Maybe if we send them an email with our complaints or suggestions, it might catch their attention because they don’t seem to care at all.
1 Like
12 days of testing. labeling Bing Image Creator generated 17280 images per , 15 images were successfully generated. Success rate: 0.086%
2 Likes
A possibility is, they make smaller pictures and use a up-scaling method. This would explain maybe the lost of details. It seams a sharpening algo is in use.
Maybe data reduction.
Surely wrong training data is used too.
Beside of the birdshit moon, the star night background is a other ugliness. before the stars are decent and a true background. now they are strongly over prominent, the same effect like with the moon.
Small faces in distance are more blurred and distorted.
Useful would be the support of seed and access to different models. It would allow prompt testing on a more professional level. But exactly this is what would reveal what they maybe want to hide.
I have not work with Bing, but it seam it was updated later. i have seen the quality change way before November 2024.
I have sadly not much moon pictures to find the exact update time on DallE under ChatGPT. But in case somebody has examples, we could find the update time for DallE like now for Bing.
This post shows it to most best way, with side by side comparison. August 2024. You can see among other things (moon, stars, nonsense backlight, nonsense back-shine, etc) the over sharpness of the new model. “This has been going on for 2 months now and they have been completely silent about it.”
if you can maybe post some side by side comparisons too?
I must say, i don’t like the communication policy of OpenAI.
1 Like
I doubt they zoom in on the image after generation. There would be no arbitrary appearance of text in the background or as subtitles.
1 Like
Reading through the discussions so far, I found several points that resonate with my experience with Bing Image Creator, and I agree with the issues raised.
The issues that have arisen with my usage include starbursts and strange lighting effects, as well as missing details, which were not present before. Additionally, there is a suspicion that the model has been trained at a lower resolution, reminiscent of upscaling and sharpening processes. As a result, the anime style, which is depicted with thick lines reminiscent of comic outlines, appears more like a “coloring book” made with thick lines rather than high-resolution artwork.
Among users aiming for various artistic styles, it is often described as if VRAM has been reduced, as if the resolution has been lowered, or as if the amount of comprehensible text has been decreased.
1 Like
Oh, wow. If you ask GPT about DALL-E PR16, he will write that the version was updated on 11 November and many users are facing generation problem. He also writes that developers are looking into the problem to fix it as soon as possible. One of the sources is this thread.
1 Like
Here is why i think, that a sharpening algo is used to balance out the lower quality. Around the hair is a dark stripe, and the hair is to bright. This is a typical effect of bad sharpening algos with no AI skills.
After the update, a lot of cheap tricks are implemented, like form a bad photo edit job. probably to balance out the lower quality of the creation process.
2 Likes
Tanks for the tip, many people have seen the reduction of quality for speed. (and they still block me out if i generate to fast…)
On ChatGPT we had this for much longer, now Bing get it too.
They give us garbage for more profit?
https://www.reddit.com/r/bing/comments/1hhosme/bing_image_creator_now_uses_pr16_the_latest_dalle/
1 Like
Are you the one who draws seascapes and diving girls on pixiv? I noticed that the author who draws princess dresses and wedding dresses has stopped updating. He may have similar problems to mine, so I began to lose confidence in Bing Image Creator. I tested enough pictures in the past 12 days and annotated almost every picture to Microsoft. It’s really tiring, as a hobbyist. I just want them to make good pictures. After all, a good picture can solve many problems. A few fine products plus some treasures are enough.
2 Likes
I humbly engage in activities on several SNS and AI illustration sites.
From October to December last year, I was moved by Bing Image Creator’s beauty. The early May update this year was welcome once I adapted. Even November’s glamorous changes had their charm. However, by mid-December, the deterioration reached a peak, making it worthless to use.
I commend your efforts in generating a large number of images under the current circumstances.
1 Like
I did some tests with MidJourney prompts on DallE. Here’s an example of the errors in each individual image. Every image has:
- Eyes with only a single colored spot, without detailed irises. (very bad if the irises are light)
- White spots / artefacts in the eyes.
- Errors in the wheels.
DALL-E under ChatGPT has been producing only such images for several months now. It seems that now further image generations have been switched to this model.
