The Art of Quality Tags in Image Prompting
When creating AI art, it’s tempting to overload prompts with every imaginable quality tag. But blindly stuffing prompts causes more harm than good. In this post, we’ll explore nuanced tactics for quality tags that truly enhance your images.
The Pitfalls of Tag Stuffing
Phrases like “masterpiece” and “best quality” seem beneficial, but often just alter the output unpredictably. Tags like “4k HD” or “cinematic” are frequently copied prompts rather than technical directives for AIs. Without proper context, the AI has no inherent understanding of these vague subjective terms.
Prompt: A lone humpback whale swims through crystalline turquoise waters under the midnight sun. The majestic whale's skin shows scars telling a story of survival and resilience. Its eye reflects the glimmering northern lights starting to dance across the sky. Every detail of the scene captured in stunning photorealism - from the whale's barnacle-encrusted skin to the subtle ripples in the icy water. Despite the challenging lonely environment, the whale radiates stoic grace, power and ancient wisdom. This stock image conveys themes of strength, beauty, and our connection to nature. It inspires us to have faith during difficult times and forge ahead with quiet courage.
Created with: Stable Diffusion XL (SDXL)
Seed: 459108
Guidance Scale: 7.5
Dimensions: 1024x1024
Steps: 40
Sampler: Euler Ancestral
CLIP Skip: 2
Now, the same prompt with "masterpiece" and then "4k" then "8k"
As you can see the tag "Masterpiece" didn't add any discernible quality to the image. The file size did change from 1.25mb to 1.27mb.
This is the 4k image where the file size dropped from 1.27mb down to 1.20mb This would suggest that the tag removed something from the image.
This is the 8k image where the there is no discernable quality increase just some changes to the image composition. The size of the file also went down from 1.25mb to 1.22mb.
More meaningful triggers are specialized embeddings like “easynegative” or “ng_deepnegative_v1_75t”. These can refine images, but only if you have the exact models they’re designed for. Otherwise, they simply add noise and randomness.
Rather than piling on every tag found online, focus prompts on language used in real image descriptions and captions. That organic contextual data is what AIs comprehend best. After that foundation, sprinkle in targeted triggers suited for your models.
The Principles of Concise Prompting
With quality tags, aim for 1-2 well-chosen phrases that set the tone like “intricate details” or “soft lighting”. Resist cramming in a laundry list of generic terms hoping one will work. Place your most important tag near the start to guide the image.
Avoid assumptions that more tags equals higher quality. Lengthy prompts often dilute the AI’s focus. Hunter S. Thompson said, “Anything longer than 3 sentences gets boring.” AIs too function best with crisp, tight prompting.
Tailor your tags to the specific subject matter and desired style. For example, a sweeping landscape needs cues like “epic sense of scale” and “atmospheric light”, while a textured still life may use “rich color palette” and “brilliant details in the foreground”. Standard Diffusion XL can understand sentence structure better than previous models so if you struggle to find the different tags your after, describe it as best you can an SDXL will do a better job of understanding you.
Assess the Image First
No amount of tagging replaces directly evaluating the AI output and refining based on that. Look at factors like lighting, resolution, color tone, textures and tweak subsequent prompts accordingly.
True quality comes from choosing a few targeted tags, fitting them to the subject and style, and reading the AI results. Not blindly stuffing prompts hoping to magically increase fidelity.
Activate Your AI’s Unique Strengths
Learn which embeddings and model-specific triggers you have access to, and focus prompts on those. Rather than pasting prompts from MidJourney or Dall-E 2 suited for those AIs, use tags designed to get the most out of your environment.
Understand the materials your AI was trained on to pick fitting vocabulary. For example, if your model learned from artwork captions, tags like “impressionist lighting” or “Baroque composition” may resonate. But likely not camera technical terms, such as the difference between DSLR and SLR or f1.6 and f22. I have even seen ISO tags used in prompts and conducted my own research to see if any difference was made and found little to no difference in the image at all.
Know Your Model's Artistic Knowledge
When using artist names in prompts, understand that AIs won't inherently recognize every artist, especially less mainstream ones. Models trained on image-caption datasets may only "know" artists that were explicitly tagged frequently. Unless your AI was specifically trained to identify niche artists through labeled data, dropping random names may not guide it effectively.
For example, a model trained on general internet images likely recognizes "Picasso" or "Van Gogh" if they appeared in associated texts. But artists like Wendy Sullivan Green, Deven Rue or Edritch Blep would be unknown without focused training. Their names may alter the output unpredictably rather than in the desired stylistic direction. Get to know which artists your model will comprehend before relying on their names to define a style.
A list of known recognized artists can be found here :
https://stablediffusion.fr/artists
The Path to Prompt Mastery
Like any skill, excellent prompting comes through practice and analysis. Try prompts focused on different styles and subjects, and pay attention to which tags produce the desired effects.
Over time, you’ll learn the “secret sauce” tailored prompts that help your models shine. Mastering this advanced technique elevates your AI art to new levels.
Quality comes not from prompt stuffing, but through intentional concise tags activated for your specific environment. What tips have you discovered for prompts that really make images pop? Share your insights below!