MiniMax H3 workflow: REF2VA visuals + FL2VA audio + LightX2V speed, 198s on RTX 5090
gabxav · reddit · 2026-09-05
After testing MiniMax H3 since release, gabxav found REF2VA and FL2VA each excel at different things and shares a ComfyUI workflow combining their strengths.
Findings
- REF2VA: noticeably better visuals — natural skin texture, lighting, environments.
- FL2VA: over-smooth "AI look," but handles high-motion characters better, and its voice/audio cloning is significantly cleaner (less echo, noise, artifacts).
The workflow
- REF2VA for main video generation (quality)
- FL2VA for audio refinement
- LightX2V 8-step LoRA for a major speed boost
- H3 AudioRefine node to pass generated audio through FL2VA for cleaner sound
All reproduction assets are shared: H3 models and LoRA on Hugging Face, workflow JSON, voice/image reference files, KJNodes and AudioRefine. Reference performance: 198.4 seconds per generation on an RTX 5090.
More from Multimodal
- Voxtral 3 ships a LiveKit-based voice agent SDK streaming listen-think-speak — MistralAI · 2026-10-21
- Voxtral Mini TTS 3 hits SOTA expressiveness in 35+ languages with tag-based control — MistralAI · 2026-10-21
- Mistral launches Voxtral 3 speech models and new Voice Agents API — MistralAI · 2026-10-21
- Open-sourced pixel art normals generator ships as Rust CLI and wasm library — HugoDuprez · 2026-10-20
- Fibo 2.0 is strongest for ecommerce and creator content, weakest at UIUX design — ArtificialAnlys · 2026-10-20
- Fibo 2.0 leads old Fibo on all 9 capabilities, best at reasoning and anatomy — ArtificialAnlys · 2026-10-20