MiniMax H3 Evening Report 2026-11-17 | Introducing Sakuga Style — Anime Action & Fight Choreography

This issue reviews 7 new versions of H3, with 1 S, 1 A, 3 B, and 2 C ratings. Top recommendation: Mobius_0105's sakuga action LoRA, which pushes videos toward the cutting rhythm, impact pauses, and hard-edge smears of high-quality Japanese anime action sequences; trigger words, frame rules, and prompt writing guidelines are clearly documented. FourBunny's dance LoRA includes visible examples and scene-specific strength settings, rated A; the same author's dance beat-sync slider has all examples hidden, KinetiX's impact frame strobe flash slider has complete parameters but is paid, and the keyframe workflow v5 has all examples hidden—all rated B. The LoL version of KinetiX Combat Bridge lacks any description, and the paid camera shake slider v2 has all examples hidden—rated C.

Zhenzhen's AI Radar · MiniMax H3 · 2026-11-17 Evening Report

This issue features 7 new versions: Strongly Recommended 1 · Recommended 1 · Worth Watching 3 · Listed Only 2

Editor's Note This issue reviews 7 new versions of H3, with 1 S, 1 A, 3 B, and 2 C ratings. Top recommendation: Mobius_0105's sakuga action LoRA, which pushes videos toward the cutting rhythm, impact pauses, and hard-edge smears of high-quality Japanese anime action sequences; trigger words, frame rules, and prompt writing guidelines are clearly documented. FourBunny's dance LoRA includes visible examples and scene-specific strength settings, rated A; the same author's dance beat-sync slider has all examples hidden, KinetiX's impact frame strobe flash slider has complete parameters but is paid, and the keyframe workflow v5 has all examples hidden—all rated B. The LoL version of KinetiX Combat Bridge lacks any description, and the paid camera shake slider v2 has all examples hidden—rated C.

Strongly Recommended (1)

[S] Sakuga Style — Anime Action & Fight Choreography · v1.0

LoRA · New Model · Author Mobius_0105 · Published 11-17 13:22 · Downloads 102

A LoRA that pushes H3 videos toward high-quality Japanese anime action sequences (sakuga): alternating extreme close-ups and wide shots, cuts timed to the action beats, pauses before impact and overshoot aftermath, and hard-edge smears on moving limbs.

Use the same action sequence prompt to compare with the LoRA toggled on/off, observing differences in cutting and smearing.

Images

Video 1504x832

Reasons for Recommendation

  • New capabilities: The author has annotated each training video segment with specific animation terminology, including shot size, camera height, camera movement amplitude and speed, light direction, smear, impact frames, and held frames, so prompts containing these terms will be recognized.
  • Supports 2D animation fight scenes, magical duels, and action shorts featuring single-light hard silhouettes; audio is jointly trained, meaning audio descriptions also influence the visuals.
  • The author has clearly documented frame count rules, training parameters (rank 16, 2,200 steps), and prompt writing guidelines. 3 example videos are visible, it is free, and the license permits selling merged models.

How to Use

  • Place the trigger word sakuga_style at the beginning; recommended prefix: 2d anime style, sakuga_style, ... to prevent drift toward realism.
  • Frame count must satisfy frames % 17 == 5 (240 frames are invalid); 158 frames (6.58 seconds) is the most stable training length, 243 frames (10.1 seconds) is also usable, but longer durations may drift.
  • Specify shot size and camera height at the start of the sentence (e.g., low angle close shot); in two-person scenes, only one body should move; do not leave overall_soundscape and non_diegetic_music empty.
  • The author did not specify strength settings.

Risks and Limitations

  • 6 sensitive-level examples are hidden.
  • The description only lists strengths, with no failure examples; the author did not specify strength settings.

Trigger Word: sakuga_style

Recommended (1)

[A] MiniMax H3 Dance LoRA · v1.0

LoRA · New Model · Author FourBunny · Published 11-17 18:36 · Downloads 122

A LoRA that enhances H3 dance movements: transforms erratic upper-body shaking into coordinated full-body choreography, suppresses finger sticking during fast arm swings, and strengthens chest-waist-hip separation, foot landing, and the inertial swing of clothing and hair.

Images

Video 896x1184

Reasons for Recommendation

  • Serves TikTok, K-Pop, street dance, and otaku-style solo or group dance videos; the author claims it works stably for both solo and synchronized group dances.
  • The author provides scene-specific strength settings: one tier for synchronized group dances and another for solo dances; 2 example videos are visible, and it is free.
  • The author states it only enhances movements H3 already knows, cannot teach obscure dance styles, and complex choreography still requires multiple generations.

