DALL-E 3 Generating Incorrect Colors and Details Since November 11, 2024

[kaienkala]

Hello,
I’ve noticed that DALL-E 3 has started generating unusual and incorrect images since around November 11, 2024. Before this date, the model worked perfectly. Now, it has issues like:

1. Colors are wrong (e.g., green or desaturated tones appear instead of natural ones).
2. Details are missing (e.g., textures or small elements vanish from images).
3. Strange artifacts (e.g., text-like symbols appear on the moon or other parts of the image).
4. Starbursts and odd lighting effects that were not present before.

For example, when I use a prompt like this:
“In an anime style: A beautiful girl with long blonde side ponytail and red eyes, wearing a gothic black and purple satin gown with layers of ruffles. She is sitting on a field under the moonlight, smiling playfully. The dress has a shiny silk texture with three layers of ruffles and purple rose decorations, making her look like a princess.”

Before November 11, the result was beautiful and matched the description perfectly. However, after November 11, the generated images have:

• Wrong colors (greenish tones).
• Artifacts on the moon (strange text-like patterns).
• Lost details (e.g., missing decorations, distorted elements).

Here is a comparison:

• Correct image from November 4: (Attach or link the correct image)
  
  

. Current Issues (After November 11):

Now, the images have the following problems:

• Wrong colors: The image appears greenish or desaturated instead of vibrant.
• Artifacts: The moon has strange text-like patterns on its surface.
• Missing details: Decorative elements like ruffles, roses, or textures are lost or distorted.
• Lighting issues: There are sharp starburst effects that look unnatural.

I’ve attached the correct image (from November 4) for reference. Since I cannot upload the faulty image, here’s a description:

• The generated girl now appears with faded or incorrect colors.
• The moon has unreadable symbols or random text artifacts.
• Decorations on the dress are missing or blurry, and the overall image quality has dropped.

This issue also happens in Bing Image Creator and when using DALL-E 3 through ChatGPT, so I believe it is a model-level problem. Could you please investigate and fix this issue? Many others seem to be experiencing similar problems.

Thank you!

[kaienkala]

“I’ve uploaded sample images to this OneDrive folder:

OneDrive [/f/s!AloLpD9-uBSJgghrP48u6pkQsmgK?e=JKsN5F] .

The folder is divided into Correct Samples (before Nov 11, 2024) and Error Samples (after Nov 11, 2024). All images are in their original state to preserve metadata for analysis.”

[_j]

@kaienkala

I’ve uploaded sample images

I’ve utilized that as https://onedrive.live.com/f/s!AloLpD9-uBSJgghrP48u6pkQsmgK?e=JKsN5F - however, I get “This item might not exist or is no longer available”.

I intended to retrieve your “bad” images and help you share.

If you browse and explore forum topics for ten minutes, you should receive a forum trust upgrade to where you can post more info yourself.

If on the API, you can use the style parameter to get a distinctly-different result, almost a different DALL-E 3. Send “natural” instead of the default “vivid”.

[kaienkala]

Apologies for the confusion earlier. It seems there was a mistake with the link prefix. The correct one should be

‘*ttps://1drv.ms//f/s!AloLpD9-uBSJgghrP48u6pkQsmgK?e=JKsN5F’.

I apologize for any inconvenience caused. Regarding the forum trust upgrade, I appreciate your advice and will spend some time browsing and engaging in forum discussions to increase my trust level, so I can post more detailed information myself.

[_j]

URL is then https://1drv.ms/f/s!AloLpD9-uBSJgghrP48u6pkQsmgK?e=JKsN5

The “bad” image folder image appearing most similar to your “good”, perhaps from the same prompt.

It certainly seems to give the impression of additional or altered upscale passes of “make extremely detailed and contrast-y, regardless of consequence”.

[kaienkala]

"Thank you for your reply! You’re right, the phrase ‘additional or altered upscale passes of “make extremely detailed and contrast-y, regardless of consequence”’ describes the problem perfectly. The crucial point is that both the ‘good’ and ‘bad’ images in my OneDrive link were generated using the exact same prompt within Bing Image Creator. This eliminates the possibility of the prompt itself being the cause; the issue lies with how Bing Image Creator (which uses DALL-E) is processing the prompts.

Here’s my prompt (in English): A beautiful anime-style girl with long blonde side ponytail and red eyes, wearing a gothic lolita-like gorgeous court dress in black and purple satin tones, with a moonlit wasteland as the background. The outside of the skirt is decorated with three layers of ruffles, and the silk texture of the skirt is very glossy. As beautiful as a princess, she is smiling playfully.

