AI E-Commerce Photography — A Workbench Built by a One-Person Company, Four Libraries and One Click

I was talking with an investor friend recently. He said the AI companies that raised money over the past few years have all had a rough time, with headcount and operating costs pressing down on them while they still have to iterate fast.

Here’s my take: a product built by a one-person company isn’t necessarily worse than one built on venture money. AI has moved so far this past year that what used to take a team can now be shipped by one person plus a few agents.

What I’m demoing today is the AI e-commerce image workbench a designer friend and I spent a bit over half a month polishing.

1. Why this isn’t a “one-click e-commerce image” tool

Most AI e-commerce image tools on the market work like this: upload one product photo, get five e-commerce images in one click. And the result? Every model swap gives you the same face, the scenes are interchangeable, and the logo is a blur.

The hard part of e-commerce imagery isn’t “generating.” It’s accumulating.

Style, scene, model, product — every e-commerce team is building these up, and not one tool helps you store them and reuse them. So we built four libraries.

2. Style library: upload one street-style shot, the AI pulls out the tone

Once you’re in a project, the first module is the style library.

I upload a street-style photo and the AI analyzes its tone, lighting, color temperature, composition, and texture. Once those parameters are banked, every downstream generation step can call that style, so the whole set of e-commerce images shares one visual register.

This isn’t slapping on a filter. It’s decomposing “why does this image look good” into reusable parameters.

3. Scene library + model library: upload an image and it expands

The scene library works the same way. Upload a photo of a tree-lined park path and the AI expands that scene into different angles and different framings — wide, medium, close-up, beside the bench, at the end of the path — a whole set in one pass. Later, when you’re picking a scene, you just choose from it.

The model library follows the same logic. Upload a photo of a distinctive model and the AI extracts her face, then generates full-body looks from the front, three-quarter, profile, and back. She becomes your “digital model,” ready to be re-signed and called up again and again.

E-commerce companies with models under contract can upload them through a digital asset licensing flow too. One real model, unlimited shoots.

4. Product library: one outfit photo, individual items parsed out

The cleverest piece is the product library.

In the demo I upload a single photo of a styled model and the AI parses it automatically: the top is a white strapless bustier, then the skirt, the shoes, the bag. Four separate items parsed out and filed.

And it checks against the database: if that image already exists in the product library, it flags “already recognized” and skips it. Which means a team can keep accumulating its item library, and it gets cheaper the more you use it.

5. One-click generation: pick a model, a scene, a focal length, a light

With all four libraries in place, the next step is generating.

Same logic as an actual shoot — pick the model, pick the head (hat or no hat), pick the top, the bottom, the shoes, the bag, then the aspect ratio (1:1 / 3:4), the framing (full body / half body / portrait / product close-up), the focal length (portrait / wide), the camera angle (high / eye level / low), the color temperature, contrast, hard light or soft light, expression.

Every one of those is a photographer’s vocabulary.

6. Two engines, each doing what it does best

We went with two generation engines:

• Nano Banana 2: extremely strong at holding character consistency, and the overall style feels natural
• GPT Image: slightly stiffer, but noticeably better at controlling clothing logo detail

After generation, if the logo is still muddy, you can marquee-select it and hit logo refinement — it isolates the logo region and regenerates it alone, preserving the brand detail.

Wrapping up

If you’re in e-commerce, running an apparel brand, a photo studio looking to pivot to AI, or a one-person company trying to build a brand — this workbench genuinely works.

Want to try it? Let’s talk in the comments.

Source: AI拍摄电商图.mp4 · ASR model gemini-2.5-pro · full text of the 11-minute original recording

[00:00] Hello, today let’s talk about some uses of AI in e-commerce. We’re going to use AI to make some e-commerce product shots. Honestly, look — a one-person company sometimes builds products that aren’t any worse than what those funded AI companies build. AI has moved so fast lately.

[00:19] Uh, a few days ago I was talking with an investor friend, and he said that over the past few years the AI star companies that raised money have actually had a pretty rough time. Because they carry so much headcount cost, so much operating cost, and their product has to iterate really fast for them to have any chance of holding a leading position in the market.

[00:37] So what I’m sharing today is a workbench for making e-commerce assets, something I’ve been polishing over the past half month or so together with a designer friend. Let me walk you through it.

[00:50] So once you’re into this workbench you can create a project, and you can also invite — ah, WeChat login here — you can invite your own colleagues and partners into the project. And once you’re in the project, it’s split into a few sections. This section here, it’s a style library. We can go find image styles online that we really like. And here we can upload the reference image style we want.

[01:24] Hmm, let me see. Say I find a street photo. Ah, it warns me it’ll deduct credits — we hit confirm, and it analyzes the image: its tonality, its lighting, its color temperature, and its texture and composition. Then later in the image-generation step we reference that style, and it learns it fast — you can see it come through very directly in the generated results.

[01:58] And in the scene section we can also find images to pull their scenes out of. For example, here I upload — ah, a natural environment, a natural environment. You can see it’s a natural setting, like a park. It analyzes the image and generates shots of that scene from different angles and different framings, which can then serve as background references.

[02:29] And in the model section we can find some good model looks. Say we see a very distinctive model online — we can upload their photo, and it can extract a model from it. Then later in the generation process we can have that model try on our clothes and products.

