返回
RCreddit.com
19
·17小时前·开发者社区 · RSS

Reproducible canvas-aligned low-level patterns in somerandomllm-generated images and their possible relation to iterative editing artifacts [D]

查看原文
OpenAI模型发布

热度趋势

趋势数据积累中

百分比基于当前可用热度信号,而非评论数或独立用户人数。

推荐理由

OpenAI 相关模型动态已经出现,适合跟踪能力变化、生态影响和后续可用性。

AI 摘要

一位用户在ChatGPT图像生成和编辑中发现重复出现的伪影,这可能也适用于其他模型。他们注意到生成图像中存在主导空间频率,峰值大约在2.45像素和5.57像素。这一发现表明,这些可重现的画布对齐低级模式可能与迭代编辑伪影之间存在关联。

I may have stumbled onto something interesting while trying to figure out a recurring artifact in ChatGPT image generation and editing (maybe applicable to other models as well?).

After several rounds of generative editing on portraits, I would sometimes get this faint cloudy / mottled texture in areas that should have stayed smooth — backgrounds, walls, skin, and other low-detail regions.

At first I wrote it off as normal denoising or regeneration noise. But the more I tested it, the less random it looked.

That made me wonder whether different parts of the image were being handled differently during editing — preserved in some areas, regenerated in others, perhaps based on some internal mask or segmentation step.

Instead of repairing the image in place, I shifted the entire image by a fixed amount before running the repair. I eventually settled on 20 px for testing.

If some hidden spatial pattern is tied to the output canvas, moving the image relative to that pattern should change how strongly it shows up on the subject or background.

- in one case, simply removing the final instruction to “shift back -20 px” improved the result dramatically

That was the first point where this stopped looking like ordinary random noise to me.

The face and body tended to remain more stable than the wall, which made me suspect that some regions were being preserved while others were being re-synthesized.

That still didn’t explain the artifact itself, but it could explain why the artifact builds up unevenly.

Pixel-wise, though, it wasn’t actually all zeroes. There were sparse non-zero pixels and tiny variations throughout the image.

- dominant spatial frequencies were very similar in both images, including peaks around 2.45 px and 5.57 px

The result was surprisingly striking: both revealed a very similar large-scale cloud-like structure.

The cross-correlation peaked at zero lag, meaning the structured pattern was already aligned at the same canvas coordinates across independent generations.

So whatever this low-level signal is, it doesn’t look purely random. At least part of it appears to be reproducible and locked to the canvas coordinates.

I’m not claiming that this proves OpenAI watermarking, SynthID, or any particular proprietary mechanism.

Generated images appear to contain a weak, reproducible, canvas-locked spatial pattern — even when the image looks completely black.

- After several passes, those signals may begin to reinforce or reveal themselves as visible mottling in smooth areas.

Repeated generative editing can accumulate or expose a weak structured signal that is fixed in output-image coordinates, eventually making it visible as cloudiness or mottling in otherwise smooth areas.

- Have you seen this kind of cloudy / mottled artifact after repeated AI image edits (I mean, come on, who doesn't)?

- Has anyone tested whether supposedly “black” images from other generators contain reproducible spatial structure?

- Does this look more like watermarking, dithering, decoder bias, quantization, or something else (go figure!)?

- Has anyone analyzed something similar in frequency space, after heavy blurring, or using phase shifts?

- If you’ve run into this before: what turned out to be the most reliable way to prevent it during iterative editing?