MiniMax H3 Evening Report 2026-11-12 | First Recommendation: Minimax Female Accents (North American)

This issue reviews 6 new H3 versions, with 1 S, 1 A, 2 B, and 2 C ratings. The top recommendation is a female accent audio LoRA: it modifies only the audio component of H3, adding Utah, Atlanta, and New Jersey North American accents to dialogue without affecting the visuals. Fight Flow V2 reduces weapon "noodling" and character distortion/blur, includes all three interaction trigger words, but all examples are hidden, earning it an A rating. The X-Ray Skeleton Flash LoRA has detailed parameters but is paid and lacks visible examples, earning a B rating; Nola is a fully parameterized Krea 2 character port, earning a B rating; KinetiX Combat Semantic Bridge v1.1 only has external link descriptions, and Bertie Wells is a standard original character, both earning C ratings.

Zhenzhen's AI Radar · MiniMax H3 · 2026-11-12 Evening Report

This issue features 6 new versions: Strongly Recommended 1 · Recommended 1 · Worth Watching 2 · Listed Only 2

Editor's Note This issue reviews 6 new H3 versions, with 1 S, 1 A, 2 B, and 2 C ratings. The top recommendation is a female accent audio LoRA: it modifies only the audio component of H3, adding Utah, Atlanta, and New Jersey North American accents to dialogue without affecting the visuals. Fight Flow V2 reduces weapon "noodling" and character distortion/blur, includes all three interaction trigger words, but all examples are hidden, earning it an A rating. The X-Ray Skeleton Flash LoRA has detailed parameters but is paid and lacks visible examples, earning a B rating; Nola is a fully parameterized Krea 2 character port, earning a B rating; KinetiX Combat Semantic Bridge v1.1 only has external link descriptions, and Bertie Wells is a standard original character, both earning C ratings.

Strongly Recommended (1)

[S] Minimax Female Accents (North American) · v1.0

LoRA · New Model · Author imb101 · Released 11-12 10:03 · Downloads 141

An accent LoRA that affects only the audio component of H3: it makes female dialogue clearly carry Utah, Atlanta, and New Jersey North American accents.

Image

Video 1184x1760

Reasons for Recommendation

  • New model, serving text-to-video / image-to-video with dialogue: the author states that H3's native accents are relatively flat, and this LoRA strengthens the accents.
  • The author claims it was trained only on the audio component, leaving the video component untouched, so visual output should remain unaffected; different accents can be assigned to two characters in the same scene.
  • Usage instructions are clear: prompt format, weights, and sampling suggestions are provided, along with 1 visible example videos.

How to Use

  • Weight 0.4–0.5; add "in a <accent name> accent" after the dialogue, e.g., The woman says <d>[English]……</d> in a utah accent; currently available options are utah, atlanta, and new jersey.
  • If strange vocal fry occurs, the author suggests increasing the audio sampling rate, believing 3 is appropriate.

Risks and Limitations

  • The author describes it as an experimental model; the accents are far from accurate, merely better than the base model; currently only 3 accent types are available, targeting female voices only.
  • When two characters in the same scene have different accents, they sometimes blend together.
  • 4 examples are hidden, with only 1 visible; the license does not permit the release of derivative models.

Recommended (1)

[A] H3 Fight Flow | Continuous Fight Interaction LoRA · V2.0

LoRA · Version Update · Author FourBunny · Released 11-12 10:09 · Downloads 200

An action interaction LoRA V2 that enables two or more characters to engage in continuous combat within the same fight scene; this version reduces weapon "noodling" and character distortion/blur.

Image

Diagram 1672x941

Reasons for Recommendation

  • Version update: the author claims V2 significantly reduces issues with weapons turning into "noodles" and characters becoming distorted or blurred.
  • It does not teach new moves but ensures actions flow through "initiation → response → contact → result → initiative shift," where the position and center of gravity from the previous move become the starting point for the next, making enemy rotations in multi-person fights more natural.
  • The author suggests using it as an interaction layer, stacked with specialized fighting, weapon, or action repair LoRAs: specialized LoRAs determine which moves are performed, while this LoRA connects them into a coherent fight.

