Pika Flex LoRA showdown: official keeps audio quality, HD 8-step wins consistency
Reasonable-Aide1259 · reddit · 2026-11-17
After generating many 20-second videos with refv2a and various turbo LoRAs, a user compares the official Pika Flex LoRA with Pika Flex HD:
- The official pika-flex LoRA is excellent; switching from 6-step sigmas to 8 steps yields great quality
- When using video references for face/hair consistency, changing a character's outfit with the official LoRA also changes the hair — distinctive hairstyles get normalized to plain long straight hair. Quality is unreliable below 0.7MP. Keeping clothes unchanged preserves hair with high probability
- Pika-flex HD 8-step holds reference consistency without prompt modifications, but results show subtle unwanted changes, and audio quality degrades noticeably versus the official LoRA
Scheduler, sigma, prompt, and resolution changes did not break these patterns.
More from Multimodal
- AI horror short 'The Day She Arrived' built with Midjourney, Hailuo H3 and ElevenLabs — Kyrannio · 2026-11-17
- Solo dev ships H3-FaceRefine-Plus, doubling face detail on small faces with Claude Code — AnniCodex · 2026-11-17
- AI-animated Chinese fable 'The Fox Who Fooled a Tiger' in stunning ink-painting style — Vrchcab_2027 · 2026-11-17
- AI-Generated Spanish Sci-Fi Trailer 'Enki Chronicles' Gets Director's Cut — Excellent-Emphasis25 · 2026-11-17
- Krea2TextEncoder hands-on: reference-based gens with zero sampling overhead and minimal prompting — ThetaCursed · 2026-11-17
- Koyal pivots from NeurIPS audio-to-video paper to full-blown AI film company — DavidmComfort · 2026-11-17