Separate three sources of a recurring background
A merchant wants an Ideogram custom model to learn the exact green kettle, not the marble wall behind it. Suppose 18 of 24 approved photos share that wall. The first problem is visual repetition in the training images; a caption edit alone cannot remove those pixels. A second problem is caption wording that describes the wall as if it were part of the kettle. A third, only observable after a real training run, is whether the generated candidates actually reproduce unwanted scenery. Diagnose these stages separately.
| Stage | Evidence to inspect | Failure | Decision |
|---|---|---|---|
| Image distribution | Contact sheet and count of repeated settings | One wall appears in most images | Replace near-duplicates with authorized varied photos before training |
| Caption accuracy | Each caption against its actual file | Wrong product name, angle or background | Correct the text and filename mapping |
| Post-training observation | Outputs against approved sources | Unwanted wall recurs or kettle geometry changes | Document actual behavior; reconsider source set, captioning or production route |
Keep product identity and scene context in different columns
The approved kettle's handle, spout, lid, color and mark are invariants. Camera angle, tabletop, light and distance should vary where genuine source photos allow. Review the complete set for repeated handwritten cards, watermarks or reflections before captioning. Do not synthesize new physical views to create false evidence of a real SKU.
For an illustrative training ledger, use these records:
| Filename | Product ID | Background | View | Caption audit |
|---|---|---|---|---|
| kettle-01.jpg | KETTLE-G | Wood counter | Three-quarter | Approved example |
| kettle-02.jpg | KETTLE-G | Dark table | Side | Approved example |
| kettle-03.jpg | KETTLE-G | Marble wall | Front | Flag if overrepresented |
The ledger contains reviewer-only fields. Ideogram's actual bulk-caption CSV requires exactly the columns filename and caption, with filenames matching uploaded images. A proposed CSV extract is:
filename,caption
kettle-01.jpg,"KETTLE-G green kettle on a wood counter, three-quarter view"
kettle-02.jpg,"KETTLE-G green kettle on a dark table, side view"
kettle-03.jpg,"KETTLE-G green kettle before a marble wall, front view"Quotation marks keep each caption in one CSV field despite commas. These are fictional records, not uploaded data. Check exact spelling and file extensions, and resolve any unmatched-caption warning before training.
Plan an observation without claiming a result
After an authorized training run, propose two prompts using the same product name: a side view in a plain studio and a three-quarter view in a different kitchen. Compare any resulting wall repetition, cap shape and label against real source photographs. If the wall persists, examine the source distribution before changing adjectives. If the kettle changes, record a separate identity defect. Neither failure has been observed in this exercise.
Decide whether a custom model is necessary
Training may make sense for a series of exploratory scenes; a single exact catalog hero may be better served by retaining verified product pixels. See generation-time reference roles for a different mechanism, and protected product-layer editing when factual details cannot change. Hold training when source permissions, visual variety or filename-caption integrity remain unresolved. Do not promise that better captions guarantee product fidelity.