09:05
AI product photography
Lundström walks through a real 2025 client project for the Finnish canned-water brand Arcti to show his end-to-end AI product-image workflow (Pinterest reference sourcing to ChatGPT prompt writing to generation on Higgsfield) and a beginner-level Photoshop method for fixing the label-text and background errors that AI image tools introduce.
The video answers the two questions the channel gets asked most: the AI image workflow/prompt-creation process, and how to fix label text errors on AI-generated images.
Case study is a real client project: Arcti, a newly launched Finnish canned still/sparkling water brand, for which the client supplied only two rendered product files and a mood board.
Workflow starts with 10, 20 or even 30 minutes browsing Pinterest for inspiration and reference images.
A screenshot of the chosen reference image is sent to ChatGPT to be broken down into an image-generation prompt.
Product photos and a style reference can be sent to ChatGPT together so it describes the product rendered in that style as the final image.
Lundström uses a custom-trained GPT (linked in the video description) that automatically converts an image into a correctly formatted prompt for different image models.
All images shown were generated using 'Cadream' or Nanobanana Pro on Higgsfield, with 4K resolution recommended for a photorealistic look.
The video states it is not sponsored by Higgsfield, though a 10% discount code is included in the description.
Tip: always request condensation on cans/bottles in the prompt, since it makes renders look far more realistic and avoids the 'too plastic' look of a bare product render.
The label-fix demo starts from an image (generated with 'Cream 4') he likes despite its flaws: an infinite black background with cans half-submerged in a water tank.
Text fix: select and remove the AI-garbled text with the polygon lasso and remove tool, then place the real text layer from the original label design file on top, aligning it to the blurred text still visible underneath; repeat for the logo and the second can.
Background fix: remove the visible water tank using Photoshop's generative fill, then re-select and re-request to remove what the first pass didn't fully clear.
Graphic fix: import the correct graphic element from the design file, align it to the label, paint out the unneeded parts, and adjust brightness so it matches the surrounding label.
The full retouch took about 10 to 15 minutes and turned the flawed original into the finished image.
Lundström states the method won't work if the label error is bigger or more complex than a simple line of text.
Takeaway framed for viewers: an AI image with a small error doesn't need to be discarded — it can often be salvaged with light Photoshop editing.
Pinterest reference sourcing — Spending 10-30 minutes browsing Pinterest to find inspiration and reference images before starting an AI image project. Apply: Browse Pinterest for roughly 10 to 30 minutes to find a style or composition to build the prompt around before opening an AI tool.
ChatGPT prompt breakdown — Taking a screenshot of a reference image and asking ChatGPT to convert it into a written image-generation prompt. Apply: Screenshot the chosen reference image, send it to ChatGPT, and ask it to break the image down into a prompt for an AI image model.
Product + style reference combination — Sending both a product image and a separate style-reference image to ChatGPT so it writes a single prompt describing the product rendered in that style. Apply: Send the product render and the style reference image together to ChatGPT and ask it to describe the final image with the product rendered in that style.
Custom GPT for image-model prompting — A custom-trained GPT (linked in the video description) trained on instructions for how to phrase prompts for different image models. Apply: Feed a reference image to the custom GPT so it automatically converts it into a correctly formatted prompt for the specific image model being used.
Condensation prompting tip — A prompting trick where requesting condensation on cans or bottles makes AI-generated renders look markedly more realistic and less 'plastic.'. Apply: Add a request for condensation on the can or bottle to prompts whenever generating beverage-container product images, especially when starting from a bare product render.
4K resolution requirement — Using 4K output resolution as a stated requirement for achieving a photorealistic look in generated images. Apply: Set or request 4K resolution when generating product images that need to read as photorealistic.
Label text replacement (Photoshop) — A beginner-level Photoshop method for fixing AI-garbled label text by removing it and overlaying the real text from the original label design file. Apply: Use the polygon lasso tool and the remove tool to erase the AI-generated text area, then place the real text layer from the design file on top and align it to the blurred text still visible underneath before flattening.
Photoshop generative fill for background cleanup — Using Photoshop's generative fill to remove or replace unwanted background elements, such as a visible water tank, in a generated image. Apply: Select the unwanted background area and run generative fill, repeating the selection and request on remaining bits if the first pass doesn't fully remove the element.
Design-file graphic overlay — Fixing distorted or broken label artwork by importing the correct graphic element from the original design file, aligning it, and blending it into the label. Apply: Bring in the correct graphic from the design file, align it with the corresponding label section, remove the excess parts, paint out the edges to match the surrounding label, and adjust brightness so it blends seamlessly.
The condensation trick is framed as a fix for a specific named failure mode (bare product renders looking 'too plastic'), not a generic realism tip.
The AI-generated garbled text isn't erased first and then rebuilt — the blurred text is deliberately left visible underneath so the real text layer can be aligned to it, which is what makes the fix fast.
The pipeline always converts a visual reference into an editable text prompt (via ChatGPT/custom GPT) before generation, which is what lets him tweak the description afterward rather than regenerating blindly.
Even inside an AI-generation workflow, the client's original design files remain load-bearing: the actual label fix is re-importing real vector/text assets in Photoshop, not a better prompt.
He deliberately sets a low skill floor for the segment ('I myself is by no way, shape, or form a super seasoned veteran in Photoshop'), positioning the fix as accessible to non-designers rather than a professional retouching tutorial.
He pre-empts the obvious objection himself, stating upfront that the method only fixes small text/visual errors and would not work on bigger or more complex label problems.
«In this video, I'm going to show you a real client project where I used AI tools to produce professionallooking product images like these.»
— 00:00
«So, we're going to take a real client example from last year, take a look at the workflow more in detail, and then I'll show you how to fix some of the errors that sometimes appear when you use AI generated tools.»
— 00:58
«Now, whenever I create anything using AI tools, I usually start at Pinterest looking for inspiration and reference images to start the creative process.»
— 01:44
«And I use a custom GPT that I have trained on some instructions of how to prompt image models.»
— 03:05
«Now, a quick tip when generating images for cans or bottles is to always ask for condensation on the can or bottle itself.»
— 03:59
«I want to start out this section by saying that you definitely don't need to be super advanced and expert level in Photoshop to fix these errors.»
— 04:35
«I myself is by no way, shape, or form a super seasoned veteran in Photoshop.»
— 04:48
«And with just a quick 10 to 15 minutes Photoshop retouching, we ended up with an image that looked like this.»
— 07:46
«And of course, you would be completely correct. But this is just a quick fix method to fix those small errors in text and maybe visuals that would otherwise ruin an image that has something really nice in it.»
— 08:04
Reception
Highly enthusiastic audience expressing genuine gratitude and appreciation, with constructive feature requests and strong engagement throughout.
The video delivers a genuine, reproducible protocol — reference sourcing, prompt writing, generation, and a specific Photoshop repair method — rather than abstract theory, but the demonstrated label fix works cleanly only because the errors were small and isolated and the client's original design assets were on hand to paste back in.
09:05