My Basic AI Image Work Process For Bing Image Creator (same text in image description)
The following is the same text as contained in the image:
- Normally, I create an average of 125 prompts (4 images per prompt yielding 500 total images) in about 1 1/2 hrs. and then stop AI work for the day. The way to get great AI images is to make the best possible decisions for selecting those 125 prompts, and a few of those images will be much better than the rest and will likely be world class in quality. The daily limit for Bing Image Creator varies, but seems to be around 500 images per day as of now. Using the ‘resize’ option cuts sharply into this daily limit, based on how frequently it is used.
- I work fast enough that the time for coming up with a new prompt is several seconds at most, and there is usually no real delay at all. I copy (for pasting) the most recent prompt I came up with and click around the finished images to speed up the display of the newly completed images to save time. I am basically inputting prompts nonstop for about 1 1/2 hrs and creating a large volume of images.
- I create photo-realistic images by adding ‘1980s grainy color film photo’ to the end of the prompt. Example: ‘person in desert, 1980s grainy color film photo’. The importance of ‘1980s grainy color film photo’ can’t be overstated. There are other similar prompt endings that would create a similar effect, but just adding ‘photo’ to the end of a prompt will not create a similar result. The difference when it comes to images of people is profound, both in terms of image quality, and in terms of avoiding a bunch of faces mixed together incoherently into one. I should note that I have been simply using ‘photo’ at the end of my prompts for the majority of my recent surreal images with strawberries.
- I make decisions about trying a completely new prompt or building on the most recent one based on how successful the most recent prompt was.
- I pay very close attention to every single word in the prompt. This is very different than how most people seem to make AI image prompts. This allows for much more precise control, and for much less typing with each prompt. Minimizing typing is key when inputting a huge volume of prompts.
- This whole thing has to be treated like (non-staged) photography where you can go in with an idea, but what really matters is what is actually available and workable. If you go to the forest to photograph a certain animal but its not there, you need to work with what actually is there. Trying to force the AI model to do things it doesn’t want to do is something to always avoid.
- A major goal is to look for something predictable in the AI model. For example, I know that ‘face made out of X’ works well in a predictable way. ‘X’ could be birds, lobster, gold, saxophones or anything.
- I have no music playing or distractions of any kind while creating AI images. I need every fiber of my brain focused to do this work at a high level. With AI, the only limits are the limits of the imagination, and every bit of brainpower I have gets poured into creating the images.
- Quality over quantity is critical. Having more than several hundred images on a Flickr page means that most people can’t take the time to go through it all. This means that increased quantity is a bad thing at that point. I usually upload around 10% of the images I create. I only want to upload my best work. If you are not actually getting paid for this, then quality is all that should matter, not daily view counts or follower numbers. The reason people on youtube and instagram care about these things is because they only care about making money (usually not much money) from what they do, and not about the quality of what they output.
- I always try to learn something from every single prompt attempt. It is only possible to understand how an AI model really works through extensive experience using it. I only use DALL-E largely because I don’t want to divide up my time among different AI models – I need to master just one.
- It is often completely unpredictable how changing an element in a prompt will change the quality of the image. Changing the type of animal in an image for example can improve or ruin the realism of the image - and in completely unpredictable ways.
- I prioritize not getting banned from Bing Image Creator and work cautiously in this regard. Another image post will cover this topic in more detail. Avoid getting multiple flagged images in a row at all costs. Use a completely safe prompt like ‘cat’ or ‘forest’ following a flagged image to avoid getting multiple flagged images in a row. This is always priority #1 when using this AI tool.
- I backup my most recent output to an external hard drive as soon as I finish the batch of images.
- I restrict modifications of images to necessary cropping and using the clone tool to remove image errors. I need to operate in a very efficient way to handle the massive volume of images, meaning retouching beyond this is to be avoided. Also, I don’t like the idea of it not being clear how much of an image was AI generated vs. edited later by a human, and this is another reason for limited editing of the images.
- I always wait 24 hours to start curating the images so I am not doing it in a state of being over-tired or over-excited. This is essential to examine the huge volume of images in a sober-minded state. The huge volume of images I create in a short period of time is something the human brain can’t quite handle, when it comes to evaluating them. This is why I wait 24 hours before starting the curation process.
- The massive volume of images is a major challenge to curate. More of my time is spent curating the images than generating them in the first place. The idea is to only upload the very best images. I only care about quality, not quantity. Most images have good and bad things about them, which is different than most real photography, and makes it challenging to decide is worth uploading.
- Unlike with real photography, it is well worth creating hundreds of minor variants of a single image to ‘hit a home run’. This is an extremely important (and not very obvious) thing to understand. Taking 20-50 real photos of a static subject usually is enough to get all the value that the subject has to offer. Taking 500 real photos of a chair is unlikely to yield anything that taking 50 images did not. AI image generation is profoundly different. There will probably be a tiny number of amazing images out of that set of 500. The way to get those amazing images is to create a large volume of images as a base to work with and look through.
- I save my best prompts (based on how well they worked) in a text file. I also record AI images ideas there to use later.
