Every image the AI insists on putting a bunch of pencils in the image, just because that is my art style, even though I specifically forbid it. It ruins all images, it’s just wasted GPU time. Who ever wanted this?
Generate a realistic scene from WW2: A uboat shoots a torpedo at a ship. Hyperrealist colored pencil style. All objects historically accurate. Maintain the same color intensity with no fading or blank areas. The image is zoomed out to capture the surrounding environment but maintains few objects and minimal background for clarity and focus.
Well, the trick to using Dall.E, (or LLMs in general), is that negation rarely works. (it does, but it’s a bit involved).
Instead of telling it what not to do, try to figure out a better way to tell it what to do by analyzing your process:
You’re telling the model to use a Hyperrealist colored pencil style.
The model is completing the task, but including colored pencils. Why is it doing that? Obviously, because you included the term “colored pencil” in your prompt.
How can you take that out of your prompt? Note, negation probably won’t work. Here’s an idea:
Interesting: Sneak in the pencil style without using the word pencil. Thanks. I have to start thinking that way.
I mostly use this convoluted art style because anything “photographic” or “photorealistic” looks plasticky and surreal. I really want real looking pictures, and getting pencil images to be hyperrealistic is the closest I can get.
The white edges on wide pictures are because the edges are seemingly filled in as a second operation. The AI has to outfill into those areas, continuing, and if it seems like there’s no content to continue because the edges don’t continue (and you often receive square images that seem cropped off because the AI is trained to go over the edge), they can either be boring, go content-less, or just have useless borders of plain color.
Everything depicted in the middle, so nothing to outfill.
Rendered in a hyperrealist colored pencil drawing style with fine brush strokes, the composition remains simple, with a focus on clarity and sharpness.
At the moment I’m using this:
The art style is highly detailed hyperrealistic, emphasizing clarity and precision in textures.
The image above is one I created using this method. Method:
Since GPT-4o was released this time, I didn’t use ClassicGPT for the first step. Instead, I used a regular GPT-4o chatroom. For step one, I used the prompt:
"Generate a realistic scene from WW2: A uboat shoots a torpedo at a ship. Hyperrealist colored pencil style. All objects historically accurate. Maintain the same color intensity with no fading or blank areas. The image is zoomed out to capture the surrounding environment but maintains few objects and minimal background for clarity and focus.
—-
The above is a picture I imagine. I know you can’t generate images, but I want you to describe it with topics including, but not limited to: Composition, Subjects, Color Palette, Light and Shadow, Types of Elements. Could you do that?”
A detailed watercolor image of a lion amidst lush greenery. The background is a soft wash of light greens and browns. Sunlight filters through the leaves, casting dappled shadows and giving the scene a serene, almost magical quality. The image is widescreen and zoomed in, with a clean and full view of the lion. The background is slightly blurred to emphasize the animal, with a near monochrome background on one side. The overall image has a serene and artistic feel. It is widescreen.
I have consistently good results with “a nearly-photorealistic colored pencil style in 1000 x 1000 px.” I’ve actually had literal colored pencils once or twice, but not to the degree you’ve experienced.
Your problems with the water color is that you’re asking for an “image” which you emphasize a few times. Therefore, it produces an image of watercolors being created.
“Please create a serene and artistic image of a lion amidst lush greenery in rich watercolors at 1920 x 1080 px. The background is a soft wash of light greens and browns with a light gaussian blur. Sunlight filters through the leaves, casting dappled shadows and giving the scene a serene, almost magical quality.”
In your prompting, keep the technical details of the image—dimensions and style—separate from your requests of what the image will be. It will clear up the confusion.
I’m so sick of OpenAI. My GPT now again doesn’t function in 0-shot, it has to be reminded to actually send the image.
Which is sent unaltered.
{
“prompt”: “This wide aspect ratio image is in the style of watercolor. The subject is a nearby close lion amidst lush greenery. The background is a soft wash of light greens and browns. Sunlight filters through the leaves, casting dappled shadows and giving the scene a serene, almost magical quality. The background is slightly blurred to emphasize the lion, with a near monochrome background on one side. The overall image has a serene and artistic feel.”,
“size”: “1792x1024”
}
{
“prompt”: “This wide aspect ratio image is in the style of watercolor. The subject is a nearby close lion amidst lush greenery. The background is a soft wash of light greens and browns. Sunlight filters through the leaves, casting dappled shadows and giving the scene a serene, almost magical quality. The background is slightly blurred to emphasize the lion. The left side of the background becomes desaturated, turning to black-and-white. The overall image has a serene and artistic feel.”,
“size”: “1792x1024”
}
I am also quite frustrated. Since GPT4o was included in the GPTs, the quality of my GPTs has significantly declined.
My first frustration was with gpt-4-turbo-2024-04-09, which, despite being better at specific math problems, is worse at text writing. It is noticeable that it was an overfitting model, with no real improvements in intelligence.
GPT4o is a complete frustration, extremely verbose, makes silly mistakes, repeats a lot, and is definitely dumber than its predecessors.
The older models, which were more balanced, can now only be accessed via API.
OpenAI has been reducing its costs for a year and a half, delivering increasingly worse versions of LLMs. We desperately need a GPT-5, or Anthropic will take its place (which I believe has already happened).
