Boreal-H3

Architecture Overview:
• Post-trained from MiniMax H3, inheriting its unified generation of video with native speech, sound, and multimodal reference understanding
• Three workflows: text-to-video across 6 aspect ratios, image-to-video with first and optional last frame, and reference-to-video with up to 9 images, 3 video clips, and 3 audio clips
• Output at 768p, 1088p, and 2K

Training Methodology:
• Post-trained on Creatify's Recursive Self-Improvement loop: a closed feedback system that diagnoses failures, matches interventions across data, weights, and serving, and gates every candidate behind blind A/B comparison before release
• Evaluation decomposes advertising quality into six axes: brief execution, product integrity, creator identity, realism, motion, and continuity
• Combines targeted data collection, supervised fine-tuning, and reinforcement learning, including a GRPO adaptation for the H3 architecture
• Refined on real advertising briefs, with recurring failures setting the next round's training priority

Performance Characteristics:
• All figures from Creatify's internal evaluation of the complete advertising workflow against the H3 base on matched difficult production briefs
• Brief completion: 50.0%, up from 27.8% for the base
• Subject consistency: 94.4%, up from 83.3%
• Visible defects: 0.33 per clip, down 70% from 1.11
• End-to-end acceptability: 69.0, up from 43.7
• Reference fidelity: 85.3% of ratings scored 7 or higher out of 10 for preservation of the reference

‍

Together AI API Access:
• Access Boreal-H3 via Together AI APIs using the endpoint creatify/Boreal-H3
• Authenticate using your Together AI API key in request headers
• Generate from a text prompt, from an opening frame with an optional ending frame, or from a reference set of up to 9 images, 3 video clips, and 3 audio clips, then poll until the video completes
• Speech and dialogue are generated natively, with accents inferred from references and context
• Available on Together AI dedicated infrastructure

‍

Product & E-Commerce Ads:
• Keep packaging, labels, and product geometry accurate through reveals, close-ups, and camera moves
• Turn multiple product views into one coherent 360-degree shot from reference images
• Render transparent packaging, liquids, and cause-and-effect product demonstrations believably

UGC & Creator Content:
• Produce creator-style ads and testimonials with natural spoken performances
• Keep the same person, outfit, and framing from an approved portrait through the clip
• Run two-person conversations with dialogue and accents that fit the scene

Campaign Variants at Scale:
• Direct camera angles, movement, and scene order per brief across many variants
• Localize ads across languages while holding the product and creator constant
• Iterate toward a winning cut with controlled endpoints instead of rerolling from scratch

‍