、Tencent ARC が開発したオープンソースの画像-to-3Dモデルで、SIGGRAPH 2026 で発表されました。単一の画像を、完全なPBRマテリアルを備えた高忠実度の3Dアセットに変換します。

ほとんどの3Dネイティブジェネレーターは、正準空間で形を合成し、アテンションを通じて画像の手がかりを注入するため、ピクセルと3Dの対応関係があいまいになります。一方、Pixal3D はピクセル整合生成を採用しています。バックプロジェクションによってピクセルと3Dの直接的な対応関係を確立するため、生成済みモデルの正面は入力画像と1:1で一致し、テクスチャのゆがみやずれが発生しません。

### Pixal3D: 画像からモデルへ {#pixal3d}

単一の画像をアップロードします。入力ビューに合わせて、完全なPBRテクスチャを持つ高忠実度の3Dモデルを生成します。

このワークフローには TRELLIS.2 パイプラインも含まれています。`Boolean (Switch to Trellis2)` ノードはデフォルトで `false` となり、ここで説明する Pixal3D パイプラインが実行され、Pixal3D モデルも自動的に読み込まれます。

<img src="https://raw.githubusercontent.com/Comfy-Org/workflow_templates/main/templates/3d_pixal3d_trellis2_image_to_model-1.webp" alt="Pixal3Dワークフローのプレビュー" />

<UpdateReminder />

<CardGroup cols={2}>
<Card title="Comfy Cloudで実行" icon="cloud" href="https://cloud.comfy.org/?template=3d_pixal3d_trellis2_image_to_model&utm_source=docs&utm_medium=referral&utm_campaign=pixal3d">
このワークフローをComfy Cloudで即座に実行
</Card>
<Card title="ワークフローをダウンロード" icon="download" href="https://github.com/Comfy-Org/workflow_templates/blob/main/templates/3d_pixal3d_trellis2_image_to_model.json">
JSONをダウンロードするか、テンプレートライブラリで「Pixal3D & TRELLIS.2: Image to Model」を検索してください
</Card>
</CardGroup>

**入力マテリアル**

このファイルを `LoadImage` ノードにアップロードしてください:

<CardGroup cols={1}>
<Card title="viking_wolf_rune_axe.png" icon="image" href="https://raw.githubusercontent.com/Comfy-Org/workflow_templates/main/input/viking_wolf_rune_axe.png">
`LoadImage` ノード122 · `viking_wolf_rune_axe.png`
</Card>
</CardGroup>

<div style={{display: 'grid', gridTemplateColumns: 'repeat(2, minmax(0, 1fr))', gap: '1rem', alignItems: 'start'}}>
<img src="https://raw.githubusercontent.com/Comfy-Org/workflow_templates/main/input/viking_wolf_rune_axe.png" alt="入力画像" style={{width: '100%', height: 'auto', objectFit: 'contain'}} />
</div>

## 仕組み

Pixal3D は、カメラ対応のピクセルアライン生成と、完全なメッシュ後処理パイプラインを組み合わせます:

1. **背景除去**: BiRefNet が入力画像の背景を除去し、ワークフローは被写体を中央の 1024×1024 キャンバスに切り取ります。スイッチで背景除去をスキップできます
2. **カメラ推定**: MoGe が画像からジオメトリとカメラの画角を推定します。画角がピクセルアライン条件付けを駆動します
3. **構造生成**: スパースな構造潜在変数がサンプリングされ、ボクセルにデコードされて、粗いメッシュに変換されます
4. **シェイプの精緻化**: シェイプステージとアップサンプルステージが、ターゲット解像度 (1536) までメッシュを精緻化します
5. **テクスチャ生成**: テクスチャ拡散ステージが PBR マテリアルのボクセル (ベースカラー、メタリック、ラフネス) を生成します
6. **後処理**: DC リメッシュ、QEM デシメーション、UV アンラップ、ベースカラー・法線・アンビエントオクルージョンマップのベイクを行い、最終的なテクスチャ付きメッシュを作成します
## 実行手順

