Creative AI

Hanzo Creative AI establishes a strict architectural separation between open foundations and proprietary studio models.

Open foundation models provide public, reproducible research under permissive licenses (Apache 2.0 / MIT) with downloadable weights on Hugging Face (huggingface.co/zenlm). Hanzo Creative models are proprietary fine-tunes designed for production studio workflows, character consistency, shot continuity, and professional audio-visual generation with private weights hosted in Hanzo Cloud.

The Model Pairings

Each proprietary creative model pairs directly with an open foundation checkpoint:

Architectural Contracts

The distinction between open foundation checkpoints and Hanzo proprietary models is governed by a strict lifecycle contract:

Proprietary Model Specifications

Koma — Cinematic Video Generation

• Role: High-fidelity video generation with shot-level creative direction and character continuity across camera cuts.
• Foundation: zen-director (Wan2.2 5B base, Apache 2.0).
• Core Architecture: Spatial-temporal diffusion transformer with dual trajectory motion control and depth-guided camera trajectories.
• Native Context: Up to 120-frame temporal coherence at 24/60 fps with 1080p native generation and 4K latent upscaling.
• Key Use Cases: Pre-visualization, commercial spot generation, dynamic episodic storytelling, virtual production plates.
• Access Route: hanzo.ai/koma · Apply for Research Access

Iro — Creative Image Synthesis

• Role: Creative art direction, brand asset design, prompt adherence, and photorealistic asset generation.
• Foundation: zen3-image (Qwen-Image / Qwen-Image-2512 base, Apache 2.0).
• Core Architecture: Multimodal cross-attention flow-matching transformer with precise typography and layout rendering.
• Key Use Cases: Brand design systems, production concept art, typography-accurate editorial posters, game assets.
• Access Route: hanzo.ai/iro · Apply for Research Access

Mori — Geometric 3D Mesh Generation

• Role: High-fidelity 3D asset creation with production-ready topologies, PBR material maps, and UV unwrapping.
• Foundation: zen-3d (TRELLIS base, MIT).
• Core Architecture: Structured 3D Gaussian splatting to dual-resolution signed distance field (SDF) meshing with automated quad-dominant remeshing.
• Output Formats: GLTF, USDZ, FBX, OBJ with albedo, roughness, metallic, and normal maps.
• Key Use Cases: Spatial computing, Unreal Engine / Unity level dressing, industrial asset prototyping.
• Access Route: hanzo.ai/mori · Apply for Research Access

Oto — Foley & Environmental Audio

• Role: Synchronized foley sound effects, ambient soundscapes, acoustic reverberation modeling, and video-to-audio matching.
• Foundation: zen-foley (MOSS-SoundEffect v2 base, Apache 2.0).
• Core Architecture: Latent audio diffusion with multi-band mel spectrogram reconstruction and frame-accurate hit-point temporal synchronization.
• Output Spec: 48 kHz / 24-bit uncompressed audio (WAV, FLAC) with spatial binaural rendering options.
• Key Use Cases: Automated game sound design, video post-production foley replacement, interactive media acoustics.
• Access Route: hanzo.ai/oto · Apply for Research Access

Koe — Expressive Speech & Voice Performance

• Role: Emotionally nuanced voice performance, character acting, multilingual accent synthesis, and vocal style transfer.
• Foundation: zen3-tts (Qwen3-TTS base, Apache 2.0).
• Core Architecture: Flow-matching autoregressive acoustic model with zero-shot reference voice cloning and prosody descriptor conditioning.
• Capabilities: Sub-second voice replication from 3-second reference audio; multi-lingual dialogue with natural pacing and phoneme control.
• Key Use Cases: Game NPC dialogue trees, cinematic voiceovers, localization, audiobook narration.
• Access Route: hanzo.ai/koe · Apply for Research Access

