Zorbe

Image Distillation

Your house style, at one to four steps and a fraction of the cost — a model you own, not an API you rent. Distilled on your catalogue, verified by our judge model, deployed your way.

Your house style in one to four steps

Guidance is where the money goes. The teacher runs the full schedule and pays twice at every step to hold a prompt; a student distilled on your catalogue and your brand adapters folds that guidance into its weights and lands the same image in one to four steps. The product shots and campaign assets you generate by the thousand cost a fraction of that, and still look like you.

The architecture

Six stages, from your own data to a model that belongs to you — and a loop that keeps it current as your traffic moves.

01corpus 02teacher 03objectives 04student 05verify 06serve your catalogue phash dedupe VLM recaption held-out slice teacher · DiT/UNet full schedule + guidance ODE solver pairs guided predictions frozen teacher score harvested once, cached LCM consistency distribution matching adversarial (ADD) reflow + CFG distill student · 1–4 steps pruned + int8, brand LoRA pruned + int8 brand LoRA what actually ships judge model judge vs teacher prompt adherence safety + canary release gates — any one of these blocks the ship hosted or in VPC instant rollback batch + realtime picks flow to corpus
  1. 01

    From your catalogue to training pairs

    We ingest product shots, campaign assets and past generations with the prompts that made them, drop near-duplicates by perceptual hash and embedding distance, cluster on CLIP embeddings to see what the library covers, recaption with a VLM, and hold out a stratified slice per cluster.

    perceptual-hash dedupeCLIP embedding clustersVLM recaptioningaesthetic pre-filterheld-out eval slice
  2. 02

    The expensive pass runs once

    Your teacher — SDXL, SD3 or FLUX-class, or your own fine-tune — samples the full schedule under guidance while we harvest solver trajectories, noise–latent pairs for reflow and CFG-combined predictions. Latents and embeddings cache beside them; the frozen teacher stays loaded for the online terms.

    CFG-combined targetssolver trajectoriesnoise–latent pairscached VAE latentsfrozen teacher score
  3. 03

    Four ways to score the same image

    Consistency alone softens texture; an adversarial term alone narrows diversity. So they run together — latent consistency for the step collapse, distribution matching between teacher and learned student scores, an adversarial head for detail, reflow to straighten paths, guidance distilled in.

    LCM consistency lossdistribution matchingadversarial (ADD/LADD)rectified-flow reflowguidance distillation
  4. 04

    Small enough to run at volume

    Depth and width come down, attention runs in fused kernels, weights land at int8 — fp8 where the hardware has it — and the text encoder is quantised too. A distilled VAE decoder keeps pixels from becoming the new bottleneck. Brand LoRAs are refit on the student, merged in or swappable per line.

    block pruningint8 / fp8 weightsdistilled VAE decoderbrand + style LoRAsswappable adapters
  5. 05

    The bar an image has to clear

    Every candidate runs a prompt suite from your briefs and the held-out slice. Our judge model scores it against teacher output on adherence, brand, aesthetics and diversity; a regression set re-runs past failures; likeness, trademark and NSFW filters run on every sample. One gate red, nothing ships.

    Zorbe judge modelprompt-adherence evaldiversity + aestheticsbrand-conformance setsafety + NSFW filters
  6. 06

    Where it runs, and what returns

    It ships hosted or into your own VPC behind a versioned endpoint — weights pinned, rollback one call away, overnight catalogue batches and interactive requests on the same build. The prompts you send, the variants your team picked and the ones an editor killed flow back into the corpus.

    hosted or in-VPCpinned versionsone-call rollbackpicked / killed pairscorpus refresh

aYour domain

Distilled on your own catalogue — your products, your palette, your house style — so it is right on your assets, not on everything.

bJudge-verified

The student ships only when our judge model confirms it matches the teacher on adherence, brand and aesthetics, behind canary gates.

cDeployed your way

Hosted by us or inside your own environment — a fraction of the cost and latency at catalogue volume, and the model is yours.

We distill video, speech and text models the same way — or own your intelligence today ↗.