How to Use

  • No trigger word required; strength 0.7–1.0, synchronized group dances 0.7–0.85, solo dances 0.85–1.0.
  • The author recommends stacking with the same author's dance beat-sync slider to enhance sharp stops on strong beats.

Risks and Limitations

  • 4 sensitive-level examples are hidden.
  • The accompanying dance prompt Skill pack requires joining the author's Discord to obtain; the model card itself contains no prompt examples.
  • Complex choreography still requires multiple generations, as explicitly stated by the author.

Worth Watching (3)

[B] MiniMax H3 Infinite Iteration Insanity · v5KrazyKeyframing

Workflow · Version Update · Author Ponder_Stibbons · Published 11-17 15:00 · Downloads 25

v5 of Infinite Iteration Insanity: a H3 ComfyUI workflow specialized for keyframes, adding segment-by-segment keyframes to the v4 multi-reference image structure to stitch multiple generations into long videos exceeding H3's single-generation frame limit.

Images

Diagram 1086x1448

Reasons for Recommendation

  • The author claims v5 is the keyframe version of v4, with stable image-to-image segment continuation, offering a solution to H3's frame limit.
  • v4 structure: two character reference subgraphs with up to 5 slots each, 4 composition references, and 2 audio references; keyframes have 8 slots, with the first segment optionally using a tail frame; reference image numbering is automatically assigned by Qwen.
  • The author explicitly states this is not for beginners: those unfamiliar with ComfyUI and H3 prompt syntax will find it difficult to use.

How to Use

  • Turbo model: 8 steps for the first round + 3 steps for latent upscaling; full model: beta5 20-30 steps, requiring a subsequent upscaling round with LMS to smooth artifacts.
  • The author claims it is tuned only for keyframe mode; other modes are not guaranteed.

Risks and Limitations

  • All examples are sensitive-level (5 hidden), so effects cannot be previewed.
  • Update notes are jokingly written, with no specific changes listed; the workflow has almost no comments.
  • Dependent nodes and models must be verified by the user; this tool was not run.
  • The card image is an AI-generated diagram, not actual model output; the author has no public general-level examples.

[B] MiniMax H3 Dance Beat-Sync Slider · v1.0

LoRA · New Model · Author FourBunny · Published 11-17 19:02 · Downloads 73

A dance beat-sync concept slider: positive weights enhance sharp stops on strong beats, sudden bursts of force, and drumbeat synchronization, removing the sluggishness of slow motion.

Images

Diagram 1448x1086

Reasons for Recommendation

  • Serves dance videos requiring action to hit the beat; it is a companion piece to the same author's dance LoRA.
  • Provides a positive weight range and is free; however, no visible examples are available, so effects cannot be previewed.

How to Use

  • No trigger word required; positive weights 1 to 2.
  • The author recommends using it in conjunction with the same author's Dance LoRA.

Risks and Limitations

  • All examples are sensitive-level (3 hidden), so effects cannot be previewed.
  • The effect of negative weights is not documented by the author.
  • Precise beat alignment still requires multiple generations, as explicitly stated by the author.
  • The card image is an AI-generated diagram, not actual model output; the author has no public general-level examples.

[B] KinetiX Impact Frame & Screen Flash Series · Impact Strobe Flash v1.0

LoRA · Version Update · Author make676767 · Published 11-17 09:37 · Downloads 1 · Paid

KinetiX impact frame strobe flash slider: positive weights produce full-screen pure white overexposure and horizontal lens flares for 1~2 frames at the moment of collision; negative weights suppress lighting jumps and screen flashes during fights.

Images

Video 1216x672

Reasons for Recommendation

  • Serves action shots requiring impact frames, such as sword clashes, heavy punches landing, or ground slams; negative weights can lock environmental lighting during fights.
  • Complete parameters: three strength tiers and three copy-paste-ready prompts (with timestamp notation); 4 example videos are visible.
  • Rated B because it is a paid download.

How to Use

  • No trigger word required.
  • Strength: Action blockbuster +1.2 ~ +1.6, anime full-white impact frame +2.5 ~ +3.0, anti-flash lock -1.0 ~ -2.0.
  • Specify collision moments and descriptions like white-out flash, anamorphic lens bloom in prompts; includes a workflow resource zip file.

Risks and Limitations

  • Paid download.
  • 2 sensitive-level examples are hidden.
  • The author warns that when strength exceeds 1.0, the camera is pulled toward the character's chest/torso, requiring composition constraints in prompts.

Listed Only (2)


Data: Civitai public API (base model MiniMax H3), collection window 11-17 08:31 → 11-17 21:01 (Beijing Time). Recommendation reasons are editor-reviewed; download counts reflect attention only.

Civitai

Explore the latest AI Radar updates