Pika H3 Flex: 380% faster than 30-step H3, same seed. Spot the difference?
retained text
Playing around with Pika and running some benchmarks. Same input still, same prompt, same seed. Only the sampling settings change. \*\*Setup (both runs)\*\* \- Base: minimax\_h3\_fl2va\_pruned\_int8\_convrot \- LoRA: pika\_h3\_flex\_fl2va\_v1\_bf16 @ 1.0 \- Text encoder: qwen3vl\_32b\_minimax\_h3\_nvfp4\_awq \- Audio VAE: minimax\_h3\_audio\_vae\_fp32 \- Image-to-video from a Qwen-Image 2.1 still, 1344×768, 226 frames (9.4s @ 24fps) \- RTX 4090, ComfyUI 0.43 \*\*Top: 30-step base sampling\*\* \- Sampler: euler, simple scheduler, 30 steps \- ModelSamplingMiniMaxH3: shift 12 video / 3 audio \- Video VAE: minimax\_h3\_video\_vae\_fp16 \- BasicGuider (cfg 1) \- Render time: 27.4 min \*\*Bottom: Pika H3 Flex template settings\*\* \- Sampler: lcm, ManualSigmas (1.0, 0.973, 0.923, 0.8, 0.571, 0.333, 0.0) = 6 steps \- No sigma shift node \- Video VAE: minimax\_h3\_video\_vae\_int8\_convrot \- BasicGuider (cfg 1) \- Render time: 5.7 min The Flex LoRA is distilled, so the template's 6-step lcm schedule is what it's built for. That's about 4.8× faster. Which one did it better? Also posted on X: [https://x.com/CreaitiveLabs/status/2116988711992975434](https://x.com/CreaitiveLabs/status/2116988711992975434)
document fetched 2026-11-02 00:36 · canonical https://reddit.com/r/comfyui/comments/1xldfma/pika_h3_flex_380_faster_than_30step_h3_same_seed