Pika H3 Flex distilled LoRA matches 30-step quality in 6 steps, 4.8x faster on one RTX 4090
Important-Apple7422 · reddit · 2026-11-02
A Reddit user benchmarked Pika H3 Flex against 30-step base sampling with the same seed and prompt in ComfyUI:
- Speed: euler 30-step took 27.4 min; the Flex template's 6-step LCM manual sigma schedule took 5.7 min — about 4.8x faster
- Setup: RTX 4090, image-to-video from a Qwen-Image still (1344x768), 226 frames (9.4s @ 24fps)
- Quality: the comparison frames look nearly identical
The author notes the Flex LoRA is distilled, so the 6-step LCM schedule is its intended operating point. Full node-level settings (samplers, sigma sequence, VAE precision) are listed for easy reproduction.
More from Multimodal
- Kling Team's report: video generation bottleneck shifts from rendering to semantic control — KlingTeam · 2026-11-04
- "SHE IS THE CLOCK": a 35-second AI horror short about a perpetually late rabbit conductor — DFoundryHQ · 2026-11-04
- MiniMax H3 video gen can't keep characters dead: Reddit user's RTX 3090 workflow test — DirectSentence9823 · 2026-11-04
- Bullish take: Spawn 7 will convince everyone the 3D internet era has begun — TAbrodi · 2026-11-04
- AI video made with Seedance 2.5 on Higgsfield hailed as a 'masterpiece' — SimplyAnnisa · 2026-11-04
- Prompting "photorealistic" makes Flux images look more like 3D renders, Reddit test shows — Confident_Ad2351 · 2026-11-04