お祭りの屋台です。
門戸厄神の厄神大祭や初詣、西宮神社のえべっさんの光景などから学習させてみました。
サンプルプロンプト:
omatsuri, food stand, outdoors, scenery, bag, backpack, power lines, road, street, food, tree, orange (fruit), bare tree, fruit, multiple others, cloud, sky, utility pole, cloudy sky, people, vanishing point, shop, building, photo background, 1girl, multiple boys, city, ambiguous gender, 6+others,
サンプルプロンプト:
omatsuri, food stand, multiple girls, tree, outdoors, bag, hat, skirt, handbag, road, street, walking, multiple boys, scenery, crowd, long skirt, pantyhose, day, long hair, real world location, black skirt, 6+boys, lantern, shirt, 6+girls, backpack, festival, food, photo background, pavement, crosswalk, paper lantern, lamppost, brown hair, white shirt, short hair, holding, sun hat, realistic, photo background, photo (medium), photorealistic
サンプルプロンプト:
omatsuri, food stand, multiple boys, food, shirt, outdoors, scenery, short hair, multiple girls, tree, holding, black hair, apron, day, sandals, 3boys, standing, headband, shorts, brown hair, walking, realistic, photo background, photo (medium), photorealistic
各バージョンの違い Differences between each version
昼間の屋台の風景写真31枚+夜間の屋台の風景写真24枚からDIM128/ALPHA64/Epoch8でomatsuri_V4.safetensorsを作成しました。(※当ページの「V4」タブのモデルです)
omatsuri_V4.safetensors was created with DIM128/ALPHA64/Epoch8 from 31 daytime stall scenery photos + 24 nighttime stall scenery photos.(*This is the model in the "V4" tab of this page.)
しかし、このomatsuri_V4を使って画像を生成すると作例のようにモデルが持つ画風や塗りへの影響がかなり出ます。私はこれを解消したかったのでsd-webui-lora-block-weightを使ってU-net層別の出力を調べました。(※V4ページの作例にXYZプロットの画像があります)
However, when I generate images using this omatsuri_V4, as shown in the example, the model has a considerable impact on the style and paint. I wanted to solve this problem, so I used sd-webui-lora-block-weight to check the output by U-net layer. (*See the XYZ plot image in the example on page V4)
U-net-の層別で色々なプリセットを試した結果、OUTD(OUTD:0,0,0,0,0,0,0,0,1,1,1,1,0,0,0,0,0 ※OUT3,4,5,6のみ)を使えば、学習したお祭り屋台の概念をある程度残しつつ画風への影響がかなり抑えられそうだと思いました。
The results of trying various presets with U-net- stratification,OUTD (OUTD:0,0,0,0,0,0,0,0,1,1,1,1,0,0,0,0,0,0 *OUT3,4,5,6 only)
そこで、SuperMergerを使って omatsuri_V4にOUTDを適用して層別マージしたものが、omatsuri_V4_1.0_OUTD.safetensorsになります。
So, I applied OUTD to omatsuri_V4 using SuperMerger and merged by layer, resulting in omatsuri_V4_1.0_OUTD.safetensors.
これはsd-webui-lora-block-weightを使ってプリセットを設定・適用しなくても<lora:omatsuri_V4_1.0_OUTD:1>で簡単に使えるようになるのが利点です。
The advantage of this is that you can easily use <lora:omatsuri_V4_1.0_OUTD:1> without setting and applying presets using sd-webui-lora-block-weight.
さらにsd-scriptsのリサイズスクリプトを使ってDIM128をDIM8にリサイズしたものがV4_1.0_OUTD_ResizeDIM8になります。DIM128のomatsuri_V4_1.0_OUTDに比べ表現力は低下してしまいますが、LoRAのファイルサイズが大幅に小さくなり、モデルが持つ画風や塗りへの影響が大幅に軽減できます。ご覧頂いている作例の通り、学習したお祭り屋台の概念はある程度反映できているので、私はこれを公開してみることにしました。
In addition, the DIM128 is resized to DIM8 using the sd-scripts resize script, which is V4_1.0_OUTD_ResizeDIM8. Compared to DIM128's omatsuri_V4_1.0_OUTD, the expressive power is reduced, but the LoRA file size is much smaller, and the impact on the model's style and paint can be greatly reduced. As you can see in the examples you have seen, I have decided to publish this as it reflects to some extent the concepts of the festival stalls that I have learned.