ChatGPT Images 2.6: What's New and What GIFs Are Good For

A ChatGPT Images 2.6 thumbnail split between a still photo and a running GIF

When you use AI image generators, the first picture often looks great. It's the editing that gets frustrating. You ask for nothing but a different bag color, and the face changes, or you try to fix one finger and the furniture in the background changes too.

ChatGPT Images 2.6, released on November 4, improves image generation and editing, and adds a feature that turns photos or sketches into short GIFs. Let's look at which changes are actually worth trying.

〈OpenAI introducing Images 2.6 with new GIF and reference tools〉
* From the official OpenAI channel.

So What Is a GIF Good For?

A GIF is an image format that shows a short, silent motion on repeat. Think of the reaction images in a messenger app, like a cat blinking or a character nodding.

Compared with video, GIFs are especially small and light, so apps and web pages use them a lot to show something on repeat without a separate play button.

✦ GIFs vs. Video Generation Models

Which one fits which job

GIFFits short, silent, looping motion
Reactions, memes, product demos
Video modelFits longer action, camera moves, dialogue and music

A GIF isn't meant to replace long video. It's for simple clips that make sense even without audio. It works well for content where one movement is enough, like a character saying hello or a pet wagging its tail.

Also, the biggest strength is that you get a result right away without wrestling with a complex video editing timeline. That makes GIFs good for product intros too.

Think of a GIF of a bag zipper opening, or a shoe turning slightly to show its side, stuck on a product page. It grabs more attention than a still image, without the burden of playing a long video.

On the other hand, scenes with camera moves or several people acting in order, or content that needs dialogue and background music, clearly hit the limits of a GIF. For those, you need a video generation model.

An independent review outlet tried seven GIF prompts on the web version of ChatGPT Plus on November 5. It was an early test with each request run just once, but it shows fairly concretely which scenes GIFs suit.

✦ GIF Test Results

Output width about 249 to 439 pixels (Plus web test)

Café sign
Shop name stayed the same in every frame
Paper boat
First and last frames connect smoothly
Fox
Never got past the bus stop; kept walking in place
Night market
Sign text readable, small faces blurry

In the scene where a rainy-night café sign alternates between 'OPEN' and 'LATE' and leaves the shop's name on the window, the name stayed accurate in every frame.

The paper boat GIF, rocking on a pond, connected its last frame to its first so smoothly that looping didn't feel like an abrupt cut. With results like this, you could use it for café promo posts, short decorative images for a blog, or character reaction GIFs.

In contrast, when the model was asked to generate a scene of a fox in a yellow raincoat walking past a bus stop from left to right, the fox, raincoat and umbrella held up, but it never actually got past the stop and just walked in place.

In the complex night market scene, the sign text was readable, but the small faces were blurry. Early user reactions after launch also flagged smudges on a chessboard, flickering doors, and objects jumping position from frame to frame.

What it comes down to is that it can make short looping motions, but carrying a complex movement through smoothly is still hard.

Start From a Single Photo or Sketch

In ChatGPT, you can type some text to make a GIF from scratch, or upload an existing photo and ask for the people and objects in it to move.

✦ How a GIF Gets Made

Four steps from input to sharing

1Text or photo in
Start from scratch, or upload an existing photo
2Ask for motion
Describe the movement you want in words
3Edit in the chat
Revise the finished GIF in the conversation
4Speed, save, share
Change the playback speed, then save or share it

Say you upload a dog photo and type, "Make the dog tilt its head, then wag its tail, and repeat the motion naturally." The finished GIF can be revised again in the chat, and you can change the playback speed to save it or share it.

Instead of drawing frames one by one or adjusting a timeline in a separate animation program, you describe the motion you want in words.

By the way, the GIF feature is rolling out account by account, so keep in mind it may show up a bit later on some accounts.

The 'save reference photo' feature is also handy if you want to reuse a character. If you save a photo of your pet or mascot with a name, you can call up that name in a later conversation and use the same subject in new images.

You don't have to attach the original photo every time, so it helps when you're continuing work on a specific character.

What Users Complained About in GPT Image 2.5

✦ GPT Image 2.5: What Users Complained About

The same issues kept coming up

Backgrounds
Objects look duplicated and overlapping; erased items leave smudge-like traces
Decoration
Extra props added to empty space, making the scene busy
Hands and joints
Extra fingers or awkwardly bent arms
Text
Out-of-context promo text added again and again

Users of GPT Image 2.5 keep raising similar complaints. In complex backgrounds, leaves or objects show up duplicated and overlapping, and when you remove something, a smudge-like trace of it stays behind.