1. **画像を読み込む**: `LoadImage` ノードを使用して、オブジェクトの画像を1枚読み込みます
2. **ワークフローをキューに投入する**: Ctrl(macOS では Cmd)+ Enter キーを押します
3. **パイプラインを待つ**: 構造、形状、テクスチャの各ステージが順番に実行され、その後、後処理が続きます
4. **結果を確認する**: `Preview3DAdvanced` ノードでメッシュを確認します。GLB は `ComfyUI/output/3d/ComfyUI/` に保存されます

## モデルのダウンロード

このワークフローで使用するモデルをダウンロードします。2 つの拡散モデルの両方が必要です。スイッチがどちらを実行するかを選択します。対応する `models/` サブディレクトリに配置します。

<CardGroup cols={2}>
<Card title="Pixal3D UNet モデル" icon="download" href="https://huggingface.co/Comfy-Org/Pixal3D/blob/main/diffusion_models/pixal3d_int8_convrot.safetensors">
pixal3d_int8_convrot.safetensors: Pixal3D 拡散モデル(デフォルトで読み込み)
</Card>
<Card title="TRELLIS.2 UNet モデル" icon="download" href="https://huggingface.co/Comfy-Org/TRELLIS.2/blob/main/diffusion_models/trellis_2_int8_convrot.safetensors">
trellis_2_int8_convrot.safetensors: TRELLIS.2 拡散モデル(スイッチが `true` のときに読み込み)
</Card>
<Card title="シェイプ VAE" icon="download" href="https://huggingface.co/Comfy-Org/Pixal3D/blob/main/vae/trellis_2_shape_vae_bf16.safetensors">
trellis_2_shape_vae_bf16.safetensors: 構造とシェイプのデコード用 VAE
</Card>
<Card title="テクスチャ VAE" icon="download" href="https://huggingface.co/Comfy-Org/Pixal3D/blob/main/vae/trellis_2_texture_vae_bf16.safetensors">
trellis_2_texture_vae_bf16.safetensors: テクスチャのデコード用 VAE
</Card>
<Card title="DINOv3 CLIP ビジョン" icon="download" href="https://huggingface.co/Comfy-Org/Pixal3D/blob/main/clip_vision/dino_v3_L_naf_fp32.safetensors">
dino_v3_L_naf_fp32.safetensors: 画像条件付け用の CLIP ビジョンエンコーダー
</Card>
<Card title="MoGe ジオメトリ" icon="download" href="https://huggingface.co/Comfy-Org/MoGe/blob/main/geometry_estimation/moge_2_vitl_normal_fp16.safetensors">
moge_2_vitl_normal_fp16.safetensors: ピクセルアライン条件付け用の深度・カメラ推定
</Card>
<Card title="BiRefNet 背景除去" icon="download" href="https://huggingface.co/Comfy-Org/BiRefNet/blob/main/background_removal/birefnet.safetensors">
birefnet.safetensors: 前処理用の背景除去モデル
</Card>
</CardGroup>

### モデルの保存場所

```
📂 ComfyUI/
├── 📂 models/
│ ├── 📂 diffusion_models/
│ │ ├── trellis_2_int8_convrot.safetensors
│ │ └── pixal3d_int8_convrot.safetensors
│ ├── 📂 vae/
│ │ ├── trellis_2_shape_vae_bf16.safetensors
│ │ └── trellis_2_texture_vae_bf16.safetensors
│ ├── 📂 clip_vision/
│ │ └── dino_v3_L_naf_fp32.safetensors
│ ├── 📂 geometry_estimation/
│ │ └── moge_2_vitl_normal_fp16.safetensors
│ └── 📂 background_removal/
│ └── birefnet.safetensors
```
111 changes: 111 additions & 0 deletions ja/tutorials/3d/trellis2.mdx
Original file line number Diff line number Diff line change
@@ -0,0 +1,111 @@
---
title: "TRELLIS.2 画像からモデルへの ComfyUI ワークフロー例"
description: "Microsoft のオープンソースである 4B パラメータの画像から 3D モデルへの変換モデル TRELLIS.2 を使用して、1 枚の画像からフル PBR テクスチャ付きの高忠実度 3D モデルを生成します。"
sidebarTitle: "TRELLIS.2"
translationSourceHash: 67712115
translationFrom: tutorials/3d/trellis2.mdx
translationBlockHashes:
"_intro": 62475cee
"How it works": e79c0e78
"Steps to run": f899f7e5
"Model downloads": 125142da
---

import UpdateReminder from '/snippets/ja/tutorials/update-reminder.mdx'

