Dev.to WebDev 🛠 Dev 👁 0 📖 3 min read

GPT Image 2.6 Review: Still Images Barely Changed, GIFs Are New

OpenAI released GPT Image 2.6 on November 4, in two versions, Flare and Sunburst. Over the past few days we compared it with 2.5 and tried the new GIF option. What we found: still images look about the same as 2.5. GIFs

GPT Image 2.6 Review: Still Images Barely Changed, GIFs Are New

OpenAI released GPT Image 2.6 on November 4, in two versions, Flare and Sunburst. Over the past few days we compared it with 2.5 and tried the new GIF option.

What we found: still images look about the same as 2.5. GIFs are new. They're fine for stickers and small icons, but they don't hold up on busier scenes.

Still images look about the same as 2.5

We gave both versions the same four prompts: a fashion lookbook photo, a product shot on marble, a YouTube thumbnail and a long how-to graphic for Pinterest.

On the first three, we couldn't say which was better. Skin, fabric and reflections looked about the same, and both spelled the thumbnail headline right.

The how-to graphic showed the clearest difference. Both versions spelled every word right, but 2.6 (right) used bigger text and cleaner cards, so it's easier to read on a phone.

The same how-to graphic from GPT Image 2.5 Flare (left) and GPT Image 2.6 Flare (right)

OpenAI says 2.6 is 50% faster. In our runs, a Flare image took 12 to 31 seconds and a Sunburst image 24 to 42 seconds. We never saw the six seconds OpenAI mentions.

If you only make still images, moving from 2.5 to 2.6 won't change much.

GIFs are new

2.6 can make an animated GIF from a prompt, with or without a reference image.

You can leave the settings alone and let the model pick the size and length; our cat sticker came back as 9 frames over 1.8 seconds. Or you can set the aspect ratio, the number of frames (2 to 16) and the length (1 to 5 seconds) yourself.

Simple loops come out well:

An orange cartoon cat sticker waving one paw, made with GPT Image 2.6

A green mug turning slowly from left to right and back, made from a reference photo

The mug started as one product photo. The handle turns behind the mug and back again.

Problems with the GIFs

  • 16 frames at most, so the motion looks choppy, especially in a 3-second GIF.
  • Each frame is drawn separately, so details change from frame to frame, and edges and backgrounds shake. In the coffee GIF below, the background behind the steam keeps moving.
  • Busier scenes look worse. A plain background with one small movement is fine. People or a street scene are not.

A cup of coffee with rising steam; the background shifts slightly between frames

A red SALE badge pulsing; the edge of the circle wobbles slightly

On a sticker or a sale badge you'll hardly notice the shaking. To bring a landscape or portrait photo to life, it isn't good enough.

If you need better GIFs

To make a photo move like real video, use a photo-to-GIF tool built on a video model: upload a photo, describe the motion, and you get a 5-second animation.

This one was made from a single photo with the CVY.AI AI GIF generator. Her hair and dress move in the wind and birds fly past, and the motion stays smooth for all 5 seconds. With 16 frames at most, 2.6 can't do this.

A woman in a white dress on a cliff at sunset, her hair and dress moving in the wind while birds fly past

Which one to use

You want Use
A still image GPT Image 2.5 or 2.6
A poster or infographic with text GPT Image 2.6, slightly better
A sticker, icon or sale badge GPT Image 2.6 GIF
Your own photo moving A photo-to-GIF tool

Tips for GIFs with 2.6

  • Ask for one movement at a time: "waves one paw", not "jumps and spins".
  • Add "then returns to the starting pose" so the loop is smoother.
  • Add "fixed camera" and "plain background".
  • Keep text on the image to one or two words. Longer text changes from frame to frame.
  • Use about 8 frames for a small bounce and 12 to 16 for a slow turn.

Both versions of GPT Image 2.6, GIFs included, are in the CVY.AI image generator, no OpenAI account needed.

📰 Read the original article on Dev.to WebDev

Originally published by Dev.to WebDev. Aggregated on AIWithGhost for educational purposes — full credit and traffic to the original publisher.