▲ multi-VLM render QA

Don't ship broken AI animation.

Slate gives a frame sequence and a manifest of what should be in it to a quorum of vision-language models — and returns a PASS/FAIL verdict with a structured signal report. Stop eyeballing 200 shots.

Start free ★ View on GitHub
$ pip install slate-ai  ·  local Gemma or cloud VLMs  ·  MIT core
The gap

Exit code 0 is not a verdict.

Your render farm reports success and writes 720 frames. Then you scrub the playback and the hero is lying horizontal, the scene is all black, or the wrong character rendered. File-existence and exit codes can't see any of that. A vision model can.

before“720 frames written, exit 0” — ship it. (it was broken)
with slate“FAIL — character not upright, no ground contact, frame 360” — caught before publish.
How it works

Three inputs. One trustworthy verdict.

No phone-home. Run it locally against Ollama, or add cloud VLMs (your own API keys) for a cross-checked quorum.

01 — describe

Write a manifest

Declare what the shot should contain: characters, wardrobe, landmarks, quality thresholds. Plain JSON, per shot.

02 — verify

Quorum of VLMs

Local Gemma + cloud NVIDIA models independently analyze sampled frames and emit structured signals — frames downscaled for speed.

03 — decide

PASS / FAIL + report

A signed, structured verdict with per-signal failures and quality scores. Exit codes for CI. Block bad publishes automatically.

Detections

What Slate catches that render tools don't

✕Character lying horizontal, floating mid-air, or stuck in T-pose
✕Orientation doesn't match the direction of movement
✕No ground contact — character hovering above the floor
✕Wrong identity — manifest said hero A, hero B rendered
✕Debug-pink materials, missing textures, all-black frames
✕Lighting, composition, or atmosphere below your threshold
✕Missing landmarks the manifest required to be visible
✕Wardrobe / hair / head-covering absent when required
Slate Pro

A red team that asks: “would I publish this?”

Core gives you the model verdict. Pro adds Panel — an adversarial ensemble of expert personas that independently critique each sampled frame. Any critical flag blocks the publish. Every run can emit a single-file evidence bundle for audit. Panel runs on your own Anthropic API key — you pay Anthropic directly.

DirectorColor GraderAnimatorAudience
Pricing

Start free. Upgrade when bad renders start costing you.

Core is open source and free forever. Paid tiers unlock the Panel red team, evidence bundles, team seats, and enterprise controls. Pay once — a perpetual license, yours forever. No subscription.

Bring your own models. Every tier runs on your own keys — local Gemma via Ollama (free), plus your own NVIDIA / Anthropic keys for the cloud + Panel lanes. You pay the model providers directly; Slate is the software, and it never proxies your keys or your frames.

Core

$0
Open source. Self-host. Bring your own models.
  • ✓ Multi-VLM verdict (local Gemma + cloud NVIDIA)
  • ✓ Structured signal report + quality scores
  • ✓ Manifest-driven, CI-ready exit codes
  • ✓ Unlimited frames, fully offline-capable
pip install slate-ai

Pro

$149 one-time
For solo creators who can't keep eyeballing shots. Pay once, keep it forever.
  • ✓ Everything in Core
  • ✓ Panel 4-persona adversarial review
  • ✓ Evidence bundles (audit-ready .tar.gz)
  • ✓ Email support
Buy Pro — $149

Studio

$499 one-time · 5 seats
For teams shipping episodic or client work. One purchase, five perpetual seats.
  • ✓ Everything in Pro
  • ✓ Panel 8-persona + custom personas
  • ✓ Audit history
  • ✓ Team seats
Buy Studio — $499

Enterprise

Custom
For studios with security + scale requirements.
  • ✓ Everything in Studio
  • ✓ Vertical persona packs
  • ✓ Self-hosted deployment
  • ✓ SSO & SLA
Contact sales
Get started in 60 seconds

Your next bad render shouldn't reach the timeline.

# install the free core
$ pip install slate-ai

# verify a render against a manifest
$ slate verify --frames ./my_render --manifest ./shot.json
Start free