NOW IN PRIVATE BETA · Q3 2026

Compress without compromise.

Vidadhati AI runs your ZIP archives, images, and video through an adaptive transformer-based entropy model — shaving up to 87% off file size while keeping perceptual quality byte-locked to the source.

▼ See benchmarks
vidadhati-ai>
0%// Avg. size reduction
0.000// SSIM vs. source
0ms// p99 latency / MB
0M+// Assets processed
// CAPABILITIES

Three media · One engine

A single endpoint handles archives, images, and video. The model picks the right codec, the right block size, and the right perceptual weighting per region — automatically.

◉
[ MODE / 001 ]

Archive compression

Re-encodes ZIP, TAR, 7z, and rar archives with context-aware entropy modeling. Lossless on the inner files; far smaller envelope.

↓ 41–68% · against brotli-11 baseline
◎
[ MODE / 002 ]

Image fidelity codec

Per-region perceptual scaling. PNG / JPEG / HEIC / RAW in — Vidadhati AI::IMG out. SSIM floor of 0.998 against the source, or it refuses to emit.

↓ 73–87% · vs. AVIF q-65
◈
[ MODE / 003 ]

Video stream compactor

Frame-graph aware. Detects static panels, motion blur regions, and high-entropy noise; allocates bits accordingly. H.265 / VP9 / AV1 in.

↓ 52–71% · vs. AV1 CRF-30
// THE PIPELINE

Four nodes · One pass

Every asset traverses the same four-node pipeline. No upload sits on disk longer than the time it takes to verify the output is bit-faithful enough to release.

[ NODE 01 ]

Ingest

Stream-accept up to 8 GiB. Format auto-detected. SHA-256 fingerprint locked before any work begins.

› ingest factory-floor.jpg
✓ 4.2 MiB · jpeg · sha256 a1b2…
[ NODE 02 ]

Analyze

Perceptual map built per 16×16 block. Edges, texture, motion, and chroma noise scored independently.

› analyze → 18 432 blocks
✓ entropy map ready
[ NODE 03 ]

Compress

Adaptive transformer entropy coder picks a per-block codec. High-detail regions stay rich; flat regions compress aggressively.

› compress · model v4.2
✓ 4.2 MiB → 612 KiB
[ NODE 04 ]

Verify

Output is decoded and re-measured. If SSIM < 0.998 or PSNR delta exceeds budget, the run is rejected and re-tuned.

› verify · ssim
✓ 0.9994 · release
// THE NUMBERS

Measured. Reproducible. Public.

Run on the union of Kodak, JPEG-AI 2024, UVG-4K and the Silesia corpus. Same hardware (single A100, batch 1). Numbers regenerated every release — last update: 2026-05-14.

Codec Size (avg) ↓ Reduction SSIM Latency / MB Lossless?
Source (uncompressed)100.0%—1.000—✓
brotli-1143.1%56.9%1.00038 ms✓
zstd-2244.7%55.3%1.00022 ms✓
WebP q-8027.4%72.6%0.96118 ms—
AVIF q-6521.2%78.8%0.97262 ms—
HEIC q-5523.0%77.0%0.96941 ms—
Vidadhati AI v4.2 OURS 13.4% 86.6% 0.998 14 ms opt-in
// IN PRODUCTION

Built for media-heavy pipelines

The four teams already running Vidadhati AI in production share one constraint: they ship a lot of bytes, and every percent saved compounds across millions of transfers.

Live-ingest broadcasting

Re-encode the master before it hits the CDN. Cut origin egress by half without a perceptible quality drop on 4K endpoints.

  • Drop-in replacement for ffmpeg presets
  • Sub-frame latency budget
  • HLS / DASH segment-aware
  • HDR10 + Dolby Vision metadata preserved
"We swapped our AV1 ladder for Vidadhati AI in the staging tier. Egress fell 48% the first week and the QoE dashboards didn't move." — Head of media infra · global streaming platform

Product imagery at scale

Catalog teams shipping millions of SKUs per market use Vidadhati AI::IMG as the last hop before the CDN. Page weight down, SEO image scores up.

