Pika H3 Flex Release: Six-Step Video Generation Model with Open-Source Weights and ComfyUI Template
Text to video
Pika releases H3 Flex, which post-trains and distills MiniMax H3 to six-step sampling, with open-source weights and ComfyUI template. The model loads onto the H3 base with two LoRA adapters, supporting text-to-video, first/last frame, and reference-condition generation, outputting 480p/768p, 5-15 seconds, and mixing in 32kHz stereo audio. Pika's self-evaluation shows it leads comparison models on five visual and audio quality metrics; community testing suggests high-frequency detail flickers at four steps, recommending keeping six steps.
Pika post-trains MiniMax H3 into a six-step model and open-sources the weights. The ComfyUI template can be loaded directly, allowing creators to evaluate the quality of six-step output accordingly.
Open original comfyui-wiki.com