OK the Minimax H3 Person Remover V1 lora repo is live:
https://t.co/X4CFAABDEv
I'm using it to generate better samples for the Character Swap V2 dataset. Training starting soon™️
Tomorrow I'm dropping an updated workflow for Character Swap that should help with videos longer than 5s.
Lots more in the pipeline for y'all 🫡
この動画で使ったプロンプト置いときます。
前半[1]と後半[2]に分けて、それぞれにセリフ・音・カメラ・動きを書いてます(res_multistep 20step/576×1024・10秒)
使ってみた感想とか「ここ変えたらもっと良くなる」みたいな指摘があれば、ぜひ教えてください🙏
Keep the exact same adult woman, face, black high ponytail with a red hair tie, bangs, grey eyes, white karate gi with navy undershirt, black belt, and plain white studio background as <Picture 1>. Do not change identity, outfit, or setting. No other person. Calm, focused, strong martial-artist mood. Soft even studio lighting.
Audio:
no music
no BGM
no soundtrack
no melody
no beat
no instruments
no song
non_diegetic_music: N/A
diegetic_music: N/A
music: none
Voice: same adult Japanese female, low calm voice, sharp and powerful when shouting, not an announcer, not Chinese, not English.
[1]
Dialogue (Japanese only): 「……よし。いくよ」
Keep only: that short line, fabric rustle of the gi, a slow exhale. No other words. No narration.
Camera: medium shot from the waist up, eye level, slow push-in toward her face.
She finishes tightening her ponytail with both hands, lowers her arms, and takes a deep breath with her eyes closed. She opens her eyes with a sharp, focused look at the camera, then steps into a fighting stance with both fists raised.
[2]
Dialogue (Japanese only): 「ハッ! …押忍!」
Keep only: one sharp kiai shout, the snap of the gi sleeve, a light foot stomp, then the short "osu". No other words. No narration.
Camera: continuing from the previous moment, the camera pulls back to a full-body shot from a slightly low angle, static.
Continuing from the previous moment: she throws one strong straight punch toward the camera with a loud kiai, her sleeve snapping. She pulls the fist back, returns to a natural stance, places her fists together, and gives a short respectful bow. She ends standing still, looking at the camera with a small confident smile.
@PixelAigc Nice tool. Splitting the prompt into shots made a huge difference for me — my karate clip only worked after I wrote the first and second half separately. Curious how the 8B version handles Japanese dialogue.