**TRELLIS.2** は、Microsoft Research が開発したオープンソースの画像から3Dモデルを生成するモデルで、TRELLIS の後継です。16倍の空間圧縮を備えたネイティブな3D VAE と、O-Voxel と呼ばれるスパースなボクセル表現に基づいて構築された 4B パラメータのモデルであり、単一の画像からフルPBRマテリアルを備えた高精細な3Dアセットを生成します。

TRELLIS.2 は Pixal3D と基本アーキテクチャを共有しています。両方とも同じ DINOv3 画像エンコーダ、同じ形状・テクスチャ VAE、同じ構造・形状・テクスチャ生成ステージを使用します。異なるのはメインの diffusion モデルと条件付けの形式のみです。[Pixal3D ワークフロー](/tutorials/3d/pixal3d) を実行したことがあれば、モデルのセットアップはほぼ同じです。

### TRELLIS.2: 画像から3Dモデルへ {#trellis2}

単一の画像をアップロードします。TRELLIS.2 を使用して、フルPBRテクスチャ付きの忠実度の高い3Dモデルを生成します。

このワークフローには Pixal3D パイプラインも含まれています。TRELLIS.2 を実行するには、`Boolean (Switch to Trellis2)` ノードを `true` に設定します。ワークフローが TRELLIS.2 のモデルと条件付けに自動的に切り替わります。

<img src="https://raw.githubusercontent.com/Comfy-Org/workflow_templates/main/templates/3d_pixal3d_trellis2_image_to_model-1.webp" alt="TRELLIS.2 ワークフローのプレビュー" />

<UpdateReminder />

<CardGroup cols={2}>
<Card title="Comfy Cloud で実行" icon="cloud" href="https://cloud.comfy.org/?template=3d_pixal3d_trellis2_image_to_model&utm_source=docs&utm_medium=referral&utm_campaign=trellis2">
このワークフローを Comfy Cloud で即座に実行します
</Card>
<Card title="ワークフローをダウンロード" icon="download" href="https://github.com/Comfy-Org/workflow_templates/blob/main/templates/3d_pixal3d_trellis2_image_to_model.json">
JSON をダウンロードするか、テンプレートライブラリで &quot;Pixal3D & TRELLIS.2: Image to Model&quot; を検索してください
</Card>
</CardGroup>

**入力マテリアル**

このファイルを `LoadImage` ノードにアップロードします:

<CardGroup cols={1}>
<Card title="viking_wolf_rune_axe.png" icon="image" href="https://raw.githubusercontent.com/Comfy-Org/workflow_templates/main/input/viking_wolf_rune_axe.png">
`LoadImage` ノード 122 ・ `viking_wolf_rune_axe.png`
</Card>
</CardGroup>

<div style={{display: 'grid', gridTemplateColumns: 'repeat(2, minmax(0, 1fr))', gap: '1rem', alignItems: 'start'}}>
<img src="https://raw.githubusercontent.com/Comfy-Org/workflow_templates/main/input/viking_wolf_rune_axe.png" alt="入力画像" style={{width: '100%', height: 'auto', objectFit: 'contain'}} />
</div>

