I Finally Fixed MiniMax H3's Plastic Skin Problem

Summary of the Video (Subtitle-Based)

The creator demonstrates a local, model-driven workflow for generating more realistic character video—specifically targeting “plastic/waxy” skin and unnatural facial/tear realism. They compare a standard baseline workflow against a “secret workflow,” showing that the secret approach produces:

• More detailed skin texture (pores/imperfections)
• More convincing tears
• More natural facial expression/acting
• Improved consistency across different character reference sheets

They also test the workflow on a separate AI influencer-style input video.

Artistic Techniques / Creative Process Shown

• A/B comparison workflow

  • Side-by-side testing of:
    • Baseline generation using a standard workflow
    • Improved generation using the creator’s custom “secret workflow”

• Character realism focus

  • Iterating specifically to reduce uncanny valley artifacts:
    • Plastic/waxy skin
    • Flat expressions
    • Unnatural tears

• Consistency testing

  • Keeping the workflow the same while swapping the character reference sheet to check whether the “locked-in” look transfers reliably.

• Cinematic output optimization

  • Emphasis on:
    • Lighting and cinematography quality
    • Facial dynamics and expressions
    • Motion and overall scene quality

Workflow Steps and Components (As Described)

Core Model (Local)

• Miniax H3 Singularity
  • Used to control/generate cinematic detail, motion, expressions, and scene quality
  • Supports SDR/HDR, restoration, dynamics, and VFX-like outcomes

Additional Fine-Tuning / Assist Modules (LoRAs)

Two key LoRAs are used:

• LMS LoRA

  • Purpose: Recover/boost fine detail when outputs become soft or blurry
  • Effect: Sharper, crisper imagery; counteracts the tendency of video generation to trade detail for motion

• People Realism LoRA

  • Purpose: Reduce wax-like “AI skin” and push face/skin appearance toward natural realism
  • Effect: More believable human skin detail and appearance

Two-Stage Generation

• Stage 1: First Pass

  • Uses Miniax Reftov Turbo 4-step LoRA
  • Includes a tunable quality/processing parameter (example given: 0.5; higher VRAM can justify increasing it)
  • Goal: Establish foundational visuals

• Stage 2: Dual Upsampling (Quality Refinement)

  • Re-processes the first-pass output to improve final image quality
  • Optional tool: latent upscaler
    • Scaling mode: scale by multiply
    • Precision: BF-16
  • Second-stage LoRA weights (example given):
    • LMS LoRA: 0.5
    • People Realism LoRA: 1.0

Performance / Testing Framing

The creator repeatedly challenges the workflow by:

• Running the same character test under baseline vs secret workflow
• Testing a new character sheet with the same pipeline
• Performing an AI influencer-style test

Creators / Contributors Mentioned

• AI creator(s):

  • The subtitles refer to the speaker as an “old lady” in dialogue, but the channel name isn’t explicitly stated in the subtitles.
  • The creator mentions “AI brief” (referring to their channel as “AI brief”).

• Models / components mentioned:

  • Miniax H3 Singularity
  • Miniax Reftov Turbo 4-step LoRA
  • LMS LoRA
  • People Realism LoRA