You have a product to launch on Friday and a budget of nothing. A photographer, a copywriter and a voiceover artist are all out of reach. What you do have is a browser tab, a few free tiers, and the discipline to stop experimenting and start producing.
Below is one workflow run end to end on a deliberately ordinary project: a 14-ounce soy candle called Fog Signal, made by a two-person outfit called Harborline Candle Co. Six steps, about two hours of real work, and a finished set of assets at the end of it.
Decide the Deliverables Before You Open a Tool
The fastest way to waste an afternoon is to type something vague into a generator and hope. Thirty images later you have no idea which one is usable, and your daily credits are gone.
Write the shot list first. For Fog Signal it looked like this:
- One hero image for the product page, 1600×900
- Three lifestyle images for the gallery
- A 12-word tagline and a 90-word product description
- Six social captions, each a different format
- A 20-second voiceover script plus a music bed
That is nine outputs. Nine fits inside free tiers. Forty does not, and forty never gets finished. If you are still deciding which platforms deserve your time, start with this roundup of ten free AI generators actually worth using in 2025 and pick one tool per output type.
Step 1: Turn a Messy Brief Into Structured Copy
Start with words, because text models are fast and their free tiers are the most forgiving. Paste in a short brief and give the model a job title, a format, and hard constraints.
The actual prompt used: You are a brand strategist. Here is the brief: Fog Signal is a smoky cedar and sea-salt candle for people who live near cold water. Return three messaging pillars with a one-line explanation each, then a 90-word product description at a grade-8 reading level. No exclamation marks. No sentence longer than 20 words.
Those last two rules matter more than anything else in the prompt. Left alone, most models produce chirpy copy full of exclamation marks and 35-word sentences. Constraints are cheaper than editing.
Step 2: Batch Your Image Ideas in Fours
Now the visuals. Ask for four variations per generation, never one, because free tiers charge per request rather than per image. Four attempts cost the same as one.
A workable first prompt for the hero shot: A single unlit soy candle in a ribbed amber glass jar, sitting on a weathered concrete ledge, thick coastal fog behind it, early morning, soft diffused light, 35mm film look, shallow depth of field, muted grey and warm amber palette.
Then change exactly one variable per round. Swap “thick coastal fog” for “thin morning mist” and generate four more. Adjust the light and generate four more. Changing three things at once teaches you nothing about which word did the work. There is a fuller breakdown of this tightening loop in this step-by-step DALL·E workflow for better AI images.
Step 3: Lock a Reusable Prompt Template
Once one image lands, do not start from a blank box again. Copy the winning prompt and rebuild it as a template with six slots:
[subject] + [setting] + [lighting] + [lens and film stock] + [colour palette] + [mood]
Fog Signal’s locked tail was: early morning, soft diffused light, 35mm film look, shallow depth of field, muted grey and warm amber palette. Every gallery image reused that tail verbatim. The result looks like one photographer shot all four frames. Drop the tail and you get four images from four different brands, which is the single most common reason free AI imagery looks cheap on a product page.
Step 4: Write the Short Copy With a Different Model
Use a separate text tool for the tagline than the one that wrote the description. Same-model output has a rhythm that repeats, and a second model breaks it up for free.
Ask for ten taglines in one request, then pick. Fog Signal’s shortlist included “Light it when the weather turns” and “Cedar, salt, and a stubborn grey sky.” The second won because it named two ingredients and a feeling in nine words.
For social, hand the model your winning tagline and the three messaging pillars, then request six captions in six fixed formats: one question, one behind-the-scenes note, one customer quote, one launch announcement, one ingredient fact, and one line of verse under 12 words. Fixed formats stop the feed from reading like six versions of the same post.
Step 5: Build the Audio in Two Passes
Voiceover first, music second, because the script length determines the track length. For 20 seconds, aim for 45 to 50 words. Read it aloud with a stopwatch before you generate anything.
With music, never ask for “a nice song.” Give a genre, a tempo and an instrument list: slow indie folk, 72 BPM, brushed drums, upright bass, muted nylon guitar, no vocals, no build until the final eight seconds. If the first pass comes back flat, this six-step Boomy workflow for finishing a raw AI idea as a track covers arrangement and mixing in more detail than a single prompt can.
Step 6: Assemble, Then Cut Hard
Four images, twelve lines of copy, one audio file. Export the hero at 1600×900, keep the gallery square, and upscale only what appears above the fold. Free image tiers often cap at 1024 pixels, and upscaling a soft image twice produces mush, so reshoot the prompt at a wider framing instead of enlarging the file.
If your timeline is tighter than two hours, the same sequence compresses well. This walkthrough on shipping a full campaign in 90 minutes with a free AI generator shows where to cut steps when the deadline moves up.
The Traps That Quietly Eat Your Afternoon
- Licensing drift. Free tiers change their commercial-use terms without much fanfare. Screenshot the current terms on the day you generate, not the day you publish.
- Credit resets. Most quotas roll on a 24-hour clock from your first generation, not at local midnight. If you burn out at 4pm, you are waiting until 4pm tomorrow.
- Invisible watermarks. Some free image tools embed a small mark in a lower corner. Zoom to 200% before you commit an image to a hero slot.
- Trusting your own eye. Plausible output passes review far too easily. A test in which AI image judges walked straight into a rigged trap is a good reminder that “looks fine at a glance” is not a quality check. Look at hands, shadows and reflections.
- Prompt drift. Saving only the final image loses the prompt that made it. Copy every winning prompt into a plain text file the moment it works.
Why the Second Launch Takes 40 Minutes
None of the value here is in the individual generation. It is in the artefacts you keep afterwards: the six-slot image template, the constraint list for copy, the six caption formats, the 45-words-per-20-seconds rule, the audio prompt with tempo and instrumentation.
Fog Signal was the slow run. The next scent in the Harborline range reused the locked image tail, the same pillar structure and the same caption rotation, and it took 40 minutes start to finish. Keep the rejects too. The candle shot that failed in March because the fog looked artificial tends to work in November for a different product entirely.
Free tiers also shift under you. Quotas shrink, models get swapped, new ones appear with better output at the same zero price. Re-read the tool landscape once a quarter, keep your templates, and treat the generators themselves as disposable.