Pika launched H3 Flex on October 29, releasing model weights and code for a video-generation system developed from MiniMax H3.
The company says the model is available through its API, ComfyUI and Diffusers, with weights on Hugging Face and code on GitHub. Its consumer creation product is scheduled to add the model November 2, according to Pika’s separate launch announcement.
H3 Flex supports text, image and reference-based generation, along with editing tasks such as movement transfer, character changes and selective revisions.
Pika describes a workflow that generates lower-resolution drafts and uses a separate finishing model for output up to 4K. It also separates generated sound into dialogue, score, sound effects, foley and ambience tracks for further editing.
The company’s technical report explains how post-training and distillation reduced generation from 50 denoising steps to six. Performance comparisons in that report are Pika’s own evaluations, rather than independent validation of every claim.
The researchers identify physics, consistency across multiple scenes and complex reference control as areas requiring further improvement.
Sources: Pika — October 29 Technical Report: https://pika.art/blog/pika-h3-flex-research Pika — October 29 Launch Announcement: https://www.streetinsider.com/Business%2BWire/Pika%2BLaunches%2BPika%2BH3%2BFlex,%2Ba%2BVideo%2BModel%2BBuilt%2Bfor%2BProfessional%2BIteration/27256853.html