Prompt Writing Guide - Part 2: ChatGPT

Repeating what I wrote in Part 1, there's a ton prompt writing guides out there. This one won't be the best, but I hope it's not the worst.

Repeating my tenets from Part 1, along with an additional based on the topic of this guide:

-There's no right or wrong way to write a prompt. A run-on sentence or a bunch of keywords separated by commas, either of those can work well. A run-on sentence equalizes every keyword, which means you need to ((push)) things that you need more of. A prompt with commas will push the earlier keywords, so they need to be rearranged to your preference.

-Don't be afraid to increase or decrease the CFG if the prompt isn't working out.

-If you generated 100 images trying to make it perfect but hate all of them and need to stop, keep the 1 image you dislike the least, upscale and save it somewhere. There might be something about it that could help prompt writing later.

-On a similar note, learn from the prompt ideas you get from ChatGPT and any other tests you're doing. Save the really unique ones. Drag and drop those into a spreadsheet and paste the prompt next to them, so you can reference it later.

The format for this guide will be to ask ChatGPT.com to write five different prompts for me based on a few words. To add a bit of a challenge to this, I'll be using the original SDXL base checkpoint and an image dimension of 1344 x 768. There are far better checkpoints out and 1024 x 1024 is an easier ratio, but like I said, I want a challenge for this guide.

That prompt to ChatGPT:

(camera angle),

(camera lighting),

(the subject),

(subject's clothing),

(subject's action or pose),

(setting),

(3 additional relevant descriptive keywords),

Based on the above structure, create 5 different 7-line prompts of a scene, all words in lowercase, all lines ending with a comma, no bullet points before each line, and no annotations. Scene is: a couple out on a date

I added "out on a date" and "highly detailed photo" to the ideas it responded with.

That's an output of five different prompts, queued at the same, over 100 images made, in less than 15 minutes on my midrange AMD/Geforce rig. I was able to do this by using the Uni_PC Normal sampler at 10 steps and 4 CFG (thank you, Stable Diffusion group on Reddit).

I did this because I wanted to see a lot of examples of each of the five prompts quickly without having to use a Turbo or Lightning checkpoint, so that I can switch over to a 30 step sampler in regular if I find one that I like, since the results will be similar. The results in a fast checkpoint would not be similar.

A word about Turbo and Lightning checkpoints: I use SDXL Turbo a lot and have decent success with it. It's really good at img2img, since SDXL is already slower than 1.5 without img2img bogging it down even more. Also if you have a simple, specific prompt in mind and want to output a lot of options, Turbo is the way to go. It's super fast with really good contrast and detail. (Turbo seems to have no censorship either.) But if you want to make images with dynamic lighting, camera angles, and character expressions, SDXL Turbo will struggle and SDXL regular will shine. I have not had any success with SDXL Lightning, and every time I try, I get frustrated and go back to either Turbo or regular. To summarize, while the fast checkpoints can get you decent prompts and detail, getting the really cool images that regular can make is a challenge.

Here's three favorites that I kept from each of the five different prompts. They're not perfect, but they'd be a good starting point.

eye level,

soft evening glow,

a couple out on a date,

casual chic outfits,

walking hand in hand,

city park,

romantic, relaxed, happy,

highly detailed photo

old, ugly, angry, open mouth, blurry, grainy, low detail, low quality, worst quality, 2d, 3d, cartoon, illustration, painting, sketch

overhead,

warm candlelight,

a couple out on a date,

elegant dinner attire,

leaning in close,

intimate restaurant,

cozy, intimate, affectionate,

highly detailed photo

old, ugly, angry, open mouth, blurry, grainy, low detail, low quality, worst quality, 2d, 3d, cartoon, illustration, painting, sketch

medium shot,

twinkling string lights,

a couple out on a date,

casual summer wear,

laughing together,

outdoor café,

joyful, carefree, vibrant,

highly detailed photo

old, ugly, angry, open mouth, blurry, grainy, low detail, low quality, worst quality, 2d, 3d, cartoon, illustration, painting, sketch

close-up,

dim movie theater lights,

a couple out on a date,

casual jeans and tees,

sharing popcorn,

cinema,

excited, playful, cozy,

highly detailed photo

old, ugly, angry, open mouth, blurry, grainy, low detail, low quality, worst quality, 2d, 3d, cartoon, illustration, painting, sketch

wide angle,

golden hour sunlight,

a couple out on a date,

beachwear,

sitting on the sand,

seaside,

serene, loving, picturesque,

highly detailed photo

old, ugly, angry, open mouth, blurry, grainy, low detail, low quality, worst quality, 2d, 3d, cartoon, illustration, painting, sketch

You can get a pretty good final product out of Uni_PC, but once you've figured out a prompt, I recommend switching to a different sampler with higher steps (if you're like me and are never satisfied, ugh). It'll be slower but with better quality. To prove that Uni_PC can get you to the finish line, the bottom of this guide will have an upscale of my favorite image from this batch.

In order to take a prompt idea and move it into a better sampler, I'll use my CFG grid to test five different levels to find its sweet spot. I'm going to test the same seed number just for the hell of it, despite Uni_PC and DPM++ 2M Karras being much different samplers. This is an unreliable conversion, but the results can be close. It's unlikely that a really good image you made in Uni will ever look the exact same in DPM. Instead, consider rolling it into an img2img and tweaking it with a better Turbo checkpoint. I'm only doing this to show how I would move a prompt into a workflow where I can fine tune.