There were also complaints that even when you ask for a clean, empty frame, it adds props like potted plants, books and lamps and makes the scene busy, and that a 'cozy mood' request gets the same mugs, blankets and plants over and over.

If you get mad at ChatGPT about this, it says sorry and tries super hard, but once GPT Image's stubborn streak kicks in... prompt adherence drops off a cliff. At that point, I find it works better to open a new session and redraw from scratch.

Hands and joints are an old complaint too. Fingers multiply or arms bend awkwardly, and when two people hand something to each other, the hands sometimes don't properly touch the object.

Some users asked for a tiny fix several times and the texture of the face or fabric kept changing anyway. Unwanted text being added to images was a problem too.

One Reddit user pointed out that GPT Image 2.5 kept inserting out-of-context text like 'A better tomorrow' into landscape and product images.

A picture can look convincing, but once props and text nobody asked for show up, it's hard to use right away in a real thumbnail or ad, isn't it? And when you ask for an edit to remove them, it changes some other part too...

So How Much Better Is 2.6?

OpenAI announced that GPT Image 2.6 generates images 50% faster than before and keeps the details of characters, products and brands better across multiple rounds of edits. It also said it improved the layout of posters and product images.

But independent testing showed that not every ability got better. Let's look at the comparison tests that Flixly and zzo.ai each published.

✦ GPT Image 2.6 Test Results

From the comparison tests by Flixly and zzo.ai

CharactersFace and scar kept, camera direction accurate
TeacupsCat placed next to the blue cup
Recipe cardCleanest spacing and alignment
FishermanFiner textures
Coffee bagMore promo text added
Coffee machineStep numbers repeated

Let's start with the character consistency test. The same woman appeared in six scenes, and the task was to keep her hairstyle, freckles and eyebrow scar. GPT Image 2.5 and 2.6 both carried over her facial features.

The earlier 2.0 version made the face a bit thinner and lost the scar. On top of that, 2.6 got the camera direction more accurate and realistic, like drawing a rowing scene in side profile.

There was also a test with seven colored teacups, where the instruction was to put a black cat next to the blue cup. 2.0 and 2.5 put the cat behind the cup, but 2.6 seated it in the requested spot, next to the blue cup.

In the recipe card test, all three models wrote the ingredient names and temperatures correctly, and 2.6 had the cleanest spacing and alignment.

In the portrait of a fisherman mending a net, both 2.5 and 2.6 drew the hands naturally, and 2.6 rendered the net needle, beard and fabric texture in finer detail.

But there were results that came out worse, too. In the coffee bag test, 2.6 added more promo text nobody asked for than 2.5 did, and got one punctuation mark wrong.

On the complicated coffee machine diagram, 2.6 organized things in big, easy-to-read text but repeated step numbers, while 2.5 packed in more parts but some of the small text came out garbled.

Overall, 2.6 stood out with improvements in character consistency, position instructions and readable layouts. On the other hand, there were scenes where 2.5 actually gave cleaner results, like product packaging text or complex infographics.

Back when GPT Image 2.5 launched, some tasks actually looked better on GPT Image 2.0, remember?

On top of that, when the version goes up, the art style itself can change even for the same picture, and users who had been working with the earlier version sometimes struggled to keep things consistent.

✦ Generation Time

High quality, 1024×1024, each run once

Flare
• 2.5: 27 seconds
• 2.6: 14 seconds
Sunburst
• 2.5: 46 seconds
• 2.6: 52 seconds
Flare and Sunburst are separate tests with different prompts and platforms, so the two groups aren't compared directly.

Speed varied by task, too. OpenAI said generation is 50% faster, remember? In one independent test that ran the same high-quality prompt once each, Flare took 14 seconds on 2.6 and 27 seconds on 2.5, so 2.6 was faster.

In another service's measurement, averaging four high-quality Sunburst images, 2.5 took 46 seconds and 2.6 took 52.

Flare, which handles fast generation, showed the improvement clearly, but Sunburst, which produces more detailed results, didn't show a definite improvement.

〈OpenAI showcasing faster generation with Flare and Sunburst〉
* From the official OpenAI channel.

Wrapping Up

Honestly, back when image generation models first came out, they often botched even simple images, right? But now image generation models have been leveled up so much across the board that it seems hard to get very excited when a new version comes out.

The tech has come a long way, and our expectations have probably risen along with it. How much better can image models still get? Or are we already near the end of the road?

That's all for this post. See you next time.

Related posts

No comments:

No comments yet. Want to leave the first one?