# Flat Vector / Game-Art Style LoRA for MiniMax H3

A style LoRA for MiniMax H3 focused on clean vector-inspired game art, polished flat illustration, modern anime-influenced character design, rich gradients and strong readable shapes.

The training set was built from roughly 100 style-focused images covering characters, environments, romance scenes, action poses, fantasy, portraits, architecture and game/visual-novel-style compositions. The goal is not one single character or one fixed scene type, but a reusable visual language.

## Activation / Prompt Wording

This LoRA was trained with the original Krea-era caption wording, not with a newly invented MiniMax-H3-specific trigger token.

Use this early in style_and_tone:

Vector flat design illustration matching the reference style

For stronger style adherence, reinforce it with some of the recurring training language:

32-bit true color, rich gradient shading, premium flat design anime features, clean vector artwork, smooth cel-shading, polished flat aesthetic, professional flat design meets modern anime style

You do not need to repeat every phrase every time. Start with the core activation wording, then add the descriptors that matter for the scene.

## How I Recommend Prompting It in MiniMax H3

Keep the visual style in style_and_tone.

Example:

style_and_tone: Vector flat design illustration matching the reference style, 32-bit true color, rich gradient shading, clean vector artwork, smooth cel-shading, polished flat aesthetic, professional flat design meets modern anime style.

Then describe the actual scene separately in continuous_shot or multi_shot_sequence.

This helps keep the LoRA responsible for the visual identity while MiniMax H3 handles motion, camera movement, acting and timing.

## Good Things to Try

The LoRA is especially fun for:

- visual-novel and Ren'Py-style scenes

- romantic game cutscenes

- fantasy RPG party scenes

- JRPG combat and boss battles

- cyberpunk characters

- modern lifestyle scenes

- stylized commercials

- character portraits

- architecture and environment shots

- music-video concepts

- game trailers and animated trailer-style sequences

## MiniMax H3 Prompting Tips

For H3, I recommend separating the prompt into clear jobs:

- style_and_tone = visual look

- continuous_shot or multi_shot_sequence = action, camera and timing

- audio_design = music, ambience and sound effects

- write spoken dialogue explicitly when dialogue matters

For dialogue scenes, state who speaks and include the exact line.

For trailers, ads and game sequences, multi-shot timing works very well because each shot has a clear purpose.

For music-led videos, describe the music and its overall development, but avoid assigning sung lyric lines to strict time segments. Let the audio act as the master timeline while shot and camera cuts can still use timecodes.

If the result drifts too far toward realism, strengthen the style wording with:

clean vector artwork, smooth cel-shading, polished flat aesthetic, and professional flat design meets modern anime style.

If the result becomes too rigid or over-stylized, shorten the style section and let H3 handle more of the scene naturally.

## Examples

The attached example videos were generated with MiniMax H3 using this LoRA. I also included a separate prompt pack with 15 complete H3 Text-to-Video prompts covering film, games, romance, action, trailers, advertising and music-video scenarios.

Have fun with it, modify the prompts, and feel free to share your results. I am very curious to see what people build with the style.