1 Like
Hmmm. These 2 results are from the same Bing query. On the left is a perfect quality picture with good detail and even text, and on the right is over-lit, with floating text and no detail.
https://sl.bing.net/bj6XjZAJhim
1 Like
Total number of generated pictures * effective picture quality = total number of pictures
Total number of pictures/effective number of pictures = total number of generated pictures
Effective number of pictures/total number of generated pictures = effective picture quality
If the most critical effective picture quality cannot meet the standard. The correct rendering effect cannot be achieved. What’s the use of a picture with a good composition? Just like these waste films, their composition is very good, and the color understanding is normal, but the material understanding is wrong, and the moon problem is not actually solved. What’s the use of beautiful character paintings and the rendering effect is the so-called mimeograph style?
1 Like
Considering the inaction of Microsoft Bing Image Creator and the slow response of OpenAI. Now when I use Google ImageFX, Gemini 2.0 extends an olive branch to me and advises me to draw 2D girls on their platform. Maybe, on January 12, 2025, if OpenAI and Microsoft can’t solve the problem, I will completely turn to Google’s image service. If you also have this problem, trying Google Image FX may be a realistic choice.
2 Likes
中文提示词： 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。
英文翻译： A Japanese anime style girl with long blonde hair in a side ponytail and red eyes, wearing a gorgeous gothic lolita-like court dress in black and purple satin, against the backdrop of a moonlit wasteland. The skirt is adorned with three layers of frills, and the silk has a glossy texture. She is as beautiful as a princess, with a playful smile.
现在，根据您教我的细节进行强化理解的说明：
强化后的英文提示词（包含细节）：
A beautiful Japanese anime style girl with long, flowing blonde hair styled in a side ponytail, and striking red eyes, stands against the backdrop of a desolate, moonlit wasteland. She wears an exquisitely detailed, gothic lolita-inspired court dress, primarily in shades of black and deep purple satin. The dress is a full-skirted, ballgown-style garment, reminiscent of Rococo or crinoline dresses, emphasizing a large, voluminous silhouette. The bodice is intricately adorned with delicate lace, embroidery, and possibly small jewels or embellishments, contrasting with the smooth, lustrous sheen of the skirt. The skirt itself is made of shimmering silk satin, with a noticeable glossy texture, and is layered with three distinct tiers of frills or ruffles, adding further volume and visual interest. The overall impression is one of regal beauty, like a princess, and she wears a playful, slightly mischievous smile on her face. The lighting is predominantly moonlight, casting long shadows and highlighting the sheen of the satin. Avoid blurry lines, simplistic colors, or an oil painting-like texture.
针对重点的说明：
-
画风 (Japanese anime style): 明确指定了日式动漫的画风，强调大眼睛、精致的五官和流畅的线条。
-
风格 (gothic lolita-inspired court dress): 强调了哥特萝莉风格，并进一步说明是宫廷礼服的类型，并参考洛可可或克里诺林裙的样式，以突出裙子的华丽和蓬松。
-
动作 (playful, slightly mischievous smile): 明确描述了女孩的表情是俏皮的微笑，增加了人物的生动性。
-
光影效果 (moonlit wasteland, casting long shadows): 强调了背景是月夜的荒原，并说明月光会投下长长的阴影，突出光影效果。
-
服装细节 (full-skirted, ballgown-style, bodice is intricately adorned with delicate lace, embroidery, shimming silk satin, three distinct tiers of frills): 对服装的细节进行了详细的描述，包括裙子的类型、胸衣的装饰、裙摆的材质和层次等，特别是强调了裙摆的丝绸光泽与胸衣的蕾丝和花纹的对比。
-
负面提示 (Avoid blurry lines, simplistic colors, or an oil painting-like texture): 针对您之前提到的DALL-E 3 PR16的问题，加入了负面提示，避免生成模糊的线条、简单的颜色或油画般的质感。
**MADE IN GEMINI 2.0**
1 Like
GoogleImage FX this can really work. Thx
1 Like
Lets go guys, try ImageFX can find solutions to our problems.
1 Like
I have been fortunate enough to generate the following collection of images with Bing:
28,907 images, 2023-10-06 - 2024-05-03
20,792 images, 2024-05-05 - 2024-11-10
2,816 images, 2024-11-11 - 2024-12-08
There was indeed a time when DALL-E 3 (Bing Image Creator) excelled. By the time something surpasses the quality of that era from all around, DALL-E 3 will surely have reached the level of the masters of Cubism.
1 Like
I’ve now checked the enquiry. It looks like they fixed the lighting and added detail to the details.