My Basic AI Image Work Process For Bing Image Creator (same text in image description)
The following is the same text as contained in the image:
- Normally, I create an average of 125 prompts (4 images per prompt yielding 500 total images) in about 1 1/2 hrs. and then stop AI work for the day. The way to get great AI images is to make the best possible decisions for selecting those 125 prompts, and a few of those images will be much better than the rest and will likely be world class in quality. The daily limit for Bing Image Creator varies, but seems to be around 500 images per day as of now. Using the ‘resize’ option cuts sharply into this daily limit, based on how frequently it is used.
- I work fast enough that the time for coming up with a new prompt is several seconds at most, and there is usually no real delay at all. I copy (for pasting) the most recent prompt I came up with and click around the finished images to speed up the display of the newly completed images to save time. I am basically inputting prompts nonstop for about 1 1/2 hrs and creating a large volume of images.
- I create photo-realistic images by adding ‘1980s grainy color film photo’ to the end of the prompt. Example: ‘person in desert, 1980s grainy color film photo’. The importance of ‘1980s grainy color film photo’ can’t be overstated. There are other similar prompt endings that would create a similar effect, but just adding ‘photo’ to the end of a prompt will not create a similar result. The difference when it comes to images of people is profound, both in terms of image quality, and in terms of avoiding a bunch of faces mixed together incoherently into one. I should note that I have been simply using ‘photo’ at the end of my prompts for the majority of my recent surreal images with strawberries.
- I make decisions about trying a completely new prompt or building on the most recent one based on how successful the most recent prompt was.
- I pay very close attention to every single word in the prompt. This is very different than how most people seem to make AI image prompts. This allows for much more precise control, and for much less typing with each prompt. Minimizing typing is key when inputting a huge volume of prompts.
- This whole thing has to be treated like (non-staged) photography where you can go in with an idea, but what really matters is what is actually available and workable. If you go to the forest to photograph a certain animal but its not there, you need to work with what actually is there. Trying to force the AI model to do things it doesn’t want to do is something to always avoid.
- A major goal is to look for something predictable in the AI model. For example, I know that ‘face made out of X’ works well in a predictable way. ‘X’ could be birds, lobster, gold, saxophones or anything.
- I have no music playing or distractions of any kind while creating AI images. I need every fiber of my brain focused to do this work at a high level. With AI, the only limits are the limits of the imagination, and every bit of brainpower I have gets poured into creating the images.
- Quality over quantity is critical. Having more than several hundred images on a Flickr page means that most people can’t take the time to go through it all. This means that increased quantity is a bad thing at that point. I usually upload around 10% of the images I create. I only want to upload my best work. If you are not actually getting paid for this, then quality is all that should matter, not daily view counts or follower numbers. The reason people on youtube and instagram care about these things is because they only care about making money (usually not much money) from what they do, and not about the quality of what they output.
- I always try to learn something from every single prompt attempt. It is only possible to understand how an AI model really works through extensive experience using it. I only use DALL-E largely because I don’t want to divide up my time among different AI models – I need to master just one.
- It is often completely unpredictable how changing an element in a prompt will change the quality of the image. Changing the type of animal in an image for example can improve or ruin the realism of the image - and in completely unpredictable ways.
- I prioritize not getting banned from Bing Image Creator and work cautiously in this regard. Another image post will cover this topic in more detail. Avoid getting multiple flagged images in a row at all costs. Use a completely safe prompt like ‘cat’ or ‘forest’ following a flagged image to avoid getting multiple flagged images in a row. This is always priority #1 when using this AI tool.
- I backup my most recent output to an external hard drive as soon as I finish the batch of images.
- I restrict modifications of images to necessary cropping and using the clone tool to remove image errors. I need to operate in a very efficient way to handle the massive volume of images, meaning retouching beyond this is to be avoided. Also, I don’t like the idea of it not being clear how much of an image was AI generated vs. edited later by a human, and this is another reason for limited editing of the images.
- I always wait 24 hours to start curating the images so I am not doing it in a state of being over-tired or over-excited. This is essential to examine the huge volume of images in a sober-minded state. The huge volume of images I create in a short period of time is something the human brain can’t quite handle, when it comes to evaluating them. This is why I wait 24 hours before starting the curation process.
- The massive volume of images is a major challenge to curate. More of my time is spent curating the images than generating them in the first place. The idea is to only upload the very best images. I only care about quality, not quantity. Most images have good and bad things about them, which is different than most real photography, and makes it challenging to decide is worth uploading.
- Unlike with real photography, it is well worth creating hundreds of minor variants of a single image to ‘hit a home run’. This is an extremely important (and not very obvious) thing to understand. Taking 20-50 real photos of a static subject usually is enough to get all the value that the subject has to offer. Taking 500 real photos of a chair is unlikely to yield anything that taking 50 images did not. AI image generation is profoundly different. There will probably be a tiny number of amazing images out of that set of 500. The way to get those amazing images is to create a large volume of images as a base to work with and look through.
- I save my best prompts (based on how well they worked) in a text file. I also record AI images ideas there to use later.