  • Bulk batch endpoint · 10k images / batch
  • Auto-srcset generation
  • Color-managed (sRGB · Display P3 · Adobe)
  • Visual diff report per release
"Largest contentful paint dropped 1.4s on mobile after we routed our hero images through Vidadhati AI. We didn't change a single template." — Performance lead · DTC marketplace

Game asset bundles

Texture atlases, audio banks, level archives — all compressed at build time. Smaller patches, faster cold installs, identical runtime behavior.

  • Reproducible builds (deterministic seed)
  • Engine plug-ins · Unity · Unreal · Godot
  • Console-cert compliant outputs
  • Delta-patch friendly
"Our patch size went from 4.1 GiB to 1.3 GiB on the same content. Steam users noticed before we even announced it." — Build engineer · AA studio

Cold archive & long-haul backup

For tier-3 storage where every gigabyte is recurring spend. Lossless mode only — Vidadhati AI::ARC re-encodes inner files, keeps the container honest.

  • S3 / GCS / Glacier compatible objects
  • Sealed container · tamper-evident header
  • Streaming verify · no full unpack required
  • Audit trail per object
"We re-compressed two petabytes of cold backups. The Glacier bill came down by 38% the following month." — Platform engineering · fintech
// ACCESS TIERS

Pricing

Pay only for what you compress. No charge for re-runs. No charge for assets the model rejects.

[ TIER / 01 ]
$0 / forever
For trial workloads & open-source projects.
  • 10 GiB / month processed
  • All three modes (archive · image · video)
  • Community support · Discord
  • Private model snapshots
  • SLA
RECOMMENDED
[ TIER / 02 ]
$0.04 / GiB
For production pipelines moving real volume.
  • Unlimited throughput
  • Bulk batch endpoint · 10k / batch
  • Webhook · S3 · GCS event sources
  • SOC 2 type II · ISO 27001
  • Pinned model version
  • 24h support response
[ TIER / 03 ]
Contact
For media platforms operating at petabyte scale.
  • Dedicated inference cluster
  • Custom model fine-tuning
  • On-prem appliance · air-gap option
  • 1h support · named SRE
  • Annual commit · invoicing
// QUESTIONS

Frequently asked

Is the lossy mode actually visually lossless?
The default mode rejects any output that falls below SSIM 0.998 against the source — typically the threshold at which a trained eye cannot distinguish original from output at 100% zoom. For audit-critical work, switch to --lossless and bit-identity is guaranteed (with a smaller compression win — typically 35–55%).
What happens to my data after it's processed?
Source assets are processed in volatile memory only and never persisted. The compressed output is written to an ephemeral object tied to your account and purged on download or after 24 hours, whichever comes first. We don't train on customer data — period. The model is trained on a curated public corpus only.
How is this different from AVIF or AV1?
AVIF and AV1 are static codecs — one set of trade-offs, baked in. Vidadhati AI runs an entropy model that picks parameters per block, then verifies the output meets a perceptual floor. The result: better compression at the same perceptual quality, or matched compression at higher quality. The benchmark table shows the head-to-head.
Is it real-time?
For images and short video, yes — p99 latency of 14 ms per MB on the standard tier. For long-form video the pipeline is streaming (you can start consuming the compressed output before the source finishes uploading), but full-clip latency obviously scales with clip length.
Self-hosted option?
Tier 03 only. The on-prem appliance ships as a sealed container with the model baked in, runs on A100 or H100, and exposes the same HTTP API as the hosted service. Air-gap deployments verified for two government customers in 2025.
What about audio?
Audio support is on the public roadmap for Q4 2026 — a perceptual audio codec built on the same entropy backbone. Currently, audio tracks embedded in video are handled as-is (passed through without re-encoding).

Enter the Grid.

Private beta is open by invitation. Drop your work email — we onboard a small cohort each week.