Ask an AI image model for a picture of a specific real person and you get one of three outcomes. A clean refusal. A silent substitution, where the API returns HTTP 200 and an image of somebody who is not the person you named. Or the picture you actually wanted. Which one you get is almost entirely a function of which model received the prompt. Prompt engineering does not move it.
The two models this comes down to in practice are Google's Nano Banana (the Gemini image family) and xAI's Grok Imagine. Terminal X has to decide between them on every image request, and the decision is not a heuristic somebody wrote on a whiteboard. It is a literal block of text inside our parser's system prompt, backed by a server-side rewrite that fires even when the parser gets it wrong. Below is that rule, quoted from the source, with the credit costs on both sides of the branch.
A refusal is the good case. You see an error, you know something went wrong, and on Terminal X the task is marked errored and excluded from the credit deduction pass, so you are not billed. The bad case is the model that returns a perfectly valid image of a stranger. Our image runners treat this as its own class of problem: a 200 response with no usable image is almost always a refusal wearing a costume, and it gets classified as a safety decline rather than a success.
Nano Banana is the model where this matters most, because it is otherwise excellent and cheap. Nano Banana Pro is our default for product photography, food, packaging mock-ups, architecture, interiors and cinematic film-still work. Plain Nano Banana is the default for editing an image you attach, swapping a background, or keeping a character consistent across shots. None of that changes. What changes is the moment a named human enters the prompt.
Grok Imagine sits at the other end. It is not the highest-fidelity model in our lineup for still-life or typography, and we do not pretend otherwise. It is the one that will actually render the person.
Rule 10a in the Terminal X parser system prompt is titled NAMED REAL PERSON IMAGES (likeness-aware pipeline), and it opens by overriding the normal image routing table: "THIS RULE OVERRIDES THE DEFAULT IMAGE ROUTING IN RULE 5. CHECK IT FIRST FOR EVERY IMAGE PROMPT." The justification block is the part nobody else has published, so here it is word for word:
"WHY GROK NOT NANO-BANANA: Google updated Nano Banana's policy in January 2026 to hard-block named celebrities/politicians/athletes server-side. The pipeline would silently fail. Grok Imagine (xAI) has the most permissive real-person policy in our lineup AND ships with strong celebrity-face training data — community-confirmed best route for "image of <named real person>" prompts. Grok also accepts URL references in its prompt for grounding less-famous faces."
The hard part is not picking Grok. The hard part is detecting that a real person was named at all, because people do not write full names. Sports fans write last names. Music fans write mononyms. So the rule enumerates the forms: full name, first name only when famous enough to identify, last name only ("Brunson", "Mahomes", "Messi"), mononym ("LeBron", "Shaq", "Zlatan"), nickname or stage name ("The Rock", "Mr. Beast", "AOC"), title plus name, and any combination of those in one prompt. The shortcut it gives the parser: "if a sports fan, news reader, or pop-culture-aware person would instantly know who you mean, it counts."
It also tells the parser which way to be wrong. "The cost of a false positive is one extra sonar-pro call (~$0.001). The cost of a false negative is the user gets a generic AI face that looks nothing like the real person — and the whole product feels broken." That parenthetical is optimistic, and we will say so on our own page: our provider-cost reference table prices a Sonar Pro call at roughly $0.031, not $0.001. The asymmetry the rule is arguing for still holds by a wide margin, but the number in the prompt is stale.
The rule is equally explicit about what does not trigger it. Fictional characters with clearly fictional names. Pure style references like "in the style of Banksy". Generic unnamed people ("a businesswoman in a suit", "a happy family at dinner") which route to Flux or Nano Banana normally. And a photo of yourself that you attached, which goes straight to Nano Banana with your upload as the reference.
A permissive policy is necessary but not sufficient. A model that is willing to draw a named athlete can still draw a generic athlete, because its training data is thin or stale on that face. So the rule does not schedule one task. It schedules two, with a dependency between them.
