Maxed Out
patreon
🧪 A Minimalist Prompt Test with Surprising Results
🕑 Added 2025-06-01 22:13:26 +0000 UTCThis one’s been on my mind for a while. I finally did an official test and the results are worth sharing. I'm keeping this post exclusive to Power Users, because it’s a real peek under the hood. 🫡
A lot of prompts I see floating around have things like (best quality:1.5), (masterpiece:1.3), (highly detailed eyes:1.8), and big negative prompts full of anatomical fails like "bad hands", "disfigured", or "extra fingers". I suspect these come from SD1.5 days.
But with SDXL? I think less can do just as much. And maybe even more.
🔍 The Test: Same Model, Two Prompt Styles
Someone recently posted a great image on my Hyper3D Civitai page. The image is good but when I looked at the prompt, I couldn’t help but wonder: what happens if I strip this down to just the essentials?
Here’s what they used:
Prompt:
(beautiful girl:1.4), beautiful eyes, best quality eyes, small eyes, (highly detailed eyes:1.2), (high quality:1.5), (best quality:1.5), 8k, 1girl, solo, looking at viewer, blue eyes, black hair, hair ornament, bare shoulders, (jewelry:1.2), upper body, flower, earrings, parted lips, hair flowers, blurry, black dress, from side, lips, (eyelashes:1.1), (makeup:1.1), blurry dark background, (feathers:1.9), nose, red lips, feather hair ornament, (mascara:1.1)
Negative:
(low quality:1.4), text, signature, sketch, (watermark:1.4),bad quality, worst quality, worst detail, censored, watermark, artist name, bad anatomy, extra limbs, extra fingers, missing limbs, missing fingers, distorted eyes, bad eyes, asymmetrical eyes, uneven eyes, poorly drawn eyes, extra eyes, bad nails, extra nails, deformed nails, (worst quality:1.5), blurry, distorted face, deformed facial features, asymmetrical face, uneven face, poorly drawn face, extra face, malformed hands, deformed hands, extra hands, missing hands, bad proportions, unnatural poses, artifacts, overexposed, underexposed, unrealistic textures, mutated body parts, disfigured,
Here's the image:
Now, this isn’t me calling them out at all. The image was good. But the prompt made me curious. What if I shortened it drastically? Would the image look worse?
So I tried. Same model, same settings, but different seed (Civitai didn't show it), and just rewrote the prompt to be leaner while trying to keep what actually describes the image:
Prompt:
beautiful girl, beautiful eyes, small eyes, detailed eyes, 1girl, solo, looking at viewer, blue eyes, black hair, bare shoulders, jewelry, upper body, flower, earrings, parted lips, hair flowers, black dress, from side, lips, eyelashes, makeup, blurry dark background, feathers, red lips, feather hair ornament, mascara
Negative:
text, signature, watermark
No weights. No “best quality”. Just directly describing the image. Notice I did not use "score_9, score_8_up...". I recommend it for Hyper3D but the control didn't use it so for the sake of testing, I didn't either.
And here’s what I got on the very first try (no cherry-picking):
💭 My Takeaway
It’s different which makes sense. Changing the prompt that much guarantees a different output. The goal is to compare quality and prompt adherence.
One thing that stood out, besides quality, was that the phrase "blurry dark background" from the prompt actually showed up in the stripped-down version.
To me, the image quality is so close between both versions that I'm not convinced to use many weights and quality tags for the sake of simplicity. I’m not claiming to be a master prompter so if you have thoughts or advice, feel free to drop them in the comments, not just for me but for other Power Users as well.
💁 To Clarify:
Weights just increase how strongly the model pays attention to a tag. Personally, I only use them when something keeps getting ignored (e.g., if I say “freckles” and they repeatedly don't show up). In ComfyUI, you can highlight a word in your prompt and press Ctrl + ↑ or Ctrl + ↓ to adjust the weight easily.
Also, words like "detailed" can help like "detailed eyes". Although with Hyper3D it tends to give the eyes a noticeable glow. Tags like "extra fingers" in the negative prompt can help if you keep seeing them (like when inpainting hands). Only put in the negative prompt to remove what you are seeing.
As for score tags like "score_9, score_8_up...". Hyper3D is a merge of models that all recommend that. It works because Pony diffusion (an SDXL fine-tune) was trained on these tags to drive quality. I have tested this a lot. If you don't use them, not the end of the world. If you do, they either will make the image better or have little effect. So I just stick with them.
Lastly, I was gonna do a deeper dive but that would have delayed this post for a few more days. If you would find a full blown prompting guide very useful, drop a comment.
⬇️ The Bottom Line:
My two cents:
When prompting SDXL, less is more.
Comments
Max
Ah I thought those were from sd1.5 days or something which I wasn't around for. To clarify, I do use scores on Hyper3D. I just didn't in this example because the control image didn't. So illustrious actually does benefit from prompting like that? This post might be a good topic to revisit when I swing back around for another SDxl workflow. To really lock in more about what works on some models and what doesn't.
Tigon
interesting, while i havent used your model and its interesting to see you are ignoring pony scores on a pony trained model, the user that posted on your Hyper3D page looks to be prompting Illustrious style as you often see `masterpiece, best quality, amazing quality, very aesthetic, absurdres, highly detailed, very awa` in the positive and `worst quality, low quality, bad quality` in the negatives on Illustrious models.
