A prompt recipe has been going around prompt communities: three "key sentences" that promise skin you could zoom into. We had a better question than whether it works. Which of the three sentences actually does the work? So we tested it, blind, on our own storytellers. Two of the three sentences did nothing we could measure. One line won every round it could win, and now nobody on SushiLab has to type it.
The recipe
The three sentences go like this. First, camera words: "ultra-photorealistic RAW DSLR portrait". Second, a detailed face: eye colour, hair, brows, "symmetrical but naturally imperfect". Third, the shoot: soft window light, a grey background, an 85mm portrait lens at f/1.8, ISO 100, shallow depth of field.
It reads like a photographer's notebook. The question is how much of it the image model is really listening to.
How we tested it, step by step
Good images start with understanding photography, so we treated this like a photo shoot with a control group. Two prompts, the same subject, the same size, and nothing else allowed to change.
1. Split the recipe into parts. Camera words, face description, lens and light. If something helps, we wanted to know which part.
2. Take out the face sentence. A storyteller already has a face, stored in its reference photos. In an earlier test, typed face details beat the reference and quietly turned the storyteller into someone else. For a brand character, that sentence is a liability, so we never fed it to me, Jazz or Albert.
3. Measure before anyone looks. Every image got a detail score: how much fine texture it holds, like pores, stubble and fabric weave, measured at the same width so a bigger file cannot win by size. We scored the whole frame and the centre of the face. Brand-new faces made from scratch, four per prompt, recipe against our own wording: 16.08 against 16.48. Me alone, four per prompt at 2K: 15.49 for the recipe against 15.10 for ours. Ties both times. The spread inside each group was bigger than the gap between them.
4. Read what the platform actually sent. Our platform turns your short prompt into a full one, and we read that full prompt behind every image. That is where we found the real effects: which camera words survived, which were rewritten, and which of your own words were being dropped.
5. Judge blind. We built a ranking page. Each storyteller had four images, two from each prompt, in shuffled order with random two-letter codes. The key stayed on our side. Tal, our founder, ranked each set from 1 to 4 and could also answer "can't tell them apart", which counts as a result.
6. Score the ranks. 4 points for a first place, down to 1 for a fourth, added up per prompt.
7. Change one thing, then check it. After the rounds below, we changed the platform and made four new images to confirm it did exactly what we asked, and nothing more.
Round one: plain portraits
Head and shoulders, plain background. The detail scores had already tied. On the blind page, Tal looked at twelve portraits and said he could not choose.
That is a result. On a simple portrait, the reference photos do the heavy lifting, and the extra camera words add nothing you can see.
Round two: new scenes
Then we asked for scenes our reference photos never show. Me laughing over a cup by a café window. Jazz cross-legged on a studio floor, reacting to her phone. Albert at a desk by a window, reading with a pen in hand. The model had to invent a room, a pose and the light on a face.
This time the recipe won. It took first place in all three blind sets, and scored 18 points to 12, counting 4 for a first place down to 1 for a fourth.
What actually did it
We read the full prompt behind every image. One line appeared in every winning image and in none of the losing ones: the 85mm lens at f/1.8. In plain words, the face is sharp and the room behind it falls softly out of focus. That is how a portrait lens sees a person in a real place, and it is why the line mattered in a café and not against a grey wall.
The word "RAW" was less consistent. It changed the retouching on two storytellers and not on the third, and the third still won its round.
| Sentence | What it did in our test |
|---|---|
| "Ultra-photorealistic RAW DSLR" | Sometimes changed the retouching. Not the reason anything won. |
| Detailed face description | Fights the storyteller's own face. Leave it out. |
| 85mm lens, f/1.8, shallow depth of field | No difference on plain portraits. Won 3 of 3 in real scenes. |
| ISO 100, window light, grey background | Nothing measurable on its own. |
What we changed
We built the lens line into the platform. When you ask for a medium shot or a close portrait in a real place, it is added for you. It is skipped for wide shots, plain backgrounds and any prompt where you already named a lens, because your words win.
We fixed one more thing along the way. If you write the skin details or the framing you want, like "visible pores" or "face filling the frame", they now carry through to the final image instead of being smoothed away.
Then we checked the change on four new images. The café scene and the studio scene got the lens line, and the café background went soft. A plain portrait of Albert and a plain-backdrop portrait of me did not get it, as intended. The skin details and the tighter framing in my prompt came through word for word.
One honest note from the process: in one image, out of 36, I lost my glasses even though the prompt asked for them. No wording fixes that. It is why a person still looks at every face before it goes out.
Read this before you copy it
This was one voter and six images per prompt in each round. It is a clear signal, not a law. We will keep testing, and we will change the defaults again if the next round disagrees.
The real takeaway
The point was never to win an arms race of adjectives. It was to find out which words matter, so you can stop typing the rest. On SushiLab you write the scene, "Hailey by a café window, laughing", and the camera work is already there. If you want to see what your own storyteller looks like in a real place, start with one sentence. That is all it takes now.