The intended effect of this LORA is to generate silhouette-alternating videos, but the success rate is not very high.
Expected Results:
The original image should switch between normal lighting and dark silhouette lighting, while keeping the main character unchanged but altering their outfit.
Full-body images have a higher chance of success.
It is recommended to include light source transformation descriptions in the prompt for better results.
This LORA does not work well with anime-style characters.
For this training, I used shorter-duration videos with fewer movements.
Although the results are still not exactly as expected, I now have a better understanding of how to train more interesting effects.
I might experiment with other special effect styles in future attempts.
このLORAの狙いは、シルエットが切り替わる動画 を生成することですが、成功率はあまり高くありません。
想定される効果:
通常の光源 と 暗いシルエットの光源 を切り替えながら、キャラクターの姿はそのまま、衣装だけが変化する。
全身が映った画像 を使用すると成功率が高くなる。
プロンプトに 光源の変化に関する説明 を含めると、より良い結果が得られる可能性がある。
アニメスタイルのキャラクターには効果があまり良くない。
今回のトレーニングでは、動きの少ない短めの動画 を使用しました。
まだ 理想通りの結果にはなっていません が、面白いエフェクトを作る方法 が少し分かってきました。
今後は 別の特殊効果スタイル も試してみるかもしれません。
要完美復刻影片原生的動作真的有難度
我再限縮圖片偵總長度再增加訓練次數試試
過去大力出奇蹟的訓練方式不靠譜了,沒這麼大的力了
Description
I spent eighteen hours retraining a set of models and it got slightly better, but the effect was limited.
FAQ
Comments (8)
「例の提示詞を教えてください、ありがとうございます。」
一名女性正在跳舞,燈光突然暗下剩下微弱的燈光
why not give some example prompts
This is a problem in the early days of the platform. When uploading videos, you cannot attach them like you can with pictures. I will provide it to you directly here. "A woman is dancing, the lights suddenly dim, leaving only a faint light, disco dance"
So out of curiousity....I notice the people change during when the lights turns off. Is that meant to happen? Or they can they stay the same?
From the perspective of the technology at that time, relying solely on this LoRA makes it unavoidable. After all, based on the database I trained, every single one changes after it gets dark. After all, the original purpose of this video was to showcase normal attire with the lights on and sexy attire with the lights off. However, there should currently be technologies available that can be adjusted to achieve the effect you want.
Would you consider making a Wan2.2 or LTX-2 version?
My device probably doesn't have enough memory for WAN 2.2, but I haven't tried training with video for LTX; maybe that would work.
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.