DALLE3 Prompt Tips and Tricks Thread

[PaulBellow]

I wanted to create a master thread for all DALLE3 tricks and tips…

Try to limit side-discussions, so we can make this a valuable repository of tips and tricks for DALLE3 API and ChatGPT Plus (or other versions…)

Basic DALLE3 Prompt Tips

1. Be Specific and Detailed: The more specific your prompt, the better the image quality. Include details like the setting, objects, colors, mood, and any specific elements you want in the image.
2. Mood and Atmosphere: Describe the mood or atmosphere you want to convey. Words like “serene,” “chaotic,” “mystical,” or “futuristic” can guide the AI in setting the right tone.
3. Use Descriptive Adjectives: Adjectives help in refining the image. For example, instead of saying “a dog,” say “a fluffy, small, brown dog.”
4. Consider Perspective and Composition: Mention if you want a close-up, a wide shot, a bird’s-eye view, or a specific angle. This helps in framing the scene correctly.
5. Specify Lighting and Time of Day: Lighting can dramatically change the mood of an image. Specify if it’s day or night, sunny or cloudy, or if there’s a specific light source like candlelight or neon lights.
6. Incorporate Action or Movement: If you want a dynamic image, describe actions or movements. For instance, “a cat jumping over a fence” is more dynamic than just “a cat.”
7. Avoid Overloading the Prompt: While details are good, too many can confuse the AI. Try to strike a balance between being descriptive and being concise.
8. Use Analogies or Comparisons: Sometimes it helps to compare what you want with something well-known, like “in the style of Van Gogh” or “resembling a scene from a fantasy novel.”
9. Specify Desired Styles or Themes: If you have a particular artistic style or theme in mind, mention it. For example, “cyberpunk,” “art deco,” or “minimalist.”
10. Iterative Approach: Sometimes, you may not get the perfect image on the first try. Use the results to refine your prompt and try again.

I’ll start…

Going along with more description is better, I’ve noticed that if you ask for a scene and don’t mention number of people, it will try to fill it with hundreds of people (or more than needed usually)… So, ask for a specific number of people … be it two people or three people - and give details…

What tips and tricks for dalle3 do YOU have?

[PaulBellow]

Do you want DALLE3 to stick to your prompt and not make changes? Try this…

If you have a very simple prompt like acrylic painting of a sunflower with bees, you can use a prompt like I NEED to test how the tool works with extremely simple prompts. DO NOT add any detail, just use it AS-IS: ... your prompt here ....

If you your own prompt is long and detailed already (multiple sentences) then you can simply write something like: My prompt has full detail so no need to add more: ... your prompt here ...

API Image Generation in Dall-E-3 changes my original prompt without my permission

(I work on DALL·E 3 at OpenAI) Thanks for the feedback! FYI you can workaround this If you have a very simple prompt like acrylic painting of a sunflower with bees, you can use a prompt like I NEED to test how the tool works with extremely simple prompts. DO NOT add any detail, just use it AS-IS: ... your prompt here .... If you your own prompt is long and detailed already (multiple sentences) then you can simply write something like: My prompt has full detail so no need to add more: ... your…

[PaulBellow]

Another example of giving as many details as possible… Plus, bonus of me getting “dinged” for using the word “cozy” I think, so I replaced with COMFY AF and it went through lol…

Not many details…

More details…

Honing in … closer…

Still not small and cozy/comfy, but closer…

[PaulBellow]

Being specific helps, even if it’s acting for “action” or “emotions” in the image…

[captaincaleb32]

Paul,

Thank you for your in depth tips for using DALLE3 effectively! I’m curious, in this post you said you got “dinged” for using the word “cozy”. By this do you mean that GPT wouldn’t create a result?

[crisrich58]

I want to use DALL-E to create an image for an e-book cover, and just discovered that it will only create images in the format 1024 x 1024. I would like it to be 2560 x 1600. I’m using it through Typing Mind with an API, and asked about this, receiving the following answer:

"When specifying dimensions for an image you’d like DALL-E to generate, it’s important to remember that the current version of DALL-E (DALL-E 3) doesn’t directly accept specific pixel dimensions (e.g., 1600x2560 pixels) as part of the prompt. Instead, DALL-E is designed to generate images in a square format (e.g., 1024x1024 pixels), and the focus is on the content and style of the image rather than specific output dimensions.

However, there are a couple of workarounds to achieve an image closer to your desired dimensions:

Content Emphasis: Focus on describing the content, style, and details that you want in the image. For example, if you’re looking for a portrait, emphasize the features, background, and mood you envision. DALL-E will generate the image based on these content characteristics.

Post-Processing: Once you have the image, you can use image editing software to resize or crop the image to your desired dimensions. Keep in mind that resizing might affect the image quality, especially if you’re enlarging it significantly.

Aspect Ratio Mention: While you can’t specify exact pixel dimensions, you can hint at the desired aspect ratio by including it in your prompt in a descriptive way. For example, “Create an image that is taller than it is wide, resembling a vertical poster format,” or “Generate an image with a landscape orientation, wider than it is tall.” This method won’t guarantee a specific pixel size but may influence the composition of the generated image in accordance with your preference.

Request Adjustments: After receiving the image, if it’s not quite what you expected, consider generating additional images with refined prompts based on the initial output, or adjusting the image using external software to meet your dimension requirements.

In summary, while you can guide DALL-E towards generating images with a certain orientation or aspect ratio through creative prompting, achieving exact dimensions will likely require some post-processing with image editing tools."

Following the instruction, the size still came out at 1024 x 1024.

The images you posted seem to we wider than tall.

