tori29umai/QwenImageEdit2511_LoRA

QwenImageEdit2511 LoRA Collection

言語 / Language: 日本語 | English

---

日本語

概要

Qwen-Image-Edit (2511 リビジョン) 用に学習・派生させた LoRA のコレクションです。 キャラクター変換、オブジェクト抽出、画風変換、吹き出し除去などの自作 LoRA に加え、サードパーティの公開 LoRA を SVD で低ランク (dim) に再エンコードしたものを含みます。

動作環境

• ベースモデル: Qwen-Image-Edit (2511 リビジョン)
• 推論フレームワーク: ComfyUI / Diffusers など LoRA 対応のもの
• 形式: kohya / musubi-tuner LoRA フォーマット (safetensors / bf16)

含まれる LoRA

自作 LoRA

QIE2511_3Dpose2Chara_V14_dim1_1e-3-000120.safetensors

3D ポーズ画像 (Image 2) を参照ポーズとし、Image 1 のキャラクターで完全に置き換える LoRA。Image 1 の絵柄・顔・髪型・衣装を保ったまま、Image 2 のポーズ・構図・表情を取り込みます。

プロンプト:

Use Image 2 as a reference for pose, composition, and facial expression, loosely following its pose, gesture, camera angle, framing, and expression, then replace the character entirely with the character from Image 1, applying the art style, facial features, hairstyle, and outfit design of Image 1, ensuring the final image looks like the character from Image 1 inspired by the pose and expression of Image 2.

QIE2511_3Dpose2Chara_faceless_V3_dim1_1e-3-000120.safetensors

上記の "faceless" 版。完全にのっぺらぼう（目・鼻・口・眉なし）のキャラクターを生成します。

プロンプト:

Use Image 2 as a reference for pose and composition,  following its pose, gesture, camera angle, and framing. Replace the character entirely with the character design from Image 1, applying only the hairstyle, outfit design, color palette, and overall art style of Image 1.  The character should have a completely blank, featureless face (no eyes, nose, mouth, eyebrows, or facial details), creating a smooth faceless appearance.

QIE2511_ObjectExtraction_V10_dim1_1e-3-000120.safetensors

単色のマスク（色は自由、原色がおすすめ）で覆われた領域を再構築し、白背景で出力する LoRA。元のオブジェクトの輪郭・比率・素材感を維持しつつ、マスクで隠れた部分を補完します。

プロンプト:

Remove or reconstruct the areas covered by the semi-transparent pink mask Reconstruct any missing parts of the object that were hidden by the mask by inferring the original shape structure material texture color and design Preserve the object's outline proportions details and overall style consistency with strict accuracy Rebuild the object as a continuous structure without broken lines or distorted geometry Set the background to white

QIE2511_image2rakugaki_V3_dim1_1e-3-000020.safetensors

詳細な画像をフラットで抽象的な落書き調に変換する LoRA。線・細部を削除し、色と光のバランスで形を表現します。

プロンプト:

Convert a detailed image into a flat abstract style using large, smooth color areas. Remove all lines and details, letting shapes blend softly. Focus on color and light balance rather than form accuracy

QIE2511_FukidashiErase_V1_dim4_1e-3-000080_dim1.safetensors

マンガの手書き文字、吹き出し内のテキスト、および吹き出し自体の線を消去し、背景として自然に再描画する LoRA。キャラクターの線画・表情・ポーズはそのまま保持します。

プロンプト:

Remove the text and dialogue inside the speech bubbles of a manga illustration. Do not preserve the outlines of the speech bubbles; instead, redraw the area naturally as background. Do not alter the character's line art, facial expression, or pose in any way.

派生 LoRA (dim 削減版)

collage_qwenedit_lora_V1_dim1.safetensors

元: do9/collage_lora_qwenedit を SVD で dim=1 に削減したもの。 使い方・プロンプトは元リポジトリを参照してください。

Qwen-Image-Edit-2511-Object-Remover_dim1.safetensors

元: prithivMLmods/Qwen-Image-Edit-2511-Object-Remover を SVD で dim=1 に削減したもの。 使い方・プロンプトは元リポジトリを参照してください。

qwen-image-edit-2511-multiple-angles-lora_dim1.safetensors

元: fal/Qwen-Image-Edit-2511-Multiple-Angles-LoRA を SVD で dim=1 に削減したもの。 使い方・プロンプトは元リポジトリを参照してください。

Qwen-Image-Edit-2511-Lightning-4steps-V1.0-bf16_dim{1,4,8}.safetensors

元: lightx2v/Qwen-Image-Edit-2511-Lightning を SVD でそれぞれ dim=1 / 4 / 8 に削減した Lightning (4-step 加速) LoRA。 使い方は元リポジトリを参照してください。

dim 削減について

派生 LoRA はすべて SVD ベースのランク削減を行っており、multiplier=1.0 で適用したときに元 LoRA の最良低ランク近似が再現されるよう調整しています。低 dim ほどメモリ・速度面で軽量になりますが、品質は元 LoRA より若干低下する場合があります。

ライセンス・クレジット

• 自作 LoRA: Apache-2.0 (ベースモデルに準拠)。
• 派生 LoRA: 元リポジトリのライセンスに従ってください。
• ベースモデル: Qwen/Qwen-Image-Edit (Apache-2.0)

---

English

Overview

A LoRA collection for Qwen-Image-Edit (2511 revision). It includes self-trained LoRAs for character transfer, object extraction, art-style conversion, and speech-bubble removal, plus SVD-based low-rank (dim) re-encodings of third-party public LoRAs.

