Text Prompt or Reference Image? Choose Your AI Image Starting Point

Start with what the image must get right

The difference is what you are asking the model to infer. “A green glass bottle on a gray surface” describes a category. A photograph supplies a particular bottle's neck, shoulder, labels and proportions. The text can explain which of those details matter and what should change; it cannot replace visual information you have never supplied.

Dreamina supports both text-to-image creation and reference-led image generation. You can begin with an idea, introduce an image when the direction becomes concrete, and refine the selected result. Choose the starting point around your deliverable, rather than treating one input method as universally better.

Use this decision before you write a long prompt

Ask: What would make a visually attractive result the wrong image? Your answer tells you where the important information lives.

If almost any well-composed woodland reading nook would satisfy the brief, words may supply enough direction. If it has to contain your own distinctive chair, a generic chair will fail even in a beautiful scene. If an approved label must remain unchanged, a recognizable redraw of the label can still be unacceptable.

These are workflow choices. A reference can make the intended subject clearer without guaranteeing an exact copy. A text-only generation can establish a useful composition without becoming a reusable identity reference for a real person or product.

When a text prompt gives you more useful freedom

The subject does not exist yet

For a fictional environment, you can specify the things that define it and let the generator propose the remaining visual detail. Suppose the brief is an inviting woodland reading nook for an illustrated magazine feature. No particular room, chair or building must be reproduced.

A text-first brief can say where the viewer looks, how the space is lit and what emotional character the image should have. You are evaluating a proposal for the scene. Introducing a random interior photograph at this stage may anchor the result to a room arrangement you never intended to keep.

Here is a planning example, rather than a tested prompt:

Create an editorial illustration of a small reading nook inside a hollow tree. Place an empty chair slightly right of center, with a curved bookshelf following the wall behind it. Warm afternoon light enters from a round window on the left. Use broad painted shapes, muted moss green and ochre, and leave a quiet area in the upper left for a headline. No lettering in the artwork.

The useful constraints are visible: chair placement, shelf shape, light direction, palette and space for type. They leave room for invention in the wood grain, book arrangement and small environmental details. You can judge whether the result serves the page without needing it to match a source photograph.

You need to compare visual ideas before choosing one

Text is also useful when you want genuine alternatives. Describe a shared subject and vary the visual treatment deliberately. A painted woodland scene, a geometric paper-cut version and a monochrome ink drawing present different ways to solve the brief.

Do this before you select an image as the reference for everything else. Once an attractive output becomes the anchor, later generations may remain close to its composition and motifs. That can be useful during production, but less useful when you are still trying to discover a direction.

For a fuller method of building and selecting alternatives, use the art-direction exploration guide. Here, the decision is simpler: if a new visual proposal is welcome, text-first leaves that proposal open.

Your only available reference contains the wrong assumptions

A reference is not automatically useful because it looks attractive. A wide landscape can communicate atmosphere while suggesting a composition that will not fit a narrow portrait illustration. A highly polished product scene can bring props, reflections and visual hierarchy that do not belong in your project.

If you want only its warm lighting, you may be better off describing that property: low afternoon light from the left, soft shadows and a warm neutral background. Use the image only when it communicates something important that words alone leave ambiguous. A reference should reduce uncertainty about the brief, not quietly replace the brief.

When a reference image becomes necessary

The result must depict a particular subject

A reference gives the model a visible starting point for a specific person, object or character. It is useful when recognition depends on details that would be difficult to enumerate: the shape of a bottle shoulder, the placement of a seam, a particular haircut or the relationship between several objects.

Name the intended subject if the image contains distractions. “Use the complete central bottle” is more precise than “use this photo” when another bottle appears along the edge. Then list the few features that would make the result unmistakably wrong if they changed.

This does not mean every visible item must survive. If you want the bottle in a studio setting, the restaurant table is context to replace. A reference-led prompt should explain that division so the model does not have to decide whether the background is part of the subject's identity.

You have already approved the arrangement

A sketch or rough layout can express positions more directly than a paragraph. If the left third must remain clear for a heading, or the main figure has to sit below a curved arch, a simple layout image makes those relationships visible.

Identify it as a composition guide. Otherwise, a generator may treat the sketch's rough line style as part of the desired aesthetic. Add the intended finish, materials and lighting in words, and compare the result with the structural relationships in the drawing.

