The demo pack's entire input: a 23-second video, one dashboard photo, and a 12-second voice memo ("it's mint, just the front-right rim").
+ +Out comes the pack: exterior shots classified into a fixed template (front 3/4 hero, front, profiles, rear 3/4, rear, close-ups…), an ad preview per portal, and a spec sheet. Missing slots aren't papered over — they're shown as gaps the dealer can fill with a single extra photo. + +
Auto-filled shot template from the walkaround. Empty slots ask for one more photo instead of pretending.
+ +
The listing text, previewed in the portal's own clothes — leboncoin, LaCentrale, or the dealer's site.
+ +### The architecture: a DAG with a VLM at the judgment points + +Under the hood CarStudio is a deterministic pipeline, not an agent: + +``` +upload → ingest → frame scoring & shot classification + → data extraction (plate → SIV specs; odometer OCR) + → listing text generation (grounded) + → operator review → enhancement (selected frames only) + → packaging → ZIP +``` + +ffmpeg extracts a working set of frames; classical scoring (sharpness on the vehicle mask, exposure, framing) ranks them; a VLM is consulted only where actual judgment is required — which of the template's 13 shot classes a frame belongs to, what the plate reads, whether the listing draft needs polish. The VLM is a Qwen2.5-VL-class model served on my own 2×A6000 box behind an OpenAI-compatible endpoint, so inference is free at the margin, dealer footage never leaves the machine, and moving to a hosted endpoint later is a one-line `base_url` change. + +Two structural choices I'd defend hard: + +- **The pipeline completes even with no VLM running.** Every model call degrades to "flag this slot for the operator" instead of failing the job. The model is a guest in the pipeline, not its backbone. +- **SQLite is the whole backend** — vehicle records *and* the job queue, executed by one polling worker. At single-node MVP scale, Redis and Celery are complexity cosplay; every stage is idempotent and re-runnable per vehicle, which is the property that actually matters when a run goes sideways. + +### Compliance is encoded, not documented + +This is the part of the project I find genuinely interesting. Used-car portals have strict and slightly contradictory rules — LaCentrale's photo charter effectively excludes background cutouts, Facebook Marketplace flags studio-look images, leboncoin forbids promotional overlays — and above all of it sits the obvious integrity question: an AI pipeline that "enhances" car photos is one prompt away from erasing a scratch someone paid to know about. + +CarStudio's answer is to make the violations unrepresentable: + +- **The enhancement stage is structurally incapable of generative edits.** It's signal processing only — denoise, upscale, white balance, one consistent tone curve — plus background *treatment* limited to a vignette. There is no inpainting path in the code. Defects survive enhancement unaltered, and the operator can flag up to three defect close-ups per pack, because French portal guidance treats showing defects as a trust signal. +- **Listing text is template-filled from the decoded spec sheet** (plate → SIV registry lookup) plus the dealer's own condition notes. The optional LLM polish pass is checked by a grounding diff — any fact that can't be traced to a source falls back to the deterministic draft. The model cannot invent an owner count or a service history because the text path won't accept one. +- **GDPR as pipeline stages:** plate and face blur on every published frame, raw videos auto-purged after 14 days, SIV lookups scoped and logged. + +### Concierge first, and an audit trail for everything + +Every pack passes through an operator console before delivery — swap a chosen frame for an alternate, confirm the odometer, mark defects, approve, package. Concierge operation is not a temporary embarrassment; it's how the quality bar gets defined before automation chases it. + +
The operator console: per-slot candidates with scores, low-confidence slots flagged, listing data confirmed by a human before anything ships.
+ +Because the pipeline is deterministic, every run leaves a complete trace — and the trace page turned out to be the artifact I show people first. Per stage: what came in, what was decided, what went out. + +
One 23-second AV1 video → 92 working frames → 6 filled slots → 2 spec flags → 6 crops, 5/5 plates blurred → a French listing → 19 packaged files.
+ +
Every crop decision on the record: detected vehicle box, spec crop vs best-effort, and the delivered photo beside its source frame.
+ +### What building it taught me + +**Constraints make better products than capabilities.** The most load-bearing code in CarStudio is code that *removes* ability — the enhancer that can't inpaint, the text generator that can't free-generate. In a market where the customer's customer is deciding whether to trust a stranger's car photos, "our pipeline cannot lie about the vehicle" is the feature. + +**Local-first VLM inference changes the economics of judgment.** When a model call costs nothing at the margin, you stop rationing classification and start using it anywhere a heuristic would be brittle — while keeping the architecture honest enough that the model's absence degrades the product instead of breaking it. + +**The demo asset is the spec.** The acceptance test I set myself was that a pack built from a *bad* input — short, shaky, single pass — still had to be publishable. The live demo pack is generated from 23 seconds of video of a Mercedes G580 doing a tank turn in a parking lot. If that yields a publishable six-photo exterior set with blurred plates and a grounded French listing, a dealer's honest 90-second walkaround is easy mode.