I’ve been doing some searching & trying to figure out if there are any tips & tricks or best work practices when using the refined realistic model. Also, if anybody has any useful structured guidelines to repeatable inputs in this description box they could share because I don’t understand why or how to stop portrait images of people from looking like Chewbacca or a hairy Chia Pet, as well as the background and all the hair growing everywhere. I don’t understand how that’s realistic or refined whatsoever, but I can’t figure it out, and I don’t know if there’s any help on here.
I tried to do a search for a post, but I didn’t see one. I don’t see much documentation from Topaz Labs on prompting. I did a Gemini Chat search to see if there was any info on best work practices, negative and positive prompts. There wasn’t much.
So coming to the source, anybody can add some input on what works and what doesn’t work.
also, I see that Topaz has released a realistic API for enterprise. I’ve been using the Bloom realistic (that’s a hit-and-miss). Like going through a Whataburger drive-thru right after the bars close up and getting your order right on a Friday and Saturday night downtown (around the DFW area). When I upload an image and use the low settings in the Bloom realistic profile, it looks like the person in the image either survived the scene in Terminator when the nuclear explosion goes off or they just spent too much time in Florida sun, looking extra crispy and flaky. So, how do the customers get access to this realistic API? It seems to be a lot better from the product advertisements and other platforms are using your API. Or is there plans to update the Redefined Realistic model that’s been in Beta for over 6 months now?
Generally speaking, descriptive prompts as sentences rather than a list of keywords works best. Check out tips from other users in the thread below:
Here are some other tips I generated with Perplexity that should help:
Structure for sentence-style prompts
A good pattern is: subject → key attributes → style/medium → mood/lighting → extras.
You can keep this sentence-like but still structured:
“A young woman walking alone on a foggy forest path at sunrise, cinematic lighting, soft focus, 35mm photography, high detail.”
“An old stone lighthouse on a stormy coast at night, dramatic waves crashing, moody, cinematic, ultra-detailed digital painting.”
Techniques to make descriptive prompts work
Put the main subject in the first clause: “A red vintage sports car parked under a neon-lit bridge at night…” not “At night, under a neon-lit bridge, there is a…”
Use commas to separate concepts inside your “sentence” so tokens group cleanly (subject, setting, style, lighting, mood).
Avoid filler words (“there is”, “you can see”, “kind of”, “really”) because they waste tokens without steering the image.
Prefer concrete visual nouns and adjectives over abstract language (“foggy pine forest, wet ground, beams of sunlight” instead of “mysterious vibe”).
If you want more control, you can still mix in light syntax while keeping sentences:
“A portrait of an elderly sailor with a weathered face, ((dramatic lighting)), sharp focus, realistic skin texture, studio photography.”
Negative prompts and constraints in natural language
You can phrase negatives naturally but keep them short and explicit:
“no text, no watermark, no logo, no extra limbs, no distortion” as a separate negative prompt or appended clearly.
Even when using sentences, separating unwanted elements in a concise negative list improves cleanliness more than burying them in prose.
Something else to be conscience of: you’ll get different results rendering locally versus rendering in the cloud with image descriptions. Local renders are far more influenced by image descriptions than cloud renders are.
