Combat BASE | Combat Action/Impact/Dramatic Enhancement/Combat BASEV2

Combat Base V2 is a major upgrade over V1, focusing not only on faster combat, but also on better action coherence, hit reactions, cumulative damage, spatial movement, and physical causality.

Combat Base V2 is no longer just a combat LoRA; it is now suitable for action and dialogue scenes, adding richer physical movements, finer audio details, and stronger inter-shot continuity to character performances.
Combat Base V2 is no longer just a martial arts LoRA—it now handles both dialogue and action, enhancing physical performance, guided voice expression, and stronger shot-and-state continuity even in dramatic scenes.

If the LoRA signal looks muddy or less effective than expected, please check the sampler settings first.

For LoRA, the choice of sampler is very important.

For optimal clarity and dynamic effects, I recommend one of the following settings:

Option A

• Sampler: res_multistep

• Scheduler: Simple

Option B

• Sampler: euler

• Scheduler: beta

Other sampler/scheduler combinations may produce noticeably blurrier or softer results.

[MiniMax H3] 12GB VRAM | Universal FL2VA/REF2VA Base Model + Low Sigma Combat Second Pass Workflow

Core Physics & Motion Logic / Core Physics & Action Logic

• Combat Base V2 abandons the "attack → reset" loop, opting instead for a continuous physics chain where missed attacks carry momentum, blocks smoothly redirect, and limbs can cleanly separate without sticking.

• Combat Base V2 flattens the "attack → reset" cyclic loop into a continuous motion chain, where missed attacks retain forward momentum, blocks naturally redirect, and contact points detach cleanly without sticking.

• Hit reactions now persist and accumulate—continuous torso bending, lowered defenses, and restricted mobility directly lead to heavy strikes.

• Hit states have persistence and cumulative effects; damage such as torso bending, slower guards, and movement constraints are naturally retained, making the final crushing blow more reasonable.

Combat Techniques & Stage Setup / Action Techniques & Shot Framing

• Upgrades include continuous spin kicks, sustained grappling and takedowns, dynamic three-stage environmental collisions, and multi-stage fall trajectories.

• Comprehensive upgrades include continuous spinning legs, sustained close-quarters grappling and throws, three-stage environmental collisions (body → environment → material), and stunt fall performance with complete weight scaling.

• Supports cinematic 1vMany choreography through staggered attacker entries, spatial corridor control, protagonist-centric continuous tracking, and motivated cross-cutting action continuity.

• Cinematic "sparse-multiple" siege direction, utilizing alternating cuts, environmental channel occlusion, protagonist-centered close tracking support, and cross-shot action continuity to enhance visual performance.

Dialogue Performance & Benchmarks / Dramatic Performance & Test Data

• Internal same-seed dialogue tests show a 19.5% increase in intra-shot motion, a 34.2% improvement in audio RMS, and a 22.4% increase in edge clarity, significantly enhancing physical performance, breathing, and vocal dynamics in non-combat scenes.

• Same-seed dramatic tests show a +19.5% increase in in-shot motion, a +34.2% improvement in audio RMS, and a +22.4% increase in edge sharpness, greatly enhancing physical performance, breathing, and vocal dynamics outside of combat sequences.

Workflows and Trigger Usage / Workflows & Trigger Word Usage

• For optimal fluidity, set the prompt time structure to "action → dialogue"; start with "no triggers" for maximum anatomical accuracy, add the trigger prfight2 for higher intensity, or prfight2, prfin1 for intense stunt impacts.

• It is recommended to set the prompt structure to "action → dialogue"; prioritize no triggers for the most stable human anatomy, add prfight2 when high aggressiveness is needed, and use prfight2, prfin1 for heavy finishers/stunt hits.

• Includes two specialized combat director workflows (general and 1vMan) with Chinese instructions, supporting rapid output in any target language.

• Comes with two dedicated martial arts director workflows ("General" and "Multi-opponent"), built-in Chinese instructions, and supports generating final prompts in any language.

Direct the combat first. Write the prompt second. Direct the action first, then write the prompt words.

Upgrade Notes:

Test Finding: When a single shot contains both combat and dialogue, the combat actions and dialogue must be fully completed first, otherwise it easily reverts to the base model's slow-motion performance.

Test Finding: When combat and dialogue occur in the same shot, please complete the attacks before the dialogue, otherwise the action may revert to the base model's slower behavior.

About this LoRA | About this LoRA

First of all: This is not a magical LoRA, nor will it automatically choreograph complete combat actions for you.
First: This is not a magical LoRA; it will not automatically choreograph complete combat actions for you.

If you only write very simple prompts like two people fighting or a woman fights a man, the model will likely just make two people wildly swing punches, resulting in a fast but illogical "windmill punch" sequence.
You still need to tell the model who attacks first, what they do, where the attack lands, how the opponent dodges or reacts, and why the next attack can naturally follow.

The prompt is responsible for deciding "how they fight", while LoRA enhances "how these actions are performed".
The prompt decides what the fighter does, while LoRA enhances the execution of those actions.

This LoRA combat confrontation has been trained and tested across multiple versions, experimenting with different data ratios, combat speeds, hit reactions, and finisher training directions.
This LoRA combat has gone through multiple training and testing versions, featuring different dataset balances, combat speed behaviors, hit reactions, and finisher-focused training.

Some older versions did make the combat noticeably faster, but at high weights, they easily destroyed H3's inherently good motion fluidity, causing videos to appear sped-up, blurry, or like a continuous chaotic brawl without recovery phases.
Some early versions could make combat noticeably faster, but at higher weights, they could also compromise H3's native fluidity, producing accelerated motion, blurriness, or endless flurries without proper recovery.

The version released currently is, through my actual testing, relatively the most balanced and easiest-to-use version.
The version released here is currently one of the most balanced and practical versions in my testing.

Instead of blindly chasing faster combat, it focuses on continuous combat, target feedback, physical imbalance, recoil, stagger, knockdown, finisher logic, and the actual results produced by attacks.
Rather than simply making combat faster, it focuses more on continuous combat, impact reactions, loss of balance, recoil, staggering, knockdown, finisher logic, and clearer physical consequences following attacks.

My goal is to preserve H3's already good motion continuity,