How to Use

  • Select trigger words based on the dominant interaction: use "Unarmed combat interaction." for hand-to-hand, "Melee weapon interaction." for swords, sticks, or guns, and "Gun-fu interaction." for close-range gun control.
  • Do not exceed strength 0.6; the author's test weights for two sampling stages were 0.65 for the first and 0.25 for the second; the author provides a corresponding two-stage latent space upscaling workflow on RunningHub.

Risks and Limitations

  • All 8 examples are hidden, so effects cannot be seen on the page.
  • The instruction "do not exceed 0.6" contradicts the first-stage test weight of 0.65, requiring user experimentation.
  • The author describes it as an experimental model; examples use the "Assistant Director" prompt skill from the author's Discord.
  • The card image is an AI-generated diagram, not the actual model output; the author has no publicly available standard-level examples.

Worth Watching (2)

[B] [Krea 2 DiT] Nola (znolanola) — Consistent Southern Street Muse · v1.0 - MiniMax H3 LoRA

LoRA · Port · Author MaisonLatentia · Released 11-12 10:05 · Downloads 42

A character LoRA for the original character Nola on H3, ported from Krea 2; it locks the face for image-to-video and supports pure text-to-video, including a ComfyUI workflow.

Image

Video 576x768

Reasons for Recommendation

  • Port: a Krea 2 version existed previously; this is the H3 video version; the author states no separate reference adapter node is needed.
  • Parameters are comprehensive: strength settings, plus two sets of sampling configurations for local low-VRAM and cloud full-precision, along with znolanola_minimax_h3_workflow.json.
  • Essentially a standard character LoRA, with 2 visible example videos.

How to Use

  • Use trigger words znolanola or znolawoman in the prompt, with strength 0.85 – 0.95 (default 0.90).
  • Local 8GB: FL2VA Q4_K_S + EasyCache, 15 steps, Euler / Simple, CFG 1.2; Cloud full-precision: 35 steps, Euler / Simple, CFG 3.5.
  • For image-to-video, connect Nola's image to the workflow's start frame; for pure text-to-video, bypass first_frame.

Risks and Limitations

  • Original fictional character, usage limited to this character; 3 examples are hidden.
  • Most of the model card is dedicated to the character's backstory.

Trigger Words: znolanola, znolawoman

[B] KinetiX Impact Frame & Screen Flash Series · X-Ray Skeleton Flash v1

LoRA · New Model · Author make676767 · Released 11-12 12:28 · Downloads 2 · Paid

An X-ray skeleton flash effect LoRA for impact moments: when a heavy punch, kick, or lightning strike hits, the character flashes into a transparent skeleton before returning to normal skin.

Image

Diagram 1086x1448

Reasons for Recommendation

  • New model, serving impact effects in fighting and action videos: the author claims the flash lasts only 2 to 5 frames before restoring full color, ensuring the character does not remain in a skeleton state.
  • Weight tiers are detailed: +0.5 to +0.8 maintains full-body composition; exceeding 1.0 zooms the camera to a chest close-up; negative weights -0.5 to -1.0 can suppress all flashing.
  • Includes three directly copyable prompts for a heavy punch to the chest, a high kick to the jaw, and a full-body lightning strike.

How to Use

  • No trigger word; keep weight below 0.8 for full-body shots, use +1.0 to +1.5 for chest close-ups.
  • Clearly specify the hit location and X-ray flash description in the prompt, adapting from the author's three examples.

Risks and Limitations

  • Paid download.
  • All 5 examples are hidden, so effects cannot be seen on the page.
  • All effect descriptions are per the author's claims.
  • The card image is an AI-generated diagram, not the actual model output; the author has no publicly available standard-level examples.

Listed Only (2)

  • KinetiX Combat Semantic Bridge (MiniMax-H3) · v1.1 — LoRA: KinetiX Combat Semantic Bridge v1.1: the author claims it enables H3 to absorb real fighting movements, with this version expanding the training set. Video 608x352
  • Bertie Wells · v1.0 H3 — LoRA: Character LoRA for the original character Bertie Wells (a "underground person" girl with pale skin and large eyes), ported from Krea 2 to H3. Video 640x960

Data: Civitai Public API (base model MiniMax H3), inclusion window 11-12 09:03 → 11-12 21:01 (Beijing Time). Recommendation reasons have been editor-reviewed; download counts reflect attention only.

Civitai

Explore the latest AI Radar updates