ByteDance's Seedream 5.0 Pro introduction documents sketch and reference-based design examples. Its own conclusion also acknowledges room for improvement in fine text and pixel-level editing consistency. Use a sketch to direct a generative interpretation; use a design or compositing tool when exact alignment is a delivery requirement.

The style is easier to show than to describe

An authorized reference can help communicate a combination of texture, shape language, palette and mark-making. Explain which properties you want to carry over. “Use the limited palette and broad ink shapes” is more useful than an unexplained “make it like this.”

Keep the content requirement separate. If the reference is a crowded street illustration but your brief needs a single tree, you want its treatment, rather than its people, storefronts and traffic. This is especially important when your subject, arrangement and style come from different images. The reference-pack guide explains how to assign those roles without giving every image equal authority.

What a real reference-led example establishes

The following example uses the saved prompt, settings and original output from the October 14, 2026 browser demonstration.

The starting image showed a complete green bottle on a restaurant table, with a partial bottle and tableware nearby. The task was to use that specific central bottle in a gray studio-style setting. The demonstration used Seedream 5.0 Pro in Dreamina, a Basic account, a 3:4 aspect ratio and the Standard 1.5K setting, producing one downloaded 1296 × 1728 JPEG.

Original reference: Philafrenzy, “San Pellegrino 500ml bottle,” Wikimedia Commons, CC BY-SA 4.0. This is a demonstration subject, not a Dreamina customer case or product endorsement.

The exact submitted prompt was:

Use only the complete central green glass bottle in the reference. Place it on a neutral light-gray studio surface. Keep the open uncapped neck, bottle proportions, three labels, visible brand lettering, red star and 500 ml marking. Exclude the partial bottle and all tableware. Soft even lighting, photorealistic, no extra text.

AI-modified adaptation of the reference photograph, shared under CC BY-SA 4.0, with attribution to Philafrenzy. The setting and bottle rendering were changed by generation; product branding identifies the example subject and implies no endorsement.

The generated image removes the restaurant setting and partial bottle while retaining recognizable features of the central subject: green glass, an uncapped neck, three label areas, a red star and a readable 500 ml marking. These are useful anchors for a creative reinterpretation.

Fine lettering, label curvature, reflections and perspective were redrawn. The image therefore illustrates a reference-guided transformation, rather than preservation of the original bottle pixels. There was no matching text-only control in this demonstration, so it does not measure how much the reference improved the result.

For the starting-point decision, the distinction is still clear. The photo supplied information about this particular object; the prompt selected its important features and changed its context. If the intended deliverable had required an exact photographed label, a preserved source region or conventional composite would have offered the necessary control.

The separate photo-and-illustration workflow develops the cross-style use of this same source. That is a different decision from whether an image reference is needed in the first place.

Build a prompt that complements your chosen input

A useful prompt gives the input a job. For text-only generation, it defines the scene. For reference-led generation, it defines the transformation and explains which evidence in the reference matters.

Consider two versions of the same request:

Text-first planning example:

Create a photographic-style image of an invented green glass bottle with an unbranded cream label, standing on a light-gray surface. Full bottle visible, soft light from the left, quiet background, no lettering.

Reference-led planning example:

Use the bottle in the uploaded photograph as the subject. Retain its distinctive neck and shoulder shape and the visible label arrangement. Replace the surrounding room with a light-gray studio setting. Exclude the other objects in the source.

These examples have not been run as a paired test. They show different intentions. The first permits a new bottle design; the second asks for continuity with a supplied object. Treating them as identical tasks would lead to an unfair quality comparison.

Avoid describing every irrelevant detail in a source image. If the background will change, a long inventory of its chairs and cutlery gives those objects unnecessary attention. Spend your words on the desired result, important preserved features and exclusions that matter.

Also avoid writing “keep everything” when the brief asks for a new angle, style or material. Those changes may require a different representation of the subject. Decide which properties define recognition and which can legitimately change before generating.

Move from input choice to a usable image

1. Define the deliverable and the non-negotiable details

Name the use: an editorial illustration, a concept image, a product scene or a composition for a page. Then identify the details that must be right. Separate visual preferences from hard requirements.

