10:06
AI product photography
Thomas Lundström demonstrates a workflow for turning simple product reference photos into 4K, photorealistic product images by pairing a custom ChatGPT prompting knowledge base with Seedream 4 (accessed via Higgsfield), using four Chance Chanel perfume color variants as worked examples.
The video presents a step-by-step workflow for turning a simple product reference image into 4K product photography using Seedream 4.
ChatGPT's deep research function is used to compile a report on the best ways to prompt Nano Banana and Seedream 4 for product images.
That report is converted into a PDF and uploaded into a ChatGPT 'project' (which allows up to five files) so the model consistently applies those best practices when drafting prompts, and can be added to later.
The prompting document created for the video is offered as a free download in the video description.
Two prompt-creation paths are shown: describing a vision directly to ChatGPT alongside the product reference photo, or converting an inspiration image found online/Pinterest into a text prompt via ChatGPT rather than feeding the image itself into the generator (to avoid directly copying someone else's work).
Seedream 4 is used via the Higgsfield platform, with settings for quality (standard vs. high), aspect ratio, and number of generations per prompt; high quality is used throughout for the crispest 4K results.
Four Chance Chanel perfume color variants (pink, purple, orange, green) are each placed into a distinct styled scene: ice/liquid, a purple-sand beach at sunset, a drink-making set, and a floral/outdoor setting.
Results are praised for product consistency and realistic lighting, reflections, and shadows, though label text accuracy declines when the label has more complex or smaller text.
Once a satisfying composition is generated in one aspect ratio, the same scene/product can be regenerated in other aspect ratios while staying consistent, yielding multiple versions for different use cases.
Getting a usable result is iterative — it typically takes several generations and prompt tweaks, and minor touch-ups (e.g., replacing text) may still be needed afterward; the presenter estimates near-perfect results almost 90% of the time if the product is simple enough.
ChatGPT Deep Research — A ChatGPT function that searches the internet and returns a detailed report on a given topic, used here to research the best ways to prompt Nano Banana and Seedream 4 for product images. Apply: Ask ChatGPT's deep research function to find and compile information on how to prompt a specific image model for product photography, then save the resulting report as a reference document.
ChatGPT Projects (custom project with uploaded files) — A ChatGPT feature that lets users upload up to five files into a dedicated project so the model follows those files' instructions across all discussions within it. Apply: Convert the deep research report into a PDF and upload it into a ChatGPT project so every chat session in that project automatically applies the documented best practices when drafting image prompts.
Image-to-prompt breakdown — A technique of feeding an inspiration image to ChatGPT and asking it to describe the scene as a text prompt, rather than feeding the inspiration image directly into the image generator, to avoid directly copying another creator's work. Apply: When you find an inspiration image on Pinterest or elsewhere, upload it to ChatGPT and ask for an image-prompt description of the scene, then feed that generated text prompt (not the image) into Seedream 4.
Seedream 4 (via Higgsfield) — An AI image generation model capable of producing up to 4K-resolution photorealistic images, accessed here through the Higgsfield platform with options for quality level, aspect ratio, and number of generations per prompt. Apply: Select 'high quality' mode in Higgsfield when generating Seedream 4 product images to get the crispest, highest-resolution results, and set the desired number of generations per prompt.
Aspect ratio switching — A Higgsfield/Seedream capability to generate the same scene and product in one aspect ratio, then switch aspect ratio and regenerate while keeping the product and scene consistent. Apply: After landing on a preferred product scene, change only the aspect ratio setting and regenerate to get landscape, portrait, or other crops of the same shot for different platform use cases.
Iterative prompt tweaking — The practice of treating image generation as a multi-pass process — generating several times and adjusting prompt wording to steer results toward the intended vision rather than expecting a correct result on the first try. Apply: Generate a batch of images from a prompt, evaluate what's off, tweak specific wording in the prompt, and regenerate — repeating until the output matches the intended vision.
Post-generation touch-up — A final manual editing step where imperfect details, particularly label text, are corrected in an image editor after AI generation. Apply: After Seedream 4 generates a near-final product image, manually replace or fix any inaccurate text or details in an image editor to reach a fully usable result.
Building a persistent ChatGPT 'project' pre-loaded with a curated prompting-technique document functions as a reusable prompt-engineering layer that decouples the one-time research step from every subsequent prompt-writing session.
Converting inspiration images into text prompts via ChatGPT (instead of feeding the image directly to the generator) serves double duty: it transfers compositional and stylistic information while avoiding what the presenter frames as directly copying another creator's work.
Aspect-ratio-preserving regeneration lets the presenter treat one successful generation as a virtual photoshoot, pulling multiple crops/formats from the same scene rather than starting over for each format.
Label text complexity, not overall photorealism, is identified as Seedream 4's main limiting factor for product photography — simple labels are nailed accurately while complex or small text boxes cause more errors.
The ~90%-perfect success rate framing implies the presenter treats Seedream 4 as a 'get most of the way there, then manually touch up' tool rather than a fully automated one-shot solution.
The video is explicitly framed as a follow-up correction to the presenter's own earlier Nano Banana vs. Seedream 4 comparison, made after he judged his prior Seedream experience insufficient — signaling the workflow shown here is a revised, more considered take.
«In this video, we're looking at my workflow of taking a simple reference image like this and using it with Cream 4 to create 4K resolution product images like these.»
— 00:00
«I went into chatgpt and used the deep research function that's available.»
— 01:05
«And instead of taking that image and just putting it directly into the image model, which I wouldn't advise doing because that's kind of copying others work directly, you can take that image and ask Chat GPT to break it down into an image prompt.»
— 02:44
«This video is not sponsored by Hicksfield, but I think I have a discount code that will give you 10% off Hicksfield if you want to use Seedream over on their platform.»
— 04:46
«Now looking at these results, I have to say that this is next level quality and I'm very impressed with what Seedream is actually capable of delivering in terms of product consistency and the quality of the environment.»
— 05:24
«Seedream has a bit more problems if the label includes more complex text or small text boxes that it needs to figure out, but the accuracy is still very impressive.»
— 06:08
«Now in this one what impressed me the most was the way it included also the shadows of the leaves above on top of the product itself.»
— 07:36
«But this is a kind of interesting also workflow. It's not taking photos of a product. It's tweaking the prompt and generating more variations so you slowly get something that is in line with what your vision is and what you want the model to create.»
— 09:08
«it's getting very close to having it perfect almost 90% of the time, if your product is just simple enough.»
— 09:35
Reception
Highly engaged and appreciative audience praising the video's helpfulness, concise format, and informative content, with constructive technical discussions around tool comparisons.
A tool-focused workflow video, framed as a follow-up correction to the presenter's own earlier comparison video, that pairs a custom ChatGPT prompting knowledge base with Seedream 4 to produce consistent, photorealistic 4K product images, while candidly noting the process is iterative and complex label text remains the main weak point.
10:06