OpenAI released ChatGPT Images 2.6 on November 4. The headline feature is one people have been faking for years: ChatGPT now makes animated GIFs. The second feature is less flashy but probably more useful. You can save reference images of yourself, your pet or a character and reuse them in new images without uploading them again (OpenAI on X, ChatGPT on X).

The launch post from OpenAI passed a million views on X within its first few hours. This article covers what is actually new, what is still unclear, and the prompting skill both features depend on: keeping a character the same from one image to the next. We show two prompts for that, with results you can make on EasyImagen today.

What's new in Images 2.6

From The Verge, The Decoder and OpenAI's own posts:

  • GIFs from a prompt or a photo. You can ask for something like "a gif of a cat dancing", or upload a photo and ask ChatGPT to animate it. After generation you can change the playback speed, look at the GIF frame by frame, edit individual frames and share the result.
  • Saved reference images. ChatGPT can keep reference images of "yourself, pets, and characters" ready for later, so a recurring character no longer means re-uploading the same picture every time (ChatGPT on X).
  • Faster generation. OpenAI says image generation in ChatGPT is faster. On the API side, two new models, GPT-Image-2.6 Flare and GPT-Image-2.6 Sunburst, generate images 50% faster than before, and Sunburst at medium or high quality "keeps details more consistent from one edit to the next" (OpenAI Developers on X).
  • Availability. Images 2.6 is rolling out to ChatGPT and Codex users on all tiers, on web, desktop and mobile. OpenAI also says people now make more than 4 billion images a week across ChatGPT and the API, up from 3 billion in September (The Verge).

Two caveats. The rollout is gradual: Decrypt could not get the GIF feature on its accounts when it tested, so you may not see it yet. And at the time of writing OpenAI hasn't published a full announcement with technical details, so frame counts, GIF length limits and API pricing are still unknown. Treat any precise numbers you see elsewhere with care until OpenAI publishes them.

Why a GIF model is really a consistency model

A GIF is a short series of still frames played in a loop. The Decoder describes OpenAI's dancing-cat demo as several slight variations of one character strung together flipbook-style. That only works if the character looks exactly the same in every frame: same face, same colors, same proportions, same outfit. Earlier models would drift a little between frames, and a flipbook made from drifting frames flickers.

Reference images solve the same problem from the other side: you lock what the character looks like once, then change only the scene. So the useful skill for both features is the same. Tell the model exactly what must stay fixed, and exactly what is allowed to change.

You can practise that with any strong image model. We made the examples below with GPT Image 2.5 Ext on EasyImagen. That is not Images 2.6, and we haven't tested 2.6 ourselves. But the prompts are written to transfer to it.

Prompt 1: a four-frame GIF sheet

Instead of asking for an animation, ask for every frame on one sheet. You get all the poses in one generation, and they share one style because they are one image.

A 2×2 grid of four animation frames for a looping GIF, on a plain off-white background with thin, even gutters between the four cells. The same character appears in every frame: a small, round orange tabby cartoon cat with a green bow tie, flat vector style with clean dark outlines, drawn at exactly the same size, feet on the same baseline, same camera angle and same lighting in every cell. Frame 1 (top left): standing upright, both paws down, smiling. Frame 2 (top right): leaning left, right paw raised high. Frame 3 (bottom left): small hop with both paws up, feet just off the baseline. Frame 4 (bottom right): leaning right, left paw raised high. Only the pose changes between frames; colors, proportions, face and bow tie stay identical. No text, no numbers, no frame labels.

A 2×2 grid of the same orange tabby cartoon cat with a green bow tie in four dance poses: standing, waving one paw, hopping with both paws up, and waving the other paw

The face, stripes, bow tie and outline weight hold across all four cells, which is the part that matters. Two things to notice. In frames 2 and 4 the cat tilts its head as well as its body, which reads as natural dancing but makes the loop a little bouncier than the prompt asked for. And frame 3 really leaves the ground, with the shadow staying behind, because the prompt said "feet just off the baseline". Phrases like "same size" and "feet on the same baseline" do a lot of work here: without them, the cat tends to change size from cell to cell and the GIF jitters.

To turn the sheet into a GIF, cut it into four equal tiles and play them in order. Any free online GIF maker can do this. With ffmpeg, it's two steps:

ffmpeg -i sheet.jpg -vf "crop=iw/2:ih/2:0:0" f1.png
ffmpeg -i sheet.jpg -vf "crop=iw/2:ih/2:iw/2:0" f2.png
ffmpeg -i sheet.jpg -vf "crop=iw/2:ih/2:0:ih/2" f3.png
ffmpeg -i sheet.jpg -vf "crop=iw/2:ih/2:iw/2:ih/2" f4.png
ffmpeg -framerate 5 -i f%d.png -vf "scale=480:-1,split[a][b];[a]palettegen[p];[b][p]paletteuse" -loop 0 dance.gif

Five frames per second suits a four-pose dance. For a smoother loop, ask for a 3×3 grid with nine poses that move in small steps, and raise the frame rate.

Prompt 2: a reference sheet you can reuse

Saved reference images are only as good as the reference. One photo from one angle leaves the model guessing about the back, the side and other expressions. A character sheet answers those questions up front.

A character reference sheet for a reusable mascot, on a plain light grey background. The character: a small, friendly round robot with a glossy mint-green body, a cream-colored face screen showing two simple black oval eyes, one short antenna with a coral-red ball on top, stubby arms and two small wheels instead of feet. Top row: the same robot shown full-body in front view, three-quarter view, side view and back view, all at the same size and evenly spaced. Bottom row: three close-ups of the face screen showing happy, surprised and sleepy expressions, plus four plain color swatch circles in mint green, cream, coral red and charcoal. Clean, consistent 3D render style, soft studio lighting, crisp edges. No text, no labels, no logos.

A character sheet of a mint-green round robot with a cream face screen and red-tipped antenna, shown from front, three-quarter, side and back, with happy, surprised and sleepy face close-ups and four color swatches

Then we passed that sheet back in as a reference image and asked for a new scene. The prompt repeats the character's defining details, because naming them is what keeps them:

Use the robot from the reference image as the main character and keep its design exactly the same: glossy mint-green round body, cream face screen with two black oval eyes, one short antenna with a coral-red ball, stubby arms and two small black wheels. New scene: the robot rolls along a sunny wooden kitchen counter carrying a tiny terracotta pot with a succulent in its arms, happy expression on its face screen. Morning light from a window, a cutting board and a bowl of lemons softly blurred in the background, shallow depth of field. Same clean 3D render style as the reference. No text, no logos.

The same mint-green robot on a sunny wooden kitchen counter, smiling and holding a small terracotta pot with a succulent, with a cutting board and lemons blurred behind it

The antenna, face screen, happy expression, wheels and colors all came through. The body came out a little taller and more egg-shaped than the round robot on the sheet, which is the usual way characters drift. If proportions matter, say so outright, for example "body exactly as wide as it is tall".

How to use this today

  • On EasyImagen: open the generator, upload your reference sheet as a reference image, and describe only the new scene plus the details that must not change. Save your best sheet so you can reuse it.
  • In ChatGPT: once Images 2.6 reaches your account, the same character sheet makes a good saved reference, and the frame-by-frame thinking above is how to describe the motion you want.
  • For more prompt ideas, browse the latest community prompts on the home page, or see how we wrote edit prompts for another new model in our MAI Image 2.7 post.

Whatever model you use, the rule is the same: list what stays fixed, then change one thing at a time.