Updated: Nov 17, 2026
conceptSakuga Style — Anime Action & Fight Choreography
A LoRA for MiniMax H3 trained to push video generation toward the cut, timing and staging of high-grade 2D anime action — the kind of footage usually described as sakuga.
Every training clip was individually captioned with a consistent craft vocabulary: shot scale, camera height, camera move with amplitude and speed, composition, material, light direction, and animation terms such as smear, impact frame, silhouette, held frames and secondary motion.
TRIGGER
sakuga_style
Put it first. Recommended prefix for keeping the look 2D:
2d anime style, sakuga_style, ...
TECHNICAL NOTES
• Frame count must satisfy frames % 17 == 5. 240 frames is invalid.
• 158 frames (6.58 s) is the training length and the safest choice. 243 frames (10.1 s) also works; longer clips drift.
• Audio is co-trained — the LoRA responds to sound description as well as picture.
WHAT IT DOES WELL
Cut rhythm and staging — alternating extreme close and wide, cuts that land on beats, action that chains: each follow-through becomes the next wind-up.
Impact feel — the pause before a strike, the follow-through that carries the body past its balance, the held aftermath.
Camera language — low angles, tracking that stays with the body, long static holds.
Magic and effects — casters standing small in a large space, straight-edged columns of light, energy crossing between two points, smoke and dust rolling under a single lamp, blast shapes built from hard angular lumps rather than smooth plumes.
Surfaces and light — wet stone, a single practical lamp, hard silhouettes.
Motion rendering — hard-edged smears on the moving limb while the rest of the figure stays sharp.
PROMPTING NOTES
• Put shot scale and camera height early: low angle close shot reads better than burying it mid-sentence.
• Write contact as edge against edge rather than sparks: the edge of one running down the length of the other.
• For speed, use smeared across two frames while the head and shoulders stay sharp.
• Keep one body in motion at a time in two-person scenes — only one is in motion at a time.
• Sound description carries real weight; write overall_soundscape and non_diegetic_music rather than leaving them blank.
TRAINING
MiniMax H3 · LoRA rank 16 / alpha 16 · 2,200 steps · lr 1e-4 cosine → 1e-5 · flowmatch scheduler · audio co-trained · 158 frames @ 24 fps.
中文说明
面向 MiniMax H3 的作画 / 打戏 LoRA,把视频生成往日本动画高质量动作段落的切法、时机与演出上推——通常被称为「作画」的那类画面。
触发词:sakuga_style(放最前)
建议前缀:2d anime style, sakuga_style, ...(抑制跑写实)
技术要点
• 帧数必须满足 % 17 == 5,240 帧非法
• 158 帧(6.58 秒)是训练长度,最稳;243 帧(10.1 秒)也可用,更长会漂
• 音频是联合训练的——声音描述会实际影响画面
擅长
• 切镜节奏:极近与极远交替、切点踩在动作节拍上
• 动作连锁:上一个动作的收势成为下一个的起势
• 打击感:命中前的停顿、过冲、余韵
• 镜头语言:低机位、跟随身体、长静止
• 魔法与特效:施术者小、空间大;硬边光柱;能量在两点之间交汇;烟与尘在单一光源下翻滚;爆炸由块状几何构成而非平滑火球
• 表面与光:湿石地面、单一灯具、硬剪影
• 运动表现:移动肢体硬边 smear,其余部分保持清晰
写法建议
• 景别和机位高度放句首:low angle close shot 比塞在中间更有效
• 接触写成刃对刃,而不是火花:the edge of one running down the length of the other
• 写「快」用 smeared across two frames while the head and shoulders stay sharp
• 双人场景保持只有一个身体在动:only one is in motion at a time
• 声音描述权重不低,overall_soundscape 与 non_diegetic_music 不要留空