(Here’s the original Chinese prompt too, just in case: 月夜的荒原为背景。美丽的日漫风格的黑色与紫色缎面为色调的哥特萝莉一样的华丽宫廷礼服的金发长发侧马尾辫与红色眼睛的日漫风格女孩，裙子的外面是有三层荷叶边结点缀的，裙子的丝绸质感很光泽。像是公主一样美丽，她正俏皮的的笑着。)

I noticed this problem started abruptly around November 11, 2024. Before this date, using this exact same prompt in Bing Image Creator produced excellent, detailed, and natural-looking images, like the ‘good’ example I shared. However, after that date, the images started exhibiting the over-sharpening and high contrast you mentioned, resulting in a strange appearance and loss of detail, as shown in the ‘bad’ example.

Because I’m using Bing Image Creator and the prompt is identical, this strongly suggests that something changed on Microsoft’s/OpenAI’s end around November 11th, affecting the image generation process. It could be a model update specifically within Bing Image Creator, a bug in their implementation, or other internal adjustments.

Here are the two comparison images again, to clearly illustrate the problem:

• **‘Good’ image (before November 11th)
  

  

• **‘Bad’ image (after November 11th)
  

  

I hope Microsoft/OpenAI investigates this issue within Bing Image Creator and resolves it promptly. Thanks again for your help!"

[_j]

If using Bing, the reported issue might originate with technology developed by OpenAI, but the specific deployment is that which Microsoft maintains.

An AI-produced analogy: “Reporting it to OpenAI is like calling a songwriter because a karaoke machine is out of tune.”

Yet the same features are seen in ChatGPT-produced DALL-E images, pointy stars, wacky moon.

All prompts are rewritten internally into English by AI language models, for improved understanding and quality. The instructions have remained pretty constant with OpenAI, but the AI language model may change. That gives something else platform-specific.

{“prompt”:“A moonlit wasteland as the background, featuring an anime-style girl with long blonde hair styled in a side ponytail and red eyes. She wears a gorgeous Gothic Lolita-style ball gown in black and purple satin tones. The dress is highly detailed with a silky sheen and features three layers of ruffled embellishments. The girl looks as beautiful as a princess, smiling playfully. The overall style is elegant and reminiscent of a luxurious, fantastical aesthetic.”}

Sending Chinese prompt to OpenAI API

OpenAI API: default style

OpenAI API: natural style

---

Comparison, another AI image creator’s service - a moon to satisfy an astronomer.

[kaienkala]

"Thanks for the analysis! My own testing confirms the issues you’ve described, and I’ve pinpointed a key date: November 11th. November 4th was the last time I saw normal results, and the image generation seemed unstable on November 11th itself. This strongly suggests that the issue is likely related to an update or adjustment made to DALL-E 3 in early November.

To provide even more context and evidence:

• Confirmed working state: I have concrete evidence that Bing Image Creator was functioning as expected on November 4th.
• Consistent issues after November 11th: The problems we’re discussing have been consistently present since November 11th until today, indicating a non-random or temporary issue.
• Direct feedback to Microsoft: I’ve already submitted feedback to Microsoft regarding this issue.
• Testing with ChatGPT and DALL-E 3: I’ve also tested generating images directly through ChatGPT using DALL-E 3 with the same prompts. This test was crucial in determining that the problem isn’t specific to Bing Image Creator’s implementation but rather lies within the core DALL-E 3 model.
• Attempt to recreate the style using ChatGPT: I even tried having ChatGPT recreate the style of the November 4th images by providing them as examples. Even this method failed to fully reproduce the previous style, further confirming a change within the model itself.

This strongly suggests a change within the core DALL-E 3 model around early November, and I hope this information helps in identifying and resolving the issue."

[kaienkala]

**‘Bad’ image

Thank you for the information about Bing Image Builder updates. I also use the name Anakin. AI’s third-party platform (*ttps://app.anakin.ai/artist), ) tested image generation, where I encountered similarly bizarre artifacts (such as green-gray tones, text on the moon, and cross-shaped sparkles). This suggests that the problem may be with calls to the underlying model itself, rather than being specific to Bing Image Creator or ChatGPT implementations.

Regarding the update of DALL-E 3 PR16, this update may introduce new errors or biases that cause these visual exceptions. The integration of PR16 into Bing Image Creator may also cause problems. However, there may also be problems before the update, or due to other factors.

prompt:A moonlit wasteland as the background, featuring an anime-style girl with long blonde hair styled in a side ponytail and red eyes. She wears a gorgeous Gothic Lolita-style ball gown in black and purple satin tones. The dress is highly detailed with a silky sheen and features three layers of ruffled embellishments. The girl looks as beautiful as a princess, smiling playfully. The overall style is elegant and reminiscent of a luxurious, fantastical aesthetic.

