The same model in every image and in video: a brand character that does not change

The problem

A brand that wants the same model in dozens of images needs her to be recognisable: same face, same hoodie. With a text prompt alone, every image comes out with a different person. In video there is one more risk: the face can change inside the clip itself.

What we did

1. We created the model from scratch, from text only. No photo of any person was used.
2. We trained our own model on 19 synthetic images of her. Training cost USD 2.75 of rented compute (that one run only, not our time or earlier tests).
3. We generated the scenes with that model. The exact hoodie is swapped in afterwards, the face is cleaned up if it ends up small, and the image is upscaled to 4K. That is what you see on the Brands page.
4. We animated one of those images. We started from the café scene, vertical, and asked for a 5-second clip.

Results

• One 5.04 s clip (121 frames at 24 per second), which takes 168 s to generate on a single machine.
• The face holds. It is the same model as the starting image through the whole clip, judged by eye. We have no similarity metric.
• The first attempt looped. With a short prompt, she drank again at the end of the clip. With a longer prompt saying the action happens once, she did not.

What still fails

• The face gets a little rounder towards the end. We took it for a gesture and it still reads as her, but you can see it if you look closely.
• A clip is a single take of about 5 s. We tried chaining two clips to make it longer and got an odd cut, with a strange hand movement at the seam. We dropped it.
• Slowing down or interpolating does not help. Interpolating for smoother motion gave a slow-motion look. We avoid it.
• Small text in motion. In other tests, the embroidered logo on the hoodie warped when it moved. That is why we do not do close-ups of the logo in video with this technique.
• No sound. The clip has no audio.

What a client gets

The same, with your product or your character. From the Brands page:

• Test with your product, EUR 150. You send three photos and in 72 hours you have six images in two scenes. It is credited if we continue.
• Your brand's model, from EUR 450. Your product, your character or your style trained and tested, delivered in 3 to 7 days.
• Video. Short product clips from the same references. Ask us for a quote at hello@grawwww.xyz.

We do this for brands

Commercial section. Grawwww sells the work this article describes.

If you prefer not to depend on a real model, we create a character of your own for your brand. If you use a real person, only with their written consent for that use.

Write to us with your case: hello@grawwww.xyz. We answer in under 48 hours.

Every image and the video in this article are AI-generated. Fictional model and brand.