What Are Generative Images and How Are They Created?

Updated

Generative images are pictures synthesized by a machine learning model rather than captured by a camera or assembled by hand in an editor. You give the model an input — a text prompt, a reference photo, or both — and it produces new pixels that match that input. They are used in practice for ecommerce visuals, headshots, and campaign imagery; Astria, for example, describes its product as AI image production for fashion and identity, covering ecommerce studio work, lookbooks, PDP and campaign images, and headshots. The trade-off is that you get speed and scene flexibility at the cost of occasional artifacts and the need to check consistency and usage rights before publishing.

How generative images differ from edited or captured photos

A photograph records light that existed. An edited photo starts from that record and changes it. A generative image starts from a model's learned representation of what images look like and produces a new one.

That distinction matters for three practical reasons:

  • Provenance. There is no original scene to point to, so you cannot verify "this is what it looked like."
  • Variation. You can generate the same subject in a scene that was never photographed, which is the core selling point of tools like Astria's "same model. Any scene."
  • Failure modes. Instead of noise or blur, generative images fail with structural artifacts — extra fingers, warped text, inconsistent logos.

The main model families

GANs (generative adversarial networks)

Two networks compete: a generator that produces images and a discriminator that tries to tell real images from generated ones. The generator improves by trying to fool the discriminator. GANs are fast at inference and were the basis of much early face and product imagery, but they are harder to steer with detailed instructions and are prone to mode collapse, where the model keeps producing near-identical outputs.

Diffusion models

A diffusion model learns to reverse a noising process: it is trained to remove noise from images step by step, then at generation time it starts from random noise and denoises toward an image that fits the prompt. Diffusion underpins Stable Diffusion, which Astria lists among its keywords, and it is generally better than GANs at following complex text prompts and at fine-tuning on a specific subject.

Fine-tuning approaches

To make a model produce a specific person or product rather than a generic one, you fine-tune it on a small set of reference images. Dreambooth, which Astria also lists, is one such method: it teaches the model a new subject so you can then place that subject in new scenes. This is the mechanism behind "train, generate, deliver" workflows.

The typical pipeline

  1. Train or fine-tune. Upload a set of reference images of the subject (a model, a product, a face). The model learns that identity. Quality and consistency of the input set largely determine the output.
  2. Prompt. Describe the scene, lighting, framing, and style you want. The prompt is the main control surface.
  3. Generate. The model produces candidate images. Expect to generate several per usable result.
  4. Select and iterate. Pick the best candidates, adjust the prompt or references, and regenerate. Selection is a real part of the work, not an afterthought.
  5. Deliver. Export at the resolution and format the destination requires — a PDP image, a lookbook spread, a campaign asset.

Where generative images are used in practice

Use case What the model is trained on What you get
Ecommerce studio Product or model references Lookbooks, PDP images, campaign shots without a physical shoot
Headshots A person's photos Professional-looking portraits in varied settings
Identity-consistent scenes One subject, many contexts The same person or product across multiple scenes

Astria's featured workspaces — Maison Urbane, Great Sports, America Basics — illustrate the ecommerce pattern: one trained subject, many generated scenes.

Common limitations to plan around

  • Artifacts. Hands, text, jewelry, and fine patterns are frequent failure points. Inspect at full resolution before publishing.
  • Consistency. Keeping a face or product identical across many images requires a good training set and often manual review.
  • Rights and usage. Training data, likeness, and commercial usage terms vary by tool and jurisdiction. Confirm what you are licensed to do with outputs before a campaign, not after.
  • Resolution and format. Generated images may need upscaling or retouching to meet print or platform specs.

If your goal is a one-off creative image, a general diffusion tool is enough. If your goal is the same subject across many scenes for commerce or identity, a fine-tuning workflow built around a trained model is the better fit.