[_j]

Just for fun, while we are trying other image generators, I solved your unrealistic moon problem and loli kawaikochan problem.

[kaienkala]

Thank you for your reply. I’m using Google’s image generator now. Hopefully this problem will be noticed by Microsoft and OpenAI and solved soon. I’m still using Poe’s anime style generator, but it’s too few times. So I really hope that DALL E-3 will recover quickly.

Your picture is really pretty. I don’t have the ability to repeat it. Because my current tools can’t produce such beautiful pictures.

The original DALL E-3 Star Trek-inspired Japanese anime style image. This is also a photo on my X page.

Drawing at 15:36 on 18 May 2024.

Drawing at 21:48 on 15 June 2024.

[kaienkala]

I’m following up on my previous post regarding issues with DALL-E 3. These problems with rendering and detail processing appear to be cross-platform, affecting both Bing Image Creator and ChatGPT’s DALL-E 3 integration. This strongly indicates a problem within DALL-E 3’s core model or related components, not specific platform implementations.

To reiterate the main issues:

• Abnormal Colors and Lighting (e.g., “Underworld Filter”): Unnatural colors, lighting, and contrast create a strange, “eerie” effect.
• Material and Texture Errors: Textures are often rendered incorrectly, appearing simplified or with the wrong material properties (e.g., silk looking like plastic).
• Facial Feature and Detail Errors (Especially in Anime/2D Styles): Unwanted facial details appear, making characters look distorted.
• Model “Hallucinations”: The model misinterprets elements, generating random symbols or unintended details.

The key evidence for this being a core DALL-E 3 problem:

• Cross-Platform Consistency: The exact same rendering errors occur across both Bing and ChatGPT.
• Regression in Functionality: DALL-E 3 previously worked correctly. These are recent issues, likely introduced by a change or update.
• Prompt Rewriting/Translation: Prompts are internally rewritten into English. Problems with these language models could lead to misinterpretations.

I suspect these possible causes:

1. Issues with the Translation/Encoding Process (of prompts):

   • NLP Model Updates/Bugs
   • Text Encoder Issues
   • Cross-Lingual Processing Issues (especially relevant for my testing with Chinese prompts)

2. DALL-E 3 Model Updates Themselves:

   • Improper Parameter Adjustments
   • Training Data Changes
   • New Bug Introduction

As detailed in my original post, I’ve tested with various prompts (English and Chinese), provided detailed error descriptions, and included comparison images. I’ve also reported this to Microsoft.

This severely impacts DALL-E 3’s quality and usability. This cross-platform nature of the issue is critical. I hope this helps in identifying the root cause.

[kaienkala]

The Chinese prompt word was successful once today, is it luck? Only twice, very accidentally. Other images are still rendering errors. My cue words are:

1. List item
   一个动漫风格的女孩，长长的金发，梳着侧马尾辫，眼睛炯炯有神。 她穿着一件豪华的舞会礼服，由光滑的黑丝胸衣和闪闪发光的紫色缎子裙摆组成。 连衣裙以三层褶皱裙为特色，闪耀着缎子的光泽，突出了洛可可风格的优雅时尚。 女孩像公主一样自信地微笑，而背景是一片简单，灯光昏暗的月色荒原。 强调她的衣服和头发中的超现实，错综复杂的细节。

2. List item

月夜的荒原为背景。美丽的日漫风格的粉红色缎面婚纱，粉红色调的哥特萝莉一样的华丽洛可可风格宫廷礼服的金发长发侧马尾辫与红色眼睛的日漫风格女孩，裙子的外面是有三层荷叶边结点缀的，裙子的丝绸质感很光泽。像是公主一样美丽，她正俏皮的的笑着。

3.List item

***一个动漫风格的女孩。***月夜的荒原为背景。美丽的日漫风格的黑色与紫色缎面为色调的哥特萝莉一样的华丽宫廷礼服的金发长发侧马尾辫与红色眼睛的日漫风格女孩，裙子的外面是有三层荷叶边结点缀的，裙子的丝绸质感很光泽。像是公主一样美丽，她正俏皮的的笑着。

It’s a coincidence that two correct images are generated, with a success rate of less than 2%, is it lucky?Compared with the English prompt, the Chinese prompt has a slightly better effect. But it’s not much different in nature. DALL E-3’s rendering logic and NPL cues seem to have changed. A lot of things don’t work normally. “Vocabulary bombing” is a very accidental way to get a 1% or 2% success sample. This is very rare, and these are the only images of my experimental success this afternoon.