Requirements

• Base model: Qwen-Image-Edit (2511 revision)
• Inference framework: ComfyUI / Diffusers, or any LoRA-capable runtime
• Format: kohya / musubi-tuner LoRA (safetensors / bf16)

Included LoRAs

Original LoRAs

QIE2511_3Dpose2Chara_V14_dim1_1e-3-000120.safetensors

Uses a 3D pose image (Image 2) as a reference and fully replaces the character with the one in Image 1. Image 1's art style, facial features, hairstyle, and outfit are preserved while Image 2's pose, composition, and expression are inherited.

Prompt:

Use Image 2 as a reference for pose, composition, and facial expression, loosely following its pose, gesture, camera angle, framing, and expression, then replace the character entirely with the character from Image 1, applying the art style, facial features, hairstyle, and outfit design of Image 1, ensuring the final image looks like the character from Image 1 inspired by the pose and expression of Image 2.

QIE2511_3Dpose2Chara_faceless_V3_dim1_1e-3-000120.safetensors

Faceless variant of the above — the generated character has a completely blank, featureless face (no eyes, nose, mouth, or eyebrows).

Prompt:

Use Image 2 as a reference for pose and composition,  following its pose, gesture, camera angle, and framing. Replace the character entirely with the character design from Image 1, applying only the hairstyle, outfit design, color palette, and overall art style of Image 1.  The character should have a completely blank, featureless face (no eyes, nose, mouth, eyebrows, or facial details), creating a smooth faceless appearance.

QIE2511_ObjectExtraction_V10_dim1_1e-3-000120.safetensors

Reconstructs the area covered by a semi-transparent pink mask and outputs the object on a white background, while preserving the original outline, proportions, and material/texture style.

Prompt:

Remove or reconstruct the areas covered by the semi-transparent pink mask Reconstruct any missing parts of the object that were hidden by the mask by inferring the original shape structure material texture color and design Preserve the object's outline proportions details and overall style consistency with strict accuracy Rebuild the object as a continuous structure without broken lines or distorted geometry Set the background to white

QIE2511_image2rakugaki_V3_dim1_1e-3-000020.safetensors

Converts a detailed image into a flat, abstract doodle-like style. Lines and fine detail are removed; shape is expressed through color and light balance.

Prompt:

Convert a detailed image into a flat abstract style using large, smooth color areas. Remove all lines and details, letting shapes blend softly. Focus on color and light balance rather than form accuracy

QIE2511_FukidashiErase_V1_dim4_1e-3-000080_dim1.safetensors

Removes both the text inside manga speech bubbles AND the bubble outlines themselves, naturally redrawing the area as background. The character's line art, facial expression, and pose are preserved.

Prompt:

Remove the text and dialogue inside the speech bubbles of a manga illustration. Do not preserve the outlines of the speech bubbles; instead, redraw the area naturally as background. Do not alter the character's line art, facial expression, or pose in any way.

Derivative LoRAs (rank-reduced)

collage_qwenedit_lora_V1_dim1.safetensors

Source: do9/collage_lora_qwenedit, SVD-reduced to dim=1. See the source repository for usage and prompt instructions.

Qwen-Image-Edit-2511-Object-Remover_dim1.safetensors

Source: prithivMLmods/Qwen-Image-Edit-2511-Object-Remover, SVD-reduced to dim=1. See the source repository for usage and prompt instructions.

qwen-image-edit-2511-multiple-angles-lora_dim1.safetensors

Source: fal/Qwen-Image-Edit-2511-Multiple-Angles-LoRA, SVD-reduced to dim=1. See the source repository for usage and prompt instructions.

Qwen-Image-Edit-2511-Lightning-4steps-V1.0-bf16_dim{1,4,8}.safetensors

Source: lightx2v/Qwen-Image-Edit-2511-Lightning, SVD-reduced to dim=1 / 4 / 8 respectively (Lightning 4-step acceleration LoRA). See the source repository for usage instructions.

Notes on Dim Reduction

All derivative LoRAs were rank-reduced via SVD, tuned so that applying them at multiplier=1.0 reproduces the best low-rank approximation of the original. Lower-dim versions are lighter on memory and faster, but may show a slight quality drop compared to the original.

License & Credits

• Original LoRAs: Apache-2.0 (matching the base model).
• Derivative LoRAs: follow the license of their source repositories.
• Base model: Qwen/Qwen-Image-Edit (Apache-2.0).