Skip to content
MadooMadooDocs

D02 — Hotel photos to social clips, screened first

A hotel sends its new photo shoot and wants short video clips for its social channels. A video clip is the expensive part — tens of credits each — and a real shoot always has a few photos that should never become one: a picture forwarded through a chat at a fraction of its size, a corridor shot with a shaking hand, a terrace at dusk without a flash. This workflow looks before it pays: free checks first, then a cheap AI judgement, and the video model only for what passed both.

Frames from the three clips the workflow made — the breakfast terrace, the lake-view room and the breakfast table — each at the start, the middle and the end of its four seconds: a slow camera push-in, curtains stirring, light moving on the lake

A reference run: 6 photos in, 3 clips out, every rejection explained — 96.2 credits. Clips: terrace · room · breakfast · the run’s outputs.

The case

Section titled “The case”

Hotel Villa Limoni (a fictional family hotel on Lake Garda) receives six photos for its summer campaign: the breakfast terrace, a lake-view room, the breakfast table, the pool, the spa and the restaurant. Three are good. The other three have the faults of every real shoot: the pool arrived at 640 pixels because someone forwarded it through a chat, the spa was shot with a shaking hand, the restaurant at dusk is far too dark.

The shoot: six hotel photos — the terrace, the room and the breakfast table sharp and bright; the pool that looks fine but is only 640 pixels wide; the spa corridor with heavy motion blur; the restaurant veranda almost black

The photos arrive as one bundle: each with its space (“breakfast terrace”, “pool”) and its rights, each checked for ownership and checksum before anything else happens. The photos are the example’s starting data, generated once (30 credits) and then degraded the way real files are. I07 screens the same photos by measuring them, with no AI step: the verdicts are the same, at no credit cost.

What it shows

Section titled “What it shows”
  • Free before cheap before paid. The workflow orders its checks by cost. Measuring a file costs nothing; an AI looking at a photo costs about 0.03 credits; a four-second clip costs 32. Each step only receives what passed the one before, so the expensive step never sees a photo it would waste money on. Without the screening, this shoot would have cost 192 credits and delivered three clips nobody could post; with it, 96, and every clip is usable.
  • Some faults are measured, not judged. The pool photo looks perfect to the eye — and to an AI — but it is 640 pixels wide, too small for a 720p clip. Media Technical Info measures it against a media policy, and a Quality Gate in route mode lets only the photos that pass go on. See Quality checks.
  • Some faults need eyes. Blur and exposure are not in the file’s metadata. Image Analysis judges each photo that passed the measurement and answers in JSON — sharp, exposure, publishable, a reason, a caption, and a direction for the camera — and Validate JSON Schema holds the answer to its contract before it decides anything. The verdicts in the reference run: “pronounced motion blur and lacks a crisp focal subject” for the spa, “significantly underexposed for professional hotel promotion” for the restaurant. See Schemas and contracts.
  • The judge also directs. The same answer carries one sentence for the video model — a slow push-in, curtains moving, light on the water, nothing added that is not in the photo — so each clip moves in a way that suits its photo, without anyone writing six prompts.
  • Every photo is accounted for. The Screening report has a row for each of the six photos: its width, whether it was big enough, the AI’s verdict and its reason. The pool’s verdict is empty on purpose: it was never shown to the AI, because the measurement had already excluded it. The Campaign keeps only the photos with a clip, with their captions (whenAbsent: skip_row). See Branches and absence.

The workflow

Section titled “The workflow”

The workflow in five stages. 1, receive the shoot: a bundle of 6 photos, one at a time. 2, free — is the file big enough: Media Technical Info measures each photo against a policy of at least 1200 by 675 pixels, a Quality Gate in route mode lets 5 pass; the pool is 640 pixels. 3, cheap — does it look good: AI vision judges each photo in one JSON verdict, about 0.03 credits a photo; JSON Schema checks the verdict; a Filter keeps the publishable ones: 3 pass, the blurred and the dark are out. 4, paid — only now the video: an image-to-video clip of 4 seconds with the verdict’s motion direction, 3 clips at 32 credits each. 5, account for every photo: the Campaign with 3 posts ready, and the Screening report with the size and verdict of all 6 photos.

The figure tells the workflow as stages, with the reference run’s results on the cards; the exact graph is the workflow (as a diagram), node by node in the table.

Node Type Why
Photo shoot input/bundle_manifest, validation: warn, verifyChecksum: always The shoot as one checked package; a file that fails its checks is reported and left out
Photos / One photo at a time utility/extract (assets), enumerate/json One iteration per photo, with its id, its space and the image
Measure media/technical_info, media policy: an image of at least 1200 × 675 The file’s real size and type, as a portable quality check
Big enough utility/quality_gate, all_must_pass, route, the photo on Value A photo too small never reaches the AI
Judge the photo ai/image_analysis, json_object, temperature 0.2 Sharp, exposure, publishable, reason, caption, motion — one verdict per photo
Verdict is complete utility/json_schema_validate, inline schema, route The verdict’s fields and limits before it decides
Publishable? / Why / Caption / Motion utility/extract The verdict’s fields
Keep publishable utility/filter on publishable The photo passes or is cut
Social clip ai/image_to_video, 4 s, 720p, 16:9, no audio The paid step, with the verdict’s motion as prompt
Campaign aggregate/json, the clip with whenAbsent: skip_row The posts ready to schedule: photo, space, caption, clip
Screening report aggregate/json Every photo, with its size and verdict

Reading a run: Screening report is where to start: six rows, one per photo. big_enough: false is a file problem (ask for the original); publishable: false with its reason is a photo problem (reshoot or skip). Bundle issues lists any file of the shoot that failed its checks before the workflow started. Campaign and the clips are what goes to the social channels.

Cost: 96.2 credits in the reference run — three clips at 32 credits (4 seconds at 8 credits a second) and five verdicts at about 0.03. The estimate before the run cannot know how many photos a bundle holds, so it reports the cost per photo and marks the count as deferred; run with admission_limits (for example a maximum of 20 iterations) when a shoot could be larger than expected. Changing the video model changes the price: faster models cost an eighth to a third of the default, at a visible loss of quality.

Install it

Section titled “Install it”

The example installs into your workspace with an API key, through the public API only — see Installing an example:

Terminal window
node madoo-install-example.mjs d02-hotel-clips --api <your Madoo API URL> --key <key.json>

It uploads the six photos, composes the bundle from shoot-spec.json (Madoo derives type, size and checksum), saves it as the default of Photo shoot, and publishes the workflow. Add --run to run it once (about 100 credits) and --save <folder> to download the clips and the reports.

Change it

Section titled “Change it”
  • Your shoot: compose a bundle from your own photos, with their spaces in the metadata, and pass it at run time.
  • Your standard: the media policy of Measure (the minimum size) and the instructions of Judge the photo (what makes a photo publishable) are the two dials.
  • A judge that only decides: when the verdict needs no caption or camera direction, Decide with a Model answers sharp, exposure and publishable with a probability each, on its own ports, with no schema to validate — see X01.
  • Another format: the clip’s aspect ratio, length and resolution are parameters of Social clip; for vertical stories, reframe the photo first.
  • Another paid step: the screening in front does not care what follows — an upscale, a relight, a catalogue layout. Put the expensive step behind Keep publishable.

Files

Section titled “Files”

manifest · workflow · shoot spec · the complete graph