[03:02] Ah, the scene is still generating. OK, and products, products work the same way in there. I can upload a product photo to get detail shots of it. Say I click into these shoes — it gives multiple angles, including some logo close-ups. I can upload a product — I can upload ten or twenty photos of the product from different angles, or I can just upload one photo of a model wearing the clothes.

[03:32] Say I upload — ah, that model from a second ago — upload a photo of a model in a scene. It can automatically recognize whether the model has a hat on, what kind of clothes they’re wearing, what kind of shoes. It can pull those individual items out of the model. Or you can just upload the product photos we paid a photo studio to shoot — that works too.

[04:06] So now it’s recognized them: the top is a white strapless corset, then the bottom is this, the shoes are this, the bag is this. It can also match against the images in the whole database — some products may have been uploaded and recognized already, and then it’ll flag here that the library may already have this image, so it won’t check that one. Since none of these 4 images are in the library, we check all of them and hit submit to library. It creates that recognition process for all four items, and it can run in the background.

[04:39] So let’s go back and look at that scene, it should be done recognizing. You can see this scene: out of the single image we uploaded, it recognized this as a tree-lined path in a park. It pulled the scene out, plus some close-ups, some shots with benches. And we’ll use that as a reference when we generate the product photos.

[05:00] Same thing, let’s look at the model. On the model side, the head is already generated, let’s wait for it to load. You can see it pulled a very distinctive, fashionable model out of that image. And now it’s generating her full-body shots. The full-body set is front, three-quarter, profile and back. And you can all see she’s wearing nude-colored underwear — that way, when it’s used as a reference image for generation, it won’t skew the image. It lets the model put on all different kinds of clothes.

[05:35] These first few are models we pulled out of some great images we found online. Of course, if our company — an e-commerce company — has models under contract, you can also license the digital-asset rights for that model. Then we can use their likeness to generate the model. And of course you can also just make one up from scratch, whatever you prefer.

[05:59] Let’s check that product from earlier. It’s still being recognized. OK, so let’s look at the generation step then. In the generation step, we generated a model, right, and we uploaded product photos, and we can upload scene photos too.

[06:20] So in this environment we go about it just like we’re shooting a photo. Look at how this image came out: for this one we can also click to zoom in, and the detail is still very clear. For this image we pick a model for the subject position, or we say this image has no model at all — it’s an invisible mannequin, right, where the clothes look like they’re being worn by an invisible person. Invisible mannequin works fine. And you can also generate a flat-lay of the product.

[06:48] Pick the model here, and then here we can choose whether they’ve got a hat on their head or not. Then what kind of clothes they’re wearing, what kind of top, what kind of bottom, what kind of shoes, whether they’re carrying a bag, whether they have any other accessories.

[07:08] Then the aspect ratio of the generated image — 1:1, 3:4, or some other ratio. And whether it’s a full-body shot, a half-body shot, a portrait, or a product close-up. And the focal length, here you can also pick a common focal length: portrait, documentary, or a wide-angle focal length.

[07:26] High angle, eye level or low angle; the color temperature of the photo; the contrast of the photo; and whether the light is hard or soft — these are pretty professional photography terms at this point. And the model’s expression, whether it’s natural or a smiling, happy expression. Or when we have special requirements for the model’s pose, or for the clothes, we can just write a description right here.

[07:49] Same thing, we can pick whether this was shot in a studio or in an outdoor scene. Hey look, the scene we generated earlier is already selectable — that tree-lined park path is already selectable. And for the style we can use a retro film look, or a modern-city street-style look. Let’s go with the retro film look.

[08:11] Then for the generation engine, here you can actually choose the Nano Banana 2 model or the GPT Image 2 model, and each of these has its strengths. Banana holds character consistency and overall style pretty naturally, and its character consistency is very strong. GPT’s new image model is a little stiffer, but it has stronger control over the details of garment logos.

[08:36] And once the image is generated, I can also click logo touch-up. It pulls out close-ups of the product logos that appear in the photo for me to check off. Whichever logo I check means, hey, this logo maybe didn’t come out very clearly, and now I want it repaired. Then it can go into a touch-up pass and fix that logo nicely.

[09:00] Let’s take a look, the full-body model shots are done. My connection is a bit slow today, the image hasn’t loaded. Right, so basically that model is now sitting in our library, and later we can have her shoot lots of our products.

[09:25] And some of that product actually came out too. Look, including this one — hey, this one came out a bit wrong. But her dress generated correctly. Right, we can see the dress the model herself is wearing. OK, so that top can just be regenerated.

[09:54] Then let’s see how her shoes came out. Look at some of the photos we generated before — the fidelity is very high. The fidelity is very high. Those shoes still haven’t come out.

[10:15] OK, let’s go back to that screen. Let’s use this newly picked model — let’s skip the hat, a white top, purple shoes. Let’s grab a bag too. A half-body outfit shot for her, let me see.

[10:40] Then hard light, high contrast, natural expression. On that tree-lined path, film style. We — for summer clothes we should be able to pick that dress from earlier. Let’s put it together like this and see what kind of result it generates.

[11:09] And there it is. The styling it put together is a little odd, honestly — a skirt on the bottom, a knit cardigan on top, and a purple bag too. Right, it just generated it straight off. The detail is still very good.

[11:34] OK, that’s it for today’s share, thanks everyone.