## 仕組み

TRELLIS.2は、3つのdiffusionステージで3Dアセットを生成し、その後、完全なメッシュ後処理パイプラインを実行します。

1. **背景除去**: BiRefNetが入力画像から背景を除去し、ワークフローが被写体を中央配置の1024×1024キャンバスにクロップします。切り替えで背景除去をスキップできます
2. **TRELLIS.2 パイプラインに切り替え**: `Boolean (Switch to Trellis2)` を `true` に設定します。ワークフローが TRELLIS.2 モデルを自動的に読み込みます
2. **構造生成**: スパース構造latentがサンプリングされ、ボクセルにデコードされてから、ラフなメッシュに変換されます
3. **形状リファイン**: 形状ステージとアップサンプリングステージが、メッシュをターゲット解像度(1536)までリファインします
4. **テクスチャ生成**: テクスチャdiffusionステージがPBRマテリアルボクセル(ベース色、メタリック、ラフネス)を生成します
5. **後処理**: DCリメッシュ、QEMデシメーション、UV展開、およびベース色・ノーマル・アンビエントオクルージョンマップの最終テクスチャ付きメッシュへのベイクを実行します

## 実行手順

1. **画像を読み込む**:`LoadImage` ノードを使用して、オブジェクトの画像を1枚読み込みます。
2. **ワークフローをキューに入れる**:Ctrl(macOSではCmd)+ Enter を押します。
3. **パイプラインの完了を待つ**:構造、形状、テクスチャの各ステージが順番に実行され、続いてポストプロセッシングが行われます。
4. **結果を表示する**:`Preview3DAdvanced` ノードでメッシュを確認します。GLBは `ComfyUI/output/3d/ComfyUI/` に保存されます。

## モデルのダウンロード

このワークフローで使用するモデルをダウンロードします。2 つの拡散モデルの両方が必要です。スイッチがどちらを実行するかを選択します。対応する `models/` サブディレクトリに配置します。

<CardGroup cols={2}>
<Card title="TRELLIS.2 UNet モデル" icon="download" href="https://huggingface.co/Comfy-Org/TRELLIS.2/blob/main/diffusion_models/trellis_2_int8_convrot.safetensors">
trellis_2_int8_convrot.safetensors: TRELLIS.2 拡散モデル(スイッチが `true` のときに読み込み)
</Card>
<Card title="Pixal3D UNet モデル" icon="download" href="https://huggingface.co/Comfy-Org/Pixal3D/blob/main/diffusion_models/pixal3d_int8_convrot.safetensors">
pixal3d_int8_convrot.safetensors: Pixal3D 拡散モデル(デフォルトで読み込み)
</Card>
<Card title="シェイプ VAE" icon="download" href="https://huggingface.co/Comfy-Org/TRELLIS.2/blob/main/vae/trellis_2_shape_vae_bf16.safetensors">
trellis_2_shape_vae_bf16.safetensors: 構造とシェイプのデコード用 VAE
</Card>
<Card title="テクスチャ VAE" icon="download" href="https://huggingface.co/Comfy-Org/TRELLIS.2/blob/main/vae/trellis_2_texture_vae_bf16.safetensors">
trellis_2_texture_vae_bf16.safetensors: テクスチャのデコード用 VAE
</Card>
<Card title="DINOv3 CLIP ビジョン" icon="download" href="https://huggingface.co/Comfy-Org/Pixal3D/blob/main/clip_vision/dino_v3_L_naf_fp32.safetensors">
dino_v3_L_naf_fp32.safetensors: 画像条件付け用の CLIP ビジョンエンコーダー。TRELLIS.2 リポジトリの同等の dino_v3_vit_l.safetensors も利用できます
</Card>
<Card title="BiRefNet 背景除去" icon="download" href="https://huggingface.co/Comfy-Org/BiRefNet/blob/main/background_removal/birefnet.safetensors">
birefnet.safetensors: 前処理用の背景除去モデル
</Card>
</CardGroup>

### モデルの保存場所

```
📂 ComfyUI/
├── 📂 models/
│ ├── 📂 diffusion_models/
│ │ ├── pixal3d_int8_convrot.safetensors
│ │ └── trellis_2_int8_convrot.safetensors
│ ├── 📂 vae/
│ │ ├── trellis_2_shape_vae_bf16.safetensors
│ │ └── trellis_2_texture_vae_bf16.safetensors
│ ├── 📂 clip_vision/
│ │ └── dino_v3_L_naf_fp32.safetensors
│ └── 📂 background_removal/
│ └── birefnet.safetensors
```
Loading
Loading