Task one is 'sonar-pro', a research lookup, with no dependencies. Its brief is fixed in shape: find the highest-quality public reference photos for the named people, return only direct photo URLs one per line, no markdown and no commentary, preferring Wikipedia, Britannica, official team or league pages, official press kits, and recent verified news photos.
Task two is 'grok-imagine', and it depends on the first task's id, so it does not start until the URLs are in hand. Its brief is your original prompt verbatim, plus an instruction to use the reference URLs to match the likenesses and render the actual people accurately with faithful faces and identifying features. That grounding step is the whole reason a less-famous face comes out recognizable rather than approximate.
This is also why the two tasks are sequential where most Terminal X work runs in parallel. Dependency ordering is not free: the image cannot start until the lookup returns, so this route takes longer end to end than a single-model image job. We take that trade because an ungrounded likeness is not a cheaper result, it is a wasted one.
One credit on Terminal X equals $0.006 of API spend, a ratio calibrated in May 2026. Two internal tables feed the numbers below. The credit column is what the executor quotes before a job runs. The dollar column is our reference table of what each provider charges for a typical call, verified against published pricing in May 2026.
The real-person pipeline quotes 15 credits: 6 for the Sonar Pro reference lookup plus 9 for the Grok Imagine generation. That is $0.09 of your allowance against roughly $0.071 of reference provider cost, so our margin on this specific route is thin by design. On the $4.99 Basic plan's 400 credits, that works out to about 26 grounded real-person images a month.
The comparison people expect to matter is that Nano Banana would have cost 2 credits and Nano Banana Pro 5. Both are genuinely cheaper. Both would also have returned nothing usable for a named person, which makes the effective cost of that path infinite rather than low. The 15-credit number is a pre-flight quote, not the final charge. Where a provider reports what a call actually cost, the executor re-meters that task and the deduction follows the real figure — down when the work came in small, up when a prompt was much larger than the flat estimate assumed.
The parser is Sonnet reading a long prompt. It is good, it is not deterministic, and we do not ship routing safety that depends on an LLM remembering a rule. There are three independent guards behind it.
First, the parser emits a policy flag rather than just a model choice. When 'is_real_person_likeness' comes back true, the server walks every scheduled task and rewrites any model in a hardcoded forbidden set to 'grok-imagine' before a runner ever sees it. That set covers every Flux image variant, Stability and both SD 3.5 sizes, all three Ideogram variants, Nano Banana and Nano Banana Pro, and every Recraft variant. Each rewritten task is stamped with the model it came from, and the server logs how many tasks it had to rewrite, so we can see how often the parser slips.
Second, the model picker in the composer is gated. If you manually select Nano Banana and your prompt names a celebrity, the request is refused before anything runs, at a cost of zero credits, with this message: "Nano Banana silently produces garbage faces for named real people (it generates generic faces instead of refusing — your credits would burn on an unusable image). Switch the model picker to Grok Imagine (the only image model in the lineup that handles real people well) or change to Terminal X auto-routing."
Third, there is a runtime fallback. If a Flux task fails for any reason, including a moderation decline, it hands off to Grok Imagine mid-run and emits a model-switch event so the result row, the error tooltip and the brain-view animation all re-attribute to the model that actually did the work. You see the swap announced in the stream rather than discovering it later in a log.
Routing everything to Grok because it is permissive would be a mistake, and an expensive one at 9 credits against Nano Banana's 2. Our routing table sends product photography, food, still-life, packaging mock-ups, architecture and interiors to Nano Banana Pro. Movie posters and key art with a baked-in title or credit block go to GPT Image 2, because typography inside an image is a different competence. Logos, icons and anything needing genuine editable vector output go to Recraft, which is the only model in the lineup that emits real SVG. Editing an image you attached, swapping an object or a background, or holding a character consistent across a set goes to Nano Banana.
Grok Imagine wins one category decisively and gets routed for exactly that category, plus explicit requests for it and prompts with an X/Twitter feel.
