MiniMax H3 – Beauty & Clarity Enhancer LoRA

This LoRA is designed to improve the overall visual quality of MiniMax H3 video generations, producing sharper, cleaner, and more visually appealing results.

It helps reduce unwanted blur, improve facial details, enhance the appearance of people, and clean up backgrounds for a more polished, professional look.

What It Improves

• Sharper Details: Improves overall image clarity and reduces softness and unwanted blur.

• Beauty Enhancement: Makes characters more attractive with refined facial features and a more flattering appearance.

• Cleaner Backgrounds: Helps reduce muddy, blurry, or messy background details.

• Improved Visual Quality: Gives videos a cleaner, more polished aesthetic while maintaining a realistic appearance.

Available Versions

Epoch 5 – Lighter Enhancement

A lighter version that provides gentle sharpening, mild beauty improvements, and cleaner backgrounds without dramatically changing the original look.

Epoch 10 – Stronger Enhancement

A more noticeable enhancement with stronger sharpening, more pronounced beauty improvements, and cleaner overall visuals.

Recommendations

Start with Epoch 5 if you prefer a more natural look. Use Epoch 10 if you want a stronger beauty and clarity effect. If you want a even lighter effect try a lower strength. All examples video's have a strength of 1. Start at 0.5 and work your way up till you get the effect you want.

Results may vary depending on your prompt and generation settings.

The example videos were made with a 2-pass workflow using these settings:

Models

• Base model: MiniMax H3 FL2VA (minimax_h3_fl2va_pruned_int8_convrot)

• Text encoder: qwen3vl_32b_minimax_h3_nvfp4_awq

• VAEs: minimax_h3_video_vae_fp16 and minimax_h3_audio_vae_fp32

Pass 1 (generation)

• Enhancer LoRA at strength 1.0 (pass 1 only)

• Resolution: 960 × 544 (half the final size)

• Steps: 20

• Sampler: res_multistep

• Scheduler: simple

• Denoise: 1.0

• Guider: BasicGuider (no CFG, no negative prompt)

Latent upscale

• Learned 3D latent upscaler (minimax_h3_latent_upscaler_3d_conv_v1_bf16), ×2

Pass 2 (refinement)

• Turbo LoRA: minimax_h3_fl2v_turbo_8step_v1.0_768p at strength 1.0 (pass 2 only, no Enhancer)

• Steps: 4

• Denoise: 0.35

• Sampler: res_multistep

• Scheduler: simple

Output

• 1920 × 1088 (vertical clips 1088 × 1920), 24 fps

• Clips are 6–8 seconds long

• Audio: MiniMax H3's built-in audio, decoded from the pass 1 latent

Before/after comparisons

• Each pair uses the same prompt and seeds. The only difference is the Enhancer LoRA on pass 1.