ERNIE Image workflows benefit from a brief you can inspect before generation. A product scene may look polished while changing the shape, material or number of objects that matter. Use a small controlled scene to evaluate prompt expansion and keep the accepted product requirements separate from optional styling.
Start from the official ERNIE Image project
The official Baidu repository describes ERNIE Image as a text-to-image model paired with a Prompt Enhancer. Its model cards provide distinct presets: the Base example uses 50 inference steps and guidance 4, while Turbo lists 8 steps and guidance 1. These are documented settings, not a speed comparison measured here.
Choose the actual checkpoint and implementation before copying a preset. A hosted interface may expose different controls from a local pipeline. Check current instructions and requirements for that route. This guide does not install weights, run inference or establish a hardware minimum for your particular environment.
Define a product scene that has clear acceptance criteria
Our original fictional subject is a cobalt ceramic desk lamp with a rounded shade, curved stem and circular base. Its scene is a cream desk under warm window light. Those few requirements make it possible to distinguish an intentional styling change from an unwanted product change.
Download the original short and expanded prompt plan and blank comparison worksheet. The expansion in the JSON was written manually for this guide. It is not an ERNIE Prompt Enhancer response, and none of the planned conditions has been generated. Output URLs remain empty and runtime remains unmeasured.
Compare a short prompt with a deliberate expansion
Short product prompt
A cobalt ceramic desk lamp on a cream desk in warm daylight. No text or logos.
Manually expanded product prompt
Original fictional cobalt ceramic desk lamp with one rounded shade, one curved stem and one circular base on a cream desk. Warm window daylight from the left with a believable contact shadow. Three-quarter product photograph, restrained cream background, natural glazed ceramic texture. Keep the lamp alone with no extra controls, lamps, text or logos.
The expansion makes product parts, light direction and framing explicit. It also narrows the scene. That can help a review identify why a result differs, but it is not evidence of better output. If your real product has a different stem or base, replace these invented details with its actual approved features before using the brief.
Plan three conditions without changing everything at once
| Condition | Prompt | Question |
|---|---|---|
| Short brief without enhancer | The original short prompt | What does the baseline interpretation invent |
| Short brief with enhancer | The same short prompt | Does the expanded interpretation preserve required details |
| Manual expansion without enhancer | The authored detailed prompt | Does explicit production direction solve the intended problem |
Keep the checkpoint, output size and other controls fixed where supported. Record each actual setting. A seed may help make a run traceable, but a matching seed across different checkpoints does not make their internal generation equivalent. If the interface hides an expanded prompt, record that limitation rather than reconstructing it from the picture.
Do not automatically combine a long manual expansion with another enhancement step. It may add a new style or scene detail that was not approved. If you test that fourth condition, label it separately so its effect is not confused with the three planned comparisons.
Review added detail as a production decision
Inspect the number of lamps, shade silhouette, stem curve, base shape, material and any added switches or controls. Then review the scene: light direction, contact shadow, background objects and usable negative space. Extra visual richness is useful only when it serves the product brief.
In the worksheet, leave no assumption disguised as an observation. Enter the actual output file, visible differences and whether it meets your delivery requirements. The supplied rows are marked not run. They do not contain predicted quality scores, invented inference times or a model ranking.
Keep product accuracy separate from artistic preference
A cobalt glaze can be visually appealing while failing to represent a real finish. A lamp can have a plausible curve while acquiring an extra joint. Review product identity first, then composition and style. For a real campaign, use approved source material and a supported reference workflow where available; this text-to-image exercise does not prove exact-product preservation.
The cover is an independently generated editorial concept, not an ERNIE output or a real product photograph. It illustrates the theme of a clearer brief. It is not one of the three comparison results, and its appearance supplies no evidence about Prompt Enhancer performance.
For separate form and appearance references, use the Lucid Origin content and style guidance plan.
Choose the next workflow from the deliverable
For a poster whose exact words are essential, use the GLM Image copy brief and local request preflight. For an existing product photograph that needs a controlled change, use the reference-edit planning workflow. Those guides solve different input jobs rather than establishing a winner from model names.
Explore Clout Image and Video for available creative operations and inspect the selected model, inputs and current quote. This is an alternative route, not a claim that Clout hosts ERNIE Image. Keep attempts, accepted files and finishing time together before expanding the test into more scenes.
