You've seen an AI-generated image you'd love to recreate, and you want to know how the prompt behind it was built. With 86% of creators now using creative AI in their daily workflows, prompt-craft has moved from a curiosity to a core creative skill – worth investing time in.
What is an image-to-prompt workflow?
What does a good image prompt include?
How do you work out what prompt created an image?
Can you skip the prompt and just use the image?
When image-to-prompt reverse engineering falls short.
Get inspired by the Adobe Firefly Gallery.
The final step in your image-to-prompt workflow: testing with Firefly.
An image-to-prompt workflow works backwards from an AI-generated image to identify the elements that likely created it – subject, style, composition, lighting and model-specific tags. You can use those elements to write a similar prompt or skip the extraction entirely by using the image directly as input with Adobe Firefly's Generative Match.
A good AI image prompt describes six layered components: the subject, the context or setting, the visual style, the composition, the lighting and mood, and any model-specific tags – roughly in that order of importance. Subject and style move the needle most; lighting and model tags fine-tune the result.
The order isn't arbitrary. Most image models weigh the front of your prompt more heavily, so put what matters most first. The longer your prompt, the more the later words act as polish rather than direction. We found that the average Firefly prompt length doubled in 2025 – a signal that creators have moved from one-line curiosity to layered creative direction.
What the image is of. The clearer the noun, the more reliable the generation. "A cat" gives you something generic; "a sleeping ginger kitten" gives the model something to hold onto.
Where the subject is, what surrounds it, what's happening. The setting carries mood, depth and supporting detail. "On a windowsill" is a setting; "on a sunlit kitchen windowsill with linen curtains catching afternoon light" is direction.
Photoreal, illustration, watercolour, anime, 3D render or line sketch. The visual style is the single biggest lever for output character and it's the element most often left implicit. Name it. "Photorealistic" reads very differently from "soft watercolour", even when the subject is identical.
How the image is framed. Wide shot, medium shot, close-up or extreme close-up. Rule of thirds, centred subject, low angle, high angle. Shallow depth of field or deep focus. Composition tells the model what the camera is doing, which dictates what you see.
Golden hour, soft natural light, dramatic sidelight, low-key, high-key or studio backlight. Lighting carries half the emotional weight of an image. Creators consistently see better results when lighting is named explicitly rather than left to the model's defaults.
The final layer. Different AI models respond to different cues. Firefly Image 5 reads natural-language description well. Ideogram 3.0 needs explicit text-in-image instructions when you're rendering signage or typography. FLUX rewards detailed photoreal description. Nano Banana 2 handles concise visual cues better than long paragraphs. Knowing the model you're prompting, and how generative AI works, is part of the craft.
You can't recover the exact original prompt another creator used, but you can reverse-engineer something close. The workflow has four steps: observe with the six prompt components in mind, identify the dominant style cues, note the elements that need precision, then test and refine inside Adobe Firefly.
Here's how that plays out step by step.
Don't open a tool yet. Just observe. Run through the six layers – subject, setting, style, composition, lighting and mood – and note down what you see in each. This is the single most important step, and the one AI image-to-prompt tools often skip. If you can describe what you're looking at in those six layers, you're already 60% of the way to a working prompt.
Is it photorealistic? Painterly? A specific era or aesthetic – 1970s film stock, Bauhaus illustration, anime cel? Style is the prompt element that most defines the output. Get this layer right and the rest tends to follow. Get it wrong and no amount of detail elsewhere will pull the result back to where you wanted it.
Text, faces, hands and specific objects are the elements that tend to fail or look wrong in AI generation, so they're often the elements that determine the AI model you should use with your prompt. If text is the primary focus, Ideogram is the likely model. Images with photoreal faces with natural skin are often rendered with Firefly Image 5. And scenes with multiple detailed elements are likely to have been generated with Nano Banana.
Take your reverse-engineered prompt to Adobe Firefly's AI Image Generator, generate a few variations, and compare against the original. Iterate. First pass is usually around 75% of the way there. Getting from 75 to 95% is where the skill shows, and it's almost always a matter of tightening one or two layers rather than rewriting the whole prompt.
When you've got a reference image and you want to recreate the look in a new image, Adobe Firefly's Generative Match lets you skip the prompt-extraction step entirely. Upload your reference image, pair it with a short text prompt to direct the model and click Generate. Firefly will capture the style or composition directly from the reference image and apply it to your new image.
For a lot of real-world use cases – recreating a style across a series, matching a brand look, building variations on a mood reference – there's no need to convert an image to a text prompt at all. The image is the prompt. Here's how to get the most out of it.
Most models only allow for one reference image but if you'd like to provide more for a sharper output, try GPT Image 2 which allows for up to six. More references usually mean better style capture, but only to a point. Two strong references that share a clear visual language will out-perform four loosely related ones. If you're not sure a reference adds something, leave it out.
Don't just upload the image, tell Firefly in your prompt which visual element you specifically care about. Is it the sunset? The way the tide smooths the sand on the beach? The shadows? The colour palette? Vague references produce vague outputs. Naming the specific element the reference is for gives the model something to lock onto.
For more control over your image, try the Firefly Image 4 model and upload two references images – one for composition (Structure reference) and one for style (Generative Match). Use the Strength slider to define the level of creative licence you'll allow the model for each reference. Higher Strength locks the style – the output will read as a near-sibling of the reference. Lower Strength gives the model room to interpret, which is where you find variations and creative happy accidents. Start in the middle and adjust based on what you see.
The reference handles the style while your text prompt handles the subject. "This style, applied to a coffee shop interior at golden hour" is a much stronger instruction than uploading a reference and hoping Firefly guesses what you want to make.
Over-specifying for each prompt component can confuse the model on what to prioritise. If you give it too much, it doesn't know what matters most. Expect to land around 75% of the way there on your first pass. The remaining 25% is iteration: try a different reference image, lower the Strength, or tighten the text prompt.
Reverse-engineering isn't a copy-paste machine. There are categories of images where extraction works well and categories where it doesn't. Knowing the difference saves you the time of fighting a workflow that was never going to land.
Reverse-engineering is a powerful starting point, not a replacement for creative judgement. The skill is in knowing where to refine – and when to stop.
A great way to learn how to prompt well, is to see how others do it and review their results. The Firefly Gallery gives you exactly that opportunity with a huge range of images and videos shared by creators.
Once you've found an image or video that aligns with the style you're targeting, click on it to see the prompt, model and other settings, like content type and visual intensity, set by the creator. Review a few different images and note elements you see consistently through the prompts so you can incorporate them into your own.
Once you've identified the components of the image you'd like to recreate and pulled together your prompt, the only thing left to do is test it inside Adobe Firefly. Once you've generated an image you're happy with, take it to the Edit tab to add the finishing touches with Markup.
Different images call for different models.
If you're creating photorealistic portraits, lifestyle scenes or product imagery, Adobe Firefly Image 5 is a strong all-round option. It performs well with human subjects, composited scenes and commercially oriented visual styles.
For highly detailed photorealism or stylised photographic imagery, FLUX models are especially strong at rendering texture, lighting and cinematic atmosphere. They work particularly well for close-up detail, painterly realism and images with a distinct visual mood.
For images that require accurate text rendering – including posters, signage, packaging and lettering – Ideogram remains one of the strongest options available.
For complex multi-element scenes where prompt accuracy matters, Google's Gemini-powered image tools are particularly useful for following detailed instructions and maintaining consistency across multiple described elements.