Uta — Multi-Track Music Composition

• Role: Structural music composition, stem generation, full-length musical arrangement, and vocal lyric synthesis.
• Foundation: zen-musician (YuE stage 1 / ACE-Step 1.5 base, Apache 2.0 / MIT).
• Core Architecture: Hierarchical acoustic codebook transformer with separate stem conditioning (drums, bass, harmonic accompaniment, vocal lead).
• Capabilities: Up to 4-minute continuous composition with structured verse-chorus progression and BPM-locked time signatures.
• Key Use Cases: Adaptive game soundtracks, advertising scoring, commercial audio production.
• Access Route: hanzo.ai/uta · Apply for Research Access

Access & Governance Lifecycle

Access to private creative models follows a mandatory multi-step approval pipeline:

┌────────────────────────┐      ┌────────────────────────┐      ┌────────────────────────┐
│  1. Model Application  │ ───► │  2. Enterprise Audit   │ ───► │ 3. Entitlement Minting │
│   hanzo.ai/research-   │      │  Intended use & volume │      │ Cryptographic access   │
│   access intake form   │      │  compliance clearance  │      │ token issued to IAM org│
└────────────────────────┘      └────────────────────────┘      └────────────────────────┘
                                                                             │
                                                                             ▼
┌────────────────────────┐      ┌────────────────────────┐      ┌────────────────────────┐
│  5. Studio Workflow    │ ◄─── │  4. Gateway Activation │ ◄────┘ Node graph enabled on  │
│  Interactive pipeline  │      │  POST /v1/creative/*   │       studio.hanzo.ai         │
│  with live latent view │      │  verified per request  │                               │
└────────────────────────┘      └────────────────────────┘

1. Intake Application: Submit an application specifying intended models and research use cases at hanzo.ai/research-access.
2. Review & Compliance: The research review board verifies institutional credentials and alignment with the Hanzo Acceptable Use Policy.
3. Entitlement Minting: An entitlement token is minted against your Hanzo IAM organization ID. Approvals are strictly scoped per model (approval for Koma does not grant access to Iro or Mori).
4. Wire-Time Enforcement: Every request to the Hanzo Platform API verifies the cryptographic entitlement before inference begins; unauthorized calls fail at the ingress layer.
5. Studio Node Activation: Models immediately unlock as native nodes inside Hanzo Studio for multi-node visual orchestration.

Hanzo Studio Integration

Creative models operate as first-class nodes in Hanzo Studio (studio.hanzo.ai), allowing developers and creators to link generative media with foundation models and deterministic Decision Programs:

# Conceptual Hanzo Studio DAG Pipeline
from hanzo.studio import Workflow, Node

wf = Workflow(name="Cinematic Trailer Generation")

# 1. Foundation reasoning generates character bible and prompt sequence
concept = wf.add(Node.chat("zen6", prompt="Generate 3-shot scene beats for sci-fi film"))

# 2. Decision node selects visual art style
style_choice = wf.add(Node.kai("style-decision", options=["cyberpunk", "solarpunk", "brutalist"]))

# 3. Creative Image model generates reference keyframes
keyframe = wf.add(Node.creative("iro", prompt=concept.out["shot_1"], style=style_choice.verdict))

# 4. Creative Video model generates cinematic motion
shot = wf.add(Node.creative("koma", image=keyframe.image, duration_frames=72))

# 5. Creative Audio models generate voice performance and foley
voice = wf.add(Node.creative("koe", script=concept.out["dialogue_1"], voice_ref="narrator_v2"))
foley = wf.add(Node.creative("oto", video_track=shot.video, mood="tension"))

# 6. Safety evaluation node checks content boundaries
guard = wf.add(Node.guard("zooai/lifeguard", video=shot.video, audio=voice.audio))

Safety & Content Standards

All Hanzo Creative model runs are evaluated against consequence-grounded safety standards:

• ZooAI Lifeguard (zooai/lifeguard): Multi-modal guardrail evaluating generated media for physical safety, non-consensual imagery, and policy boundaries.
• Provenance & Watermarking: Cryptographic provenance metadata (C2PA standard) is embedded into all generated video, audio, and images to ensure verifiable attribution.
• Deterministic Override: If a safety node issues a deny verdict, the media pipeline halts instantly and resources are quarantined.