AI asset versioning
AI asset versioning is how you name and keep takes so you can re-run them. It is a folder habit, not a product.
AI asset versioning is how you name and keep takes so you can re-run them. It is a folder habit: still, prompt, model slug, seed if you have one, take number. It is not a Versely product, not a hidden history pane you should wait for, and not Git for MP4s unless you already like Git.
Search Console still shows "ai asset versioning" (13 impressions, seven days to 4 September 2026). The query is operational. Teams lose the winner because the file is final_final_v7.mp4 and the prompt lives in a chat they closed. Then they pay again to rediscover a motion they already liked.
What you actually have to keep
A generate is not one file. It is a bundle. If any piece is missing, the re-run is a new job.
The still. For image-to-video, the still is the identity. Lose it and the pack cannot be remade. Keep the approved PNG or TIFF next to the clips, not only the compressed upload. If the still came from Z Image Turbo or another text-to-image row, keep that still as an asset too — 1 credit to remake a 4K draft is cheap; remaking it from memory is not the same picture.
The prompt, as text. Copy it out of the box. One motion per I2V take. The T2V prompt that invented a city. The Qwen instruction that removed a cup. A screenshot of the UI is a backup, not a source of truth; you cannot paste a JPEG into the next run.
The model slug. kling-video-v3-pro-motion-control is not "Kling." qwen-image-edit-2511 is not "Qwen edit." Rows change. Fast versus Pro versus a talking generate versus a mouth pass — the slug is the only name that maps to a catalog card. Keep it in the filename or in a sidecar .txt.
The knobs that changed the bill. Duration chip, aspect, resolution. Audio on or off. Credits on Versely are per the card you ran, not a USD line. If you do not record 8s versus 4s, you will re-run the expensive one by accident.
The seed, if the UI showed one. Same prompt plus same still plus different seed is a different take. If you never saw a seed, do not invent one in the notes. "We don't have a seed on this row" is a valid line.
The take number and the verdict. v03 keep, v04 no — invented fingers, v05 captioned. A folder of ten unlabelled MP4s is not versioning. It is a mood board you will re-watch in full every time a buyer asks for "the orbit one."
The audio separately when it is a separate row. Inworld TTS 2 is a 5-credit wav, not a talking generate. Keep the wav, the script, the style note. If a lipsync pass wore that wav, keep the plate too. Three files, one line.
You do not need a DAM on day one. You need a directory per campaign and a name you can sort.
A naming scheme that survives a re-run
Pick a boring pattern and do not get clever.
YYYYMMDD_campaign_sku_model_aspect_duration_take_verdict
Example: 20260916_serum_hero_i2v-kling-o3_9x16_5s_t03_keep.mp4
Next to it:
20260916_serum_hero_still.png20260916_serum_hero_t03_prompt.txt(motion only, still not re-described)20260916_serum_hero_t03_notes.txt(slug, aspect, duration, seed or "none", credits charged)
For edits: still_v02_qwen2511_remove-cup.png plus the instruction in still_v02_instruction.txt. Then I2V takes hang off still_v02, not off still_v01. Mixing takes from two stills in one folder is how a buyer picks a clip whose label no longer exists.
For motion control: keep the driving clip. The still plus the driver are both inputs. Kling V3 Pro motion control will not remake a dance you deleted from Downloads.
For captions: keep the transcript before burn-in. Burned-in captions are pixels. A spelling fix is a new render. The .txt or .srt is the version you can edit; the MP4 is the version you shipped. If you only keep the burned file, you will recaption from audio and miss the product spelling you already corrected.
For talking takes: plate, wav, lipsync output, as three siblings. Do not overwrite the plate with the talking file. You will need the closed mouth again when the line changes.
Date first so a week sorts. Campaign next so two SKUs do not collide. Take numbers as t01, t02 — not final. Verdict as a suffix so a search for _keep is the shortlist.
This is not a Versely feature request. It is the same discipline as naming stills on a packshot day. Generative work makes it more necessary because the "negative" is cheap and plentiful. Ten discards without names cost more in review than in credits.
What versioning is not
It is not a promise that the vendor will reproduce a take forever. Models move. A row can change weights behind the same public name. A discounted credit figure can end. OpenAI's discontinuation note is explicit about Sora: the API shut on 24 September 2026. If your notes say sora-2 and nothing else, the re-run path is a migration, not a click. Keep the prompt anyway. Rewrite it onto Veo 3.1 or Seedance 2.0 with the still attached if identity matters.
It is not seed lock as a religion. Some rows do not expose a seed. Some hosts do not honour one across a week. Versioning still pays: you have the still, the instruction, and the motion that won, which is enough to get close on a new row.
It is not "save every preview." The editor's 480p preview pass is free with a short cooldown; it is for looking, not for archiving. Keep the keepers and the inputs. Throw away the previews.
It is not Article 50 labelling. If you publish realistic synthetic media of real people to EU audiences, the Article 50 mark is a delivery obligation from 2 August 2026. Put the label on the shipped file. Versioning is how you find which take shipped, so you can prove what you labelled.
It is not a substitute for a locked still. If the hero photo was never approved, versioning ten I2V takes is versioning noise. Freeze the plate. Then version the motions.
A weekly loop that does not need new software
Monday: one folder, one still, three motion prompts in a text file.
Generate. Name as you download. Mark _keep / _no the same day, on mute, on a phone. Do not leave a pile for Friday.
When a keep wins in ads, copy its prompt.txt into winner.txt and do not "improve" it in the original. Branch winner_hookB.txt if you change a line.
When the line changes, re-run from the still and the motion file. Do not regenerate the still. Do not open T2V.
When the model row is wrong for a remake — sunset, quality miss, duration — change the slug in the notes and keep the still. That is failover. The AI video generator and image-to-video tools are doors, not archives. Your disk is the archive.
If you already have a DAM, put the same fields in the metadata. If you do not, the pattern above is enough to re-run a Reel in six months without a séance.
Keep the still. Keep the text. Keep the slug. Number the takes. That is AI asset versioning.
Start the next pack with a named still and a named first take in the studio rather than another export(3).mp4.
FAQ
Does Versely version my generates automatically?
Treat the catalog as a generate surface, not as your DAM. Download the keepers and the inputs (still, prompt, slug, knobs). Name them on disk. Do not wait for a history feature to become the archive.
What if the model no longer exists when I re-run?
Keep the still and the prompt anyway. Point the same brief at a live row — Veo 3.1, Seedance, Kling — and attach the still if identity matters. Sora 2's API, which shut on 24 September 2026, is the current example of a slug that will not re-run.
Should I save failed takes?
Save one example of a systematic fail (invented fingers, wrong label) plus the prompt that caused it, so you do not retry it. Delete the rest. Versioning is so you can re-run winners, not so you can hoard noise.