Most teams that make product visuals do not need one more one-off hero image. They need a repeatable look: the same light, the same surface, the same crop discipline, applied to twenty SKUs without twenty briefs. The interesting shift in image generation over the past year is not that models got prettier. It is that the workflow around them started to accept references as first-class input, which is what turns a lucky render into a house style.

This post is about how I set that up, in practical terms, using a browser-based generator as the working surface. I write it from the operator's side, not the researcher's.

The problem with prompt-only iteration

A prompt is a lossy description of a picture. If you write "soft daylight, matte ceramic, three-quarter angle", you are describing the outcome in words and hoping the model resolves the ambiguity the way you pictured it. When it does not, you rewrite the sentence. Two hours later you have a folder of almost-right images and a prompt that is 400 words long and fragile.

The fix is to stop describing and start showing. Once a model accepts reference images alongside the prompt, the reference does the heavy lifting on appearance and the prompt only has to describe intent: what is in frame, what changed, what must stay.

Three reference roles worth separating

In practice I end up using references in three distinct roles, and keeping them in separate folders helps more than any prompt trick:

1. Subject reference — the actual object or person. This is what most people mean by image-to-image.
2. Style reference — a finished image whose light, palette and texture you want, with completely different content.
3. Structure reference — a rough sketch or a layout that fixes composition while the model invents the surface.

Mixing all three into one upload usually produces mush, because the model has to guess which attribute of which image matters. Naming the role in the prompt is the single cheapest improvement available: "keep the bottle from image 1, take lighting and surface from image 2, follow the framing of the sketch" resolves ambiguity that a hundred adjectives cannot.

A workflow that survives a Monday

The loop I actually use, in order:

• Lock the constants first. Surface, light direction, lens feel, background tone. These are the attributes you want identical across the whole set, so they belong in a reference, not in the prompt.
• Vary one axis at a time. Rotate the product, change the colourway, swap the prop. If you change two axes at once you cannot tell which one broke the look.
• Batch, then cull. Generating wide and selecting hard beats generating narrow and re-prompting. The evaluation step is faster than the description step.
• Keep the reference set small and boring. Five references that all agree beat fifteen that argue. A style library is a contract, and contracts are short.

This is the part Seedream 5.0 Pro is built around: text-to-image and image-to-image on the same surface, sketch-guided generation, multiple references in a single request, and batch output, so the culling step happens inside the same session instead of across four tools. It runs in the browser, which matters more than it sounds — the asset folder, the references and the output all live in one place while you iterate.

What still needs a human

Two places where automation quietly degrades quality:

Colour fidelity. Generated surfaces drift toward whatever is pleasing rather than whatever is accurate. For product work, compare a render against a colour chip before it goes anywhere near a page. Expo markers and screens both lie, so compare numbers, not vibes.

Text in image. Short labels sometimes survive. Anything that has to be legible, spelled correctly, or on brand should be composited afterwards. Do not treat generation as typesetting.

The realistic payoff

A style library is worth building when you have a set, not a shot. If this quarter needs forty images that read as one family, an hour spent locking references and one axis of variation pays for itself the first time a new SKU arrives and the answer is "same references, new subject, done". If you need a single illustration for a blog post, skip all of this and write a longer prompt — the structure is overhead until repetition shows up.

The habit that transfers across tools: describe less, show more, change one thing at a time, and keep the reference set honest. Models will keep changing. The workflow does not have to.