slate-ai 0.1.0

Don't ship broken AI animation. Multi-VLM verdict for rendered frame sequences.

Slate

Don't ship broken AI animation.

Slate is a multi-VLM verdict service for rendered animation. Give it a frame sequence and a manifest of what should be in it. Get a PASS/FAIL verdict backed by a quorum of local + cloud vision-language models, with a structured signal report.

Built for indie animators, virtual production teams, AI-cinematic creators, and ML training-data curators who can't keep eyeballing 200 shots.

What Slate catches that AI render tools don't

• Character lying horizontal, floating mid-air, or stuck in T-pose
• Character orientation doesn't match movement direction
• Ground contact missing — character hovering above the floor
• Wrong character identity (the manifest said hero A but hero B rendered)
• Debug-pink checker materials, missing textures, all-black frames
• Lighting / composition / atmosphere quality below a configurable threshold
• Missing landmarks the manifest required to be visible

Quick start

pip install slate-ai

# Make sure Ollama is running locally with gemma4:latest, or set
# NVIDIA_API_KEY for cloud verdicts.

slate verify \
  --frames ./my_render \
  --manifest ./shot.json

Output (truncated):

{
  "verdict": "FAIL",
  "shot_id": "village_walk_001",
  "providers_consulted": ["gemma:local", "nvidia:nemotron-nano-12b-v2-vl"],
  "failures": [
    {"signal": "character_orientation", "value": "lying_horizontal", "frame": "frame_0360.png"},
    {"signal": "ground_contact_visible", "value": false, "frame": "frame_0001.png"}
  ],
  "quality_scores": {"lighting": 5, "composition": 4, "atmosphere": 3, "mood": 4, "coherence": 4}
}

See docs/quickstart.md for a full walkthrough.

How it works

Slate samples representative frames from your render (first/middle/last by default; configurable per manifest), asks each configured VLM provider a structured set of questions, fuses the answers into a single verdict, and writes a JSON report.

frames + manifest
       |
       v
+---------------+    +-------------------+    +-----------------------+
| frame sampler | -> | VLM provider(s)   | -> | signal fusion + verdict|
+---------------+    +-------------------+    +-----------------------+
                          |
                          +-- gemma (local, Ollama)
                          +-- nvidia (BYO API key)

Frames stay on your hardware. Slate Core never uploads them anywhere you didn't configure.

Editions

Edition	License	What you get
Core (this repo)	MIT	CLI, multi-VLM quorum (local Gemma + BYO NVIDIA primary + cross-check), signal set, schema validation, JSON output
Pro	Commercial	+ Panel adversarial agent (catches "passes VLMs but still looks like AI slop"), evidence bundle generator, web dashboard
Studio	Commercial	+ team seats, audit history, Slack/Discord integration, custom signal sets
Enterprise	Commercial	+ self-hosted, SSO, SLA

Pro / Studio / Enterprise are separate commercial offerings. Core is and will remain MIT.

Installing the VLM backends

Slate Core ships drivers for two backends. You configure one or both.

Local Gemma (free, no API key)

# Install Ollama (https://ollama.com)
ollama pull gemma4:latest
ollama serve  # listens on localhost:11434

NVIDIA NIM (cloud, BYO key)

export NVIDIA_API_KEY=nvapi-...
# Slate uses nvidia/nemotron-nano-12b-v2-vl as primary and
# meta/llama-3.2-90b-vision-instruct as cross-check by default.

You pay NVIDIA directly. Slate never sees your key or your frames.

Project status

Pre-release. v0.1 carves the validation logic out of a working production pipeline (JonsStudio dogfood) and packages it for general use. See CHANGELOG.md.

Contributing

Issues and pull requests welcome. See CONTRIBUTING.md once it exists.

License

MIT — see LICENSE.