For example, warm light is a preference you can judge creatively. A particular chair's arm shape is an identity requirement. A ticket price is factual information. Different failures call for different remedies, so write them down separately.

2. Prepare only the input you need

For a text-first job, begin with a concise description of subject, arrangement and treatment. For an image-first job, choose a clear reference you have permission to use and keep an untouched copy. Crop a working version if several subjects compete, but do not remove a feature you expect the model to preserve.

A source that hides the back of an object cannot supply that hidden geometry. If you need an accurate rear view, find an appropriate reference rather than asking the model to invent it and accepting the invention as fact.

3. Generate in the appropriate Dreamina workflow

Open AI Image, select the intended available model, and enter your prompt. Upload the reference when the task calls for it. Choose the ratio, resolution and output quantity for the intended placement, and review the displayed credit requirement before generating.

Check the selected model explicitly: a page about Seedream does not mean every entry screen has already selected Seedream for you. The Seedream access and app-workflow guide explains that distinction and the steps supplied by the host application.

4. Approve the structure before polishing detail

Look first at the subject, framing and visual hierarchy. A beautiful texture will not fix the wrong object or a composition that leaves no room for the heading. Reject a failed structure early instead of investing several editing rounds in it.

For a reference-led image, compare the output with the original, not just with your memory of it. Familiar colors can make an altered shape feel more faithful than it is. For a text-first image, compare with the written brief and ask whether the invented details are useful, plausible and appropriate.

5. Change one decision at a time

If the overall direction works, refine the specific weak point using the tools available in your selected workflow. Keep the source, earlier approved output and exact prompt so you can return to a sound version.

A targeted change and a new generation have different consequences. If you dislike the entire art direction, a fresh branch is sensible. If you need to correct one small element in an approved composition, replacing the whole scene introduces unnecessary opportunities for drift.

6. Download and finish for the destination

Open the downloaded file and confirm its dimensions and format. Add critical wording, logos or other exact assets through an editable design workflow when needed. A visually flat background is not proof of transparency, and an illustration with clean edges is not proof of vector paths.

Keep the prompt, reference and selected export together. That small handoff makes the starting decision reusable: the next person can see which facts came from the source and which details were deliberately invented.

Switch methods when the current starting point stops helping

A text-first image keeps inventing the wrong subject. If the subject is supposed to be specific, supply a reference. Adding more adjectives may clarify style without defining the missing geometry or identity.

A reference-first image looks too close to the source. Decide whether you need the source for subject, composition or style. Remove irrelevant references, explain the permitted change, or start a new text-led direction if nothing in the source must remain.

The scene changes successfully but small details drift. Move the exact material to a preserved layer or finishing editor. The bottle example shows why recognition and exact packaging reproduction need different standards.

Every attempt inherits the same unwanted layout. Stop refining the same reference. Use a simple sketch that expresses the arrangement you actually want, or return to text-first exploration before selecting a new anchor.

You cannot decide whether the result is correct. The problem may be the brief rather than the input. Define what can change and what cannot. Without that agreement, each attractive output becomes a different interpretation of success.

Questions about choosing your starting point

Does using a reference replace the prompt?

No. The image supplies visual information; the prompt explains its purpose and the desired change. Without that explanation, the model may preserve context you wanted removed or change a feature you considered important.

Can I start with text and add a reference later?

Yes. That is a useful progression when you first need to discover a visual direction and then develop it. Save the chosen result as a reference only after checking it against the brief, because its mistakes can otherwise become the foundation for later images.

Should I upload as many references as possible?

Use the smallest set that communicates the required information. More images can introduce incompatible subjects, viewpoints or styles. Add one because it supplies something missing, and name its role; a larger pack is not automatically a clearer instruction.

Is a reference enough for an accurate product listing?

It can guide recognition, but a generated redraw can alter lettering, shape or physical detail. When the listing must document the actual item, retain accurate source imagery and inspect the finished composition against the product. A compelling concept image and a faithful listing image serve different jobs.

Choose what to invent and what to carry forward

Use text when you want a visual proposal, a reference when you need continuity with a particular source, and a preserved asset when exactness matters. In Dreamina, make that choice before generation, then judge the result against the information the input was meant to communicate.

Start an AI image in Dreamina. Bring a clear brief, and add a reference when it has a specific job to do.