Free Intensive Nov 21–22Free 2-day hands-on Vibecoding Intensive on Nov 21–22 at 16:00 MSK (Sat–Sun) · Live Cohort 3 starts Nov 23Join Free →Cohort 3 (Nov 23)
ReleaseInsight · On RadarHabr AI→ original

GPT Image 2 and Gemini Omni 1.2 Flash Animated 15 Pixel-Art Game Assets

Frame-by-frame animation for games is slow work, and a Habr author tested whether a pair of AI models could take over. GPT Image 2 draws a storyboard in pixel-art style, and Gemini Omni 1.2 Flash turns it into a video of up to 15 seconds with sound. Across 15 random assets, from a pistol to a katana, the results were generally good, but the background was not always transparent and the motion was too smooth. The workflow is fine for prototypes, but the final result will need manual polishing.

AI-processed from Habr AI; edited by Hamidun News
GPT Image 2 and Gemini Omni 1.2 Flash Animated 15 Pixel-Art Game Assets
Source: Habr AI. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceHabr AI ↗
Editor-in-Chief
Key metric #1GPT Image 2 runs nearly twice as fast as earlier GPT Image versions
Key metric #2Gemini animated the wheel on the first try, while the free Qwen did worse
Key metric #3Asked for "choppy" eight-bit animation, the model delivered smooth motion instead

Animating a sprite in an indie game usually means drawing dozens of frames: a muzzle flash, a fan unfolding, a wheel turning. A Habr author tested whether two AI models could handle this job together. The first, GPT Image 2, draws a storyboard in pixel-art style, that is, a "staircase" of consecutive poses of the object on a single sheet.

The second, Gemini Omni 1.2 Flash from Google, turns that sheet into a video up to 15 seconds long with sound. The author ran 15 randomly chosen assets from different genres through this pipeline, from a pistol and a katana to a campfire and a rolling boulder.

It is not a lab benchmark but a subjective practical test, and it still shows well where the pipeline already works and where manual touch-up is needed.

How the pairing works

The article describes GPT Image 2 as one of the best image generation models. According to the author, it has a single-pass architecture and runs nearly twice as fast as earlier GPT Image versions. Quality and cost per generation rise from Low to High mode, so you can choose between a cheap draft and an expensive final.

Gemini Omni 1.2 Flash can not only generate a clip but also accept edits in a "dialogue mode": a finished video can be adjusted with a new line without regenerating everything. For game development the scheme is logical: first fix the style with a picture, then animate it.

The test included a pistol with a magazine, ship thrusters with fire, a catapult, a sakura tree with falling petals, a three-story house with windows, a UFO, a campfire, a magic staff, a boulder, a falling brick tower, a Japanese fan, a lightsaber, a car wheel, a meteorite, and a katana.

What worked and what didn't

The pistol, staff, campfire, fan, and katana came out best. The pistol got a good shot storyboard and quality animation, though the author found the smoke superfluous. He attributes this to prompting rather than a model error. The catapult surprised him: the request described the motion only in passing, yet the model itself showed inertia after the volley. The boulder came out without glitches or the frame jitter that some other networks suffer from. Gemini animated the fan's sprite sheet very quickly.

The author calls the wheel a litmus test. Gemini got it right on the first attempt and convincingly conveyed the stroboscopic effect. The free Qwen, he says, handles the wheel animation worse.

The shortcomings repeated across assets:

  • Background. The model was asked to render on a transparent PNG background but did its own thing. The boulder got a strange light, like a theatre spotlight, and the UFO was painted green.
  • Too smooth. For the sakura and the ttower, the author asked for "choppiness" to make the animation resemble eight-bit. Gemini made the motion more modern and smoother.
  • Extra details. In the UFO's last two seconds, an already landed saucer for some reason levitates, and an unclear haze appeared around the lightsaber.
  • Floating background. For the meteorite, the author did not get enough rigid locking of the object relative to the background. He advises explicitly requiring that the object's position on the background be fixed firmly when animating assets.

What a developer should do

The author's conclusion: Gemini Omni 1.2 Flash animates quickly, understands the task from a hint, and does not fantasize all that often. GPT Image 2 is praised for speed, good context memory, and consistent style across generations.

In practice this means the following:

  1. For prototypes, background effects, and references, the pairing already works. Fire, smoke, falling objects, and magic come out best.
  2. Write requirements for the background and for "choppy" frame rate directly into the prompt, but be prepared for the model to ignore them.
  3. Transparency and pixel sharpness will most likely have to be finished in an editor: cutting out the background and manually lowering the frame rate.
  4. Dialogue mode is best used for targeted edits, not for reworking the whole clip.

The limitations are obvious. The assets were picked at random, but the ratings are subjective, without metrics, and comparisons with other models are given only verbally. For a game's final release, the results must be checked by eye. Still, as a quick way to get a working draft animation from a single image, the pairing already deserves a place in the toolkit.

Story

Key Questions Answered

How the pairing works?

The article describes GPT Image 2 as one of the best image generation models. According to the author, it has a single-pass architecture and runs nearly twice as fast as earlier GPT Image versions. Quality and cost per generation rise from Low to High mode, so you can choose between a cheap draft and an expensive final. Gemini Omni 1.2 Flash can not only generate a clip but also accept edits in a "dialogue mode": a finished video can be adjusted with a new line without regenerating everything.

What worked and what didn't?

The pistol, staff, campfire, fan, and katana came out best. The pistol got a good shot storyboard and quality animation, though the author found the smoke superfluous. He attributes this to prompting rather than a model error. The catapult surprised him: the request described the motion only in passing, yet the model itself showed inertia after the volley. The boulder came out without glitches or the frame jitter that some other networks suffer from. Gemini animated the fan's sprite sheet very quickly.

What a developer should do?

The author's conclusion: Gemini Omni 1.2 Flash animates quickly, understands the task from a hint, and does not fantasize all that often. GPT Image 2 is praised for speed, good context memory, and consistent style across generations. In practice this means the following: For prototypes, background effects, and references, the pairing already works. Fire, smoke, falling objects, and magic come out best.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Unified AI Workspace · Nomad

Test Frontier LLMs & Autonomous Agents in One Workspace

Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.

Try Nomad AI →Launch Nomad Agent
What do you think?
Loading comments…