Training
Training teaches a model one thing from the brand’s own photos: a product or a visual style, with characters coming later. Train it once, and Cauldron keeps it the same in every image you make for that brand. Each trained product or style names the model it works with, for example Works with Z-Image Turbo, and it needs training turned on for your studio. If you don’t see Train on the brand’s Training tab, ask your studio owner.
What you can train
Section titled “What you can train”| Kind | Use it for | What to pick |
|---|---|---|
| Product | One thing the brand sells, like a sofa or a coffee bag | Photos of that one product from different angles and settings. About half plain, a third in a hand, the rest in a scene is a good mix |
| Style | The brand’s visual style: light, palette, grade, composition | Finished brand images that share the style. Different subjects in the same style teach the style, not the subjects |
| Character | A person who should look the same every time | Coming later. The Training tab says so; the pick screen lists it so you know it’s planned |
A trained product is still the product: its page in Products shows a Trained state, and Start from this product uses the trained version.
The Training tab
Section titled “The Training tab”Open a brand and select Training. Everyone on the brand except guests sees it.
- Waiting on you at the top lists only what needs you: approve a run, press Train, or look over a set Cauldron suggested.
- Every row shows the name and kind (Product or Style), its status in words and whose move it is, its training set and whether it is In sync or Out of sync, what it cost or will cost in credits (About before training, Used after, refunded if it didn’t finish), and older versions folded under the newest.
- A run that is training shows its progress and time left in the row. It’s safe to leave.
- Use by default on a finished row adds it to every new image made for the brand.
- Imported files fold under Imported. See Imported files.
Studio owners and admins also get a studio-wide Training list in the sidebar’s More menu, across every brand.
Who does what
Section titled “Who does what”| Who | Edit a run (photos, captions, trigger word) | Press Train | Approve the result |
|---|---|---|---|
| Studio owner or admin | yes | yes | yes |
| The person who started the run, if a brand creator | yes | yes | no |
| Brand manager | yes | no | no |
A brand manager who curates a set hands it off with Tell Jo it’s ready (Jo being whoever can train): they get a notification in the app and by email. Nothing trains, and no credits are spent, until someone presses Train.
Start a run
Section titled “Start a run”You can start from three places:
- Brand › Training: select Train.
- Library: open a brand collection’s ⋯ menu and select Start training…. Its photos start picked, and that collection becomes the training set.
- Products: open a product’s Training tab and select Train. Its confirmed photos start picked, approved ones first, with near-duplicates left out. Fill in the product’s What it is first. The product’s own page shows where its training stands, whose move it is and the one next step; see Products.
Cauldron can also suggest a set on its own. It shows as Suggested set on the Training tab with why it picked those photos, and nothing trains until a person looks it over.
The run, step by step
Section titled “The run, step by step”A run is one screen per step. Six steps run across the top: Photos, Captions, Training, Test grid, Approval, Ready. Each finished step says what it produced (“18 picked by Jo”, “Took 1 h 18 min”), and the bar along the bottom says what happens next, who does it, and holds the one action that is yours. When the move isn’t yours, the bar names who it’s waiting on.
1. Photos
Section titled “1. Photos”Pick a kind, a name and 15 to 60 photos. 25 to 40 is a good set; more angles and settings teach it more. Approved photos show by default; turn on Drafts to see the rest. Photos that can’t be used say why, and a photo that looks like a copy of another teaches it nothing new, so keep one of each pair. Cauldron suggests a trigger word, the short made-up word that goes into every caption and that Create adds to your prompt when you use the result; change it if you like. Training length sits under Advanced: Standard, Longer or Longest.
2. Captions
Section titled “2. Captions”Cauldron writes a caption for each photo; you can leave the page while it does. Captions say what changes from photo to photo (the pose, the setting, the light), and training learns what stays the same. Open Check the captions to edit one, leave a photo out, put one back, or Add photos. Then:
- If you can train, the bar reads Next: you press Train with the price and the time, for example about 1 h 20 min. Credits are spent when you press Train.
- If you are a brand manager, the bar reads who presses Train, and Tell Jo it’s ready sends them the run.
3. Training
Section titled “3. Training”Cauldron trains. The step shows how far it is and the time left, and it’s safe to leave: the run shows in Jobs, and you get a notification when the test grid is in. Whoever pressed Train can stop it. Stopping before training starts refunds the credits; once it has started, they aren’t refunded.
4. Test grid
Section titled “4. Test grid”When training ends, Cauldron makes a grid of test images and writes its read of them, for example that the handle is missing in one of four.
5. Approval
Section titled “5. Approval”A run you train as a studio owner or admin is ready as soon as it finishes. Anyone else’s waits for an owner or admin. On the approval screen the test grid comes first, with Cauldron’s read of it and your own photos below, and the paths sit beside it:
- Approve. It becomes available to everyone on the brand in Create.
- Retrain with this fix, priced like a run: trains the next version with the fix Cauldron suggested from the grid.
- Try it lighter, free: trains it again with a lighter touch.
- Decline, with a note. Whoever started the run sees the note, and they hear when you approve too.
The approval screen works on a phone, so an owner can decide away from a desk.
6. Ready
Section titled “6. Ready”Use in Create opens Create with it already added. It shows in Create’s Trained tab for everyone on the brand, and Use by default on the Training tab puts it in every new image for the brand.
If it didn’t finish
Section titled “If it didn’t finish”A run that stopped says so plainly: your credits are back, and the set and captions are kept, so Try again or Retrain picks up where it left off. The same goes for a run you stopped.
Training sets
Section titled “Training sets”The photos a run trains on are a collection in your Library, named after the product or style with “training set” on the end (Atlas training set), in the brand’s Training sets folder. Start from a collection and keep every photo, and that collection is the training set.
- Add or remove photos in the collection or on the run’s Captions step, and the set follows. New photos get captions written for them.
- The Training tab and finished runs show whether the set is In sync, or how many photos were added and removed since that version.
- Retrain opens the next version straight from the set as it is now, keeping your captions and the photos you left out. A run’s versions share one set, and set changes made while a version trains wait for the next one.
- The collection says it is a training set under its title, with the run’s status, whose move it is, and your next step. Its ⋯ menu has Review training set while the run is open, or Retrain once it has trained.
- Share on a training set lists the brand’s people and what each can do: train (spend credits), curate (edit the set and captions), or add photos. Copy the link for a teammate, or turn on a view-only link that shows the photos and nothing else. Sharing never changes anyone’s role.
- When the set is out of sync, Changed since vN shows the photos added since that version, with who added them and when, and the ones removed. Put back returns a removed photo to the set.
Versions
Section titled “Versions”Open the training set in your Library and every version that trained from it is listed under the collection’s header, newest first. Each version shows:
- its number and status, with Ready · in use on the version Create uses now
- how many photos it trained on, the training length, and how long the run took
- what changed since the version before: how many photos were added and removed, or same photos
- Cauldron’s read of its test grid, and who approved it and when
- the test grid itself once the run finished, and Open to go to the run
Compare on an older version puts its test grid beside the one in use, the same test scenes from each, so you can see what a retrain changed.
While the next version is in review and an older one is still in use, a line above the list says so, for example “v3 is in review”. Photos added to the set now join that review with a fresh caption, and removed photos leave it. Open the review takes you there.
See Collections and tags for working with collections.
Mature content
Section titled “Mature content”If some of a run’s photos are mature, the result is marked mature and follows the brand’s Mature content setting.