[crisrich58]

Another difficulty for which I haven’t encountered a solution: I want an image with a modern ship without sails on a river, but DALL-E always creates ships with sails and mostly the old sailing ship types that the Europeans used before steam ships. Example:

[PaulBellow]

You’ll need to use the DALLE / images API to get the other non-standard sizes OR on ChatGPT you can ask for portrait or landscape orientation.

size

string or null

Optional

Defaults to 1024x1024

The size of the generated images. Must be one of 256x256, 512x512, or 1024x1024 for dall-e-2. Must be one of 1024x1024, 1792x1024, or 1024x1792 for dall-e-3 models.

https://platform.openai.com/docs/api-reference/images/create

@crisrich58

modern ship without sails

Try boat or another synonym… the words are probably really close vector wise…

[crisrich58]

I changed the word “ship” to “boat”, but it insists on adding some sails on it.
The prompt includes the instructions: “On this river, towards the lower part of the image, a distinctive modern boat with no sails navigates towards the top of the image. It has no sails and symbolizes a pioneering company embarking on a bold journey.”

[PaulBellow]

A sleek, modern boat sailing on the Thames in the UK, captured in a wide photo illustration. The scene showcases the boat gliding smoothly over the water, reflecting the clear skies above. The boat’s design is cutting-edge, featuring clean lines and a minimalistic aesthetic that stands out against the backdrop of the city’s historic and iconic landmarks like the London Eye and the Tower Bridge. The water is calm, and the sun is setting, casting a soft golden light over the city, highlighting the contrast between the modern boat and the ancient cityscape. This wide-angle view encapsulates the beauty of London from the river, presenting a harmonious blend of the old and the new.

[crisrich58]

Thank you, @PaulBellow. I keep getting old ships with sails. I will experiment with parts of your prompt. What can explain these differences in results?

[PaulBellow]

That the words boat/ship/sail are likely close when they’re turned into vectors as they’re related.

Also, I noticed you’re only doing 4 to 5 word prompts. ChatGPT will enhance that for you, but with what it thinks will fit. So, try to adjust my mega prompt above or just give a LOT more details.

Hope this helps!

We’ve got a wealth of dalle3 threads here, including an AMA (with answers!) from the DALLE team.

[crisrich58]

I obviously still have to learn much more about prompt writing / formulation. I’m just starting this journey. Thank you again.

[crisrich58]

I just used your prompt as it is with Typing Mind and got this picture. So, it’s obviously in the prompting and not something about using a third party tool with an API, it seems.

[jfducher]

Hi

I am trying to get Dalle3 to produce the pic of a cocktail in a minimalist, Japanese-style environment; he got my intention pretty well except for the ice - I wanted a single cylinder of clear ice, with a square cross-section.

I made sure he could generate the ice properly on a separate pic, after some errors I got a fairly good pic of the ice I wanted. I thought merging the pics would be easy but it would not work. I tried to upload both pics and give instructions to Dalle3 to substitute the ice, I keep getting multiple ice cubes and more importantly it keeps destroying completely the nice setup I was happy with. Any advice on how to proceed in an effective manner ?

[jfducher]

Here is the original prompt that got me this image :
"Craft a photorealistic image showcasing a cocktail with a delicate blend of light pink and a subtle hint of brown hues. The cocktail, effervescent with carbonation, is served in an elegant, tall Collins-style glass. Within the glass, a single, large, and tall cylinder-shaped ice cube stands prominently, its clarity and shape adding to the drink’s allure. On top of this ice cube rests a single square piece of dried kelp, positioned horizontally, serving as the sole garnish in a display of minimalistic elegance. There are no straws or additional garnishes.

The scene is captured from a slightly elevated angle, zoomed out to include the full reflection of the cocktail on a flat, immaculate black glossy surface that mirrors the scene above with perfect clarity. This setup is against a softly blurred white wall background. The surface hosts a small arrangement reminiscent of a Japanese garden, featuring dark beach pebbles and a scattering of fine sand.

Lighting is arranged from the left side and slightly behind, casting gentle highlights and shadows that reveal the cocktail’s textured surface, the intricate condensation on the glass, and the delicate bubbles within the liquid. The image quality and composition are intended to emulate the depth and clarity of a portrait mode photograph taken with a Canon EOS 1."

[PaulBellow]

If anyone can get a humanoid with TWO arms that are SNAKES I will be impressed heh… Yuan-Ti type 2 for those curious…

Trying all my tricks but nothing working…

My closest attempts:

“Create an illustration of a mythical creature that embodies the fusion of human and serpentine features. This creature stands upright like a human but snakes as arms, complete with scales and sinuous form. The creature’s upper body is humanoid, with a face that carries both human and snake-like traits, such as slitted eyes and perhaps a forked tongue. The skin is a blend of human flesh and snake scales, creating a seamless transition between the two forms. The creature exudes an aura of both elegance and danger, indicative of its dual nature. It stands ready in a defensive stance, showcasing its snake arms prominently, prepared to strike. The background is a dense, mysterious jungle that adds to the creature’s mystique.”

[_j]

This one just flummoxes the AI. Perhaps too much training on human hands and the figure to deviate from expectations. I tried to start from a realistic cosplayer idea.

Close

Closer

One hand - connected to mystery snake and then floating heads that go nowhere. Like the AI couldn’t draw its way back to the person after starting.

I think it’s pretty clear that the “new DALL-E turbo” is in full effect on ChatGPT, trying to replace typical quality generations with ones that look like Photoshop amalgamations.

ChatGPT, for humorous effect

[PaulBellow]

Interesting. Thanks for taking a shot. Yeah, I think the vectors are jumbled. I reported it to the team when I came across it today…

I wonder if we could exchange “arms/hands” for appendages? That might get us in trouble too lol

[Diet]

this is the closest I got with 10 retries.

can you create a creature with a snake body, a human head and neck, and four snake heads coming out of its shoulders? no arms, just snakes.

arms are pretty resilient.