CFG > 1 still noticeably improves H3 model quality despite distillation
Murky-Relation481 · reddit · 2026-08-27
Despite being distilled using CFG, testing shows that setting guidance values greater than 1 and using negative prompts still noticeably improves prompt adherence and output quality for the H3 model, especially at lower resolutions. However, values exceeding 3 or 4 lead to over-burning artifacts.
More from Multimodal
- Detailed prompt breakdown for generating photorealistic SEEDANCE video — SimplyAnnisa · 2026-08-27
- Live, Camera, Fashion - AI MV using MiniMax H3 Lip-sync and Dance — Inner-Perspective382 · 2026-08-27
- OmniColor: A Unified Framework for Multi-modal Lineart Colorization (ECCV 2026) — 机器之心 · 2026-08-27
- Flova x Seedance 2.5 turns a simple idea into a complete film — hey_abusiddik · 2026-08-27
- AI Reimagines Magritte for the 2020s — joshua_saxe · 2026-08-27
- Dance of the Wraith — a two-act AI short film, made solo — Gianniarrenzetti · 2026-08-27