GIF 配方寫死：1:1 · 16 幀 · 12fps

提示詞

I’ve seen quite a few ChatGPT Image 2.5 GIF workflows and run a few tests myself. My favorite combination so far is a 1:1 canvas, 16 frames, and playback at 12fps.

For this one, I used the API with max + 4K to generate a sprite sheet of a cat in a fried-shrimp costume doing a skateboard flip. Then I gave the image to Codex to split into frames and assemble the GIF. The details came out noticeably clearer than the earlier versions I generated with Codex’s built-in imagegen tool.

Here’s a simple example using my 2880×2880 source image:

→ 3×3 grid, 9 frames: 960×960 per frame
→ 4×4 grid, 16 frames: 720×720 per frame
→ 5×5 grid, 25 frames: 576×576 per frame

The more frames you pack in, the fewer pixels each frame gets. That includes the background, too. A small character with lots of empty space around it leaves fewer pixels for the details.

The same 16 frames play in one second at 16fps. I ended up exporting at 12fps, with the GIF lasting about 1.33 seconds. The motion felt less rushed.

Good-looking individual frames don’t always make a smooth animation. Once it’s playing, you can still spot inconsistencies in the board’s rotation and wheel positions. There’s still a bit of luck involved, haha.

Prompt in the first reply. Paste it into a Codex session that can generate images, swap out the character and action, and give it a try 👇

中文翻譯

我見過不少 ChatGPT Image 2.5 的 GIF 工作流，自己也跑了一些測試。目前我最喜歡的組合是 1:1 畫布、16 幀、12fps 播放。

這一次我用 API 搭配 max + 4K 生成了一張家貓穿着炸蝦服玩滑板翻板動作的 sprite sheet。然後我把圖片交給 Codex 切分成幀併合成 GIF。相比我之前用 Codex 內置 imagegen 工具生成的早期版本，這次的細節明顯更清晰。

下面用我的 2880×2880 源圖舉個簡單的例子：

→ 3×3 網格，9 幀：每幀 960×960
→ 4×4 網格，16 幀：每幀 720×720
→ 5×5 網格，25 幀：每幀 576×576

塞進去的幀數越多，單幀分到的像素就越少。背景也一樣。一個周圍留有大片空白的小角色，留給細節的像素就更少。

同樣的 16 幀，在 16fps 下 1 秒就播完了。我最終按 12fps 導出，GIF 時長約 1.33 秒。動作感覺沒那麼趕。

單幀好看並不一定就能組成流暢的動畫。一旦播放起來，你還是能發現滑板旋轉和輪子位置上的不一致。多少還得靠點運氣，哈哈。

提示詞放在第一條回覆裏。把它粘貼到能生成圖片的 Codex 會話中，替換掉角色和動作，試試看 👇

溯源

原作者

Voxyz_ai

發佈日期

2026-09-10

原始出處

查看原帖 ↗

收錄來源

xianyu110/awesome-gpt-image2.5 · 2098063955096633467 · MIT