{
  "$type": "site.standard.document",
  "bskyPostRef": {
    "cid": "bafyreiceys725bnapr5c44s6eq5hhcwq5juetfcmcy3yeihurvmg6hvudu",
    "uri": "at://did:plc:lk3jfj3zq4k4wxnk474axylu/app.bsky.feed.post/3mkgrth66tt72"
  },
  "path": "/t/developing-sprite-sheets-with-gpt-image-2/1379831#post_1",
  "publishedAt": "2026-04-26T23:28:18.000Z",
  "site": "https://community.openai.com",
  "tags": [
    "(click for more details)"
  ],
  "textContent": "This issue has come up many times, but there still isn’t a clear, reliable solution. I’d like to share what I’ve learned so far and hopefully gather ideas to improve the process together. I’m not a designer or game developer, so input from more experienced folks would be very helpful.\n\n## The core challenge\n\nGetting a good reference image is key. If you ask the model to generate a full sprite sheet, it often produces incomplete results with repeated poses. The model is optimized for visual quality, but it doesn’t fully understand body positioning with consistent spatial awareness.\n\nEven the latest model, `gpt-image-2`, can struggle with left/right limb distinction. For example, it may interpret “left leg” as the leg on the left side of the screen, rather than the character’s actual left leg:\n\n## Practical tips\n\n  * Expect to iterate on prompts. If the model consistently misinterprets left/right, it can be more effective to work with its bias rather than fight it.\n  * Hallucinations still happen. Retrying the same prompt 2–3 times often fixes the issue.\n  * Avoid generating a full sprite sheet all at once. Working frame by frame tends to produce better results.\n\n\n\n## What makes a good reference image?\n\n  * Low-resolution images or pixel art can make depth and limb order unclear.\n  * A 3D mannequin-style reference helps a lot because it provides clear joints, shadows, and structure.\n  * Numbered frames make it easier to target specific poses.\n\n\n\n## Processing a full sprite sheet\n\n  * Generate the full sheet first, then refine individual frames.\n  * Use Codex to split and process frames one by one. This reduces manual effort, though errors can still occur.\n  * The API is another option for automation.\n\n\n\n## Transferring a pose to a character\n\n  * Attach the pose image as Image 1, then the character as Image 2.\n  * Ask the model to apply the pose, or use the prompt below.\n\nPrompt for pose transfer (click for more details)\n\n  * If the model swaps limbs or makes mistakes, it’s often faster to rerun the same prompt instead of asking for fixes.\n\n\n\n# Putting things together\n\nAfter generating the individual frames, they may have a wider range of variations:\n\nApply a simple prompt like this to normalize them:\n\n\n    normalize the style, character consistency and size for this sprite sheet, keeping the all the poses intact\n\n\n## Additional notes\n\n  * You can create reference images using tools like PowerPoint or an image editor with existing sprites. I used the model to convert sprites into a normalized 3D mannequin.\n  * Different characters may require different pose sets.\n\n\n\nOf course, it’s not perfect and is still a work in progress. Have you done something similar? Please share what you are doing and any ideas you have for improving the process.",
  "title": "Developing sprite sheets with gpt-image-2"
}