Video of a named real person is refused upfront, not attempted. No current video model handles it reliably, and video is the expensive end of the catalogue — Pika is 36 credits, Seedance 70, Veo 90, Kling 245 — so rather than spend that on something uncanny, the parser returns a refusal with a suggestion: upload a generic stand-in image, or describe the person generically. That is a real product limitation and we would rather say it here than have you find it at checkout.
Named franchise characters are blocked everywhere, including on Grok. Steve verified this on 2026-06-28 with a Pokemon prompt where Flux refused and the Grok fallback refused too. Nintendo, Disney, Marvel and Sanrio enforcement risk means no commercial provider will run it. Terminal X handles this by rewriting the brief to an original character in that franchise's visual style, which is style transfer rather than infringement, and telling you it did so.
A permissive content policy is not a rights clearance. Terminal X routes the request and grounds the likeness; it does not check publicity rights and does not put a human reviewer between the model and your result. If you are putting a real person's face in something commercial, that judgment is yours.
Reference grounding depends on the public web. For a person with thin public photo coverage, Sonar Pro returns weaker references and the likeness degrades accordingly. And provider policies move: the January 2026 Nano Banana change is the proof. This rule is maintained by hand against what we observe in production, which is why the page carries a date.
Or skip the comparison shopping: Terminal X routes one prompt to the right model automatically — and runs several in parallel when a job needs more than one.
| Model | Credits / image | Reference API cost | Named real person |
|---|---|---|---|
| Grok Imagine (xAI) | 9 | ~$0.040 | Routed here — most permissive |
| Sonar Pro (reference lookup) | 6 | ~$0.031 | Step 1 of the pipeline |
| Nano Banana (gemini-3.1-flash-image) | 2 | ~$0.005 | Blocked server-side by Google |
| Nano Banana Pro (gemini-3-pro-image) | 5 | ~$0.020 | Blocked server-side by Google |
| Nano Banana Lite | 2 | ~$0.003 | Same Google policy — not in the rewrite set |
| Flux 2 Pro | 6 | ~$0.030 | Rewritten to Grok before it runs |
| Ideogram V3 | 6 | ~$0.030 | Rewritten to Grok before it runs |
| Recraft | 8 | ~$0.040 | Rewritten to Grok before it runs |
| Stability (SD 3.5 Large) | 13 | ~$0.065 | Rewritten to Grok before it runs |
| GPT Image 2 | 16 | ~$0.080 | Not rewritten, still not the pick |
Grok Imagine. Our parser's routing rule states it plainly: Grok Imagine has the most permissive real-person policy in the lineup and ships with strong celebrity-face training data. Google hard-blocks named celebrities, politicians and athletes on Nano Banana server-side, and Flux, Stability, Ideogram and Recraft either decline or return a generic face. Terminal X routes named-person prompts to Grok automatically, and rewrites the model server-side if the parser picks anything on the forbidden list.
It's a policy decision at Google, applied server-side, not a limitation of the model's quality. Nano Banana Pro is still our default for product photography, food, packaging, architecture and cinematic scenes, and plain Nano Banana is our default for editing an image you attach. The block only applies to identifiable named people.
15 credits: 6 for the Sonar Pro reference-photo lookup, 9 for the Grok Imagine generation. At 1 credit = $0.006 that's $0.09, or about 26 images on the 400-credit Basic plan at $4.99/mo. Our reference table puts the provider cost of that pair at roughly $0.071. The 15 credits is a pre-flight quote — where the provider reports what the call actually cost, the deduction is re-metered against that figure rather than left at the flat estimate.
No. If your prompt names a celebrity and you manually select Nano Banana, Nano Banana Pro, Flux, Stability, Ideogram or Recraft, the request is refused before anything runs and you're charged zero credits. The refusal tells you why and suggests Grok Imagine or auto-routing. Letting the call through would return a stranger's face and still spend your credits.
No, and it refuses upfront instead of trying. No current video model animates named real people reliably. Seedance and Veo refuse; Pika and Kling produce uncanny results. The parser returns a refusal with a workaround: upload a generic stand-in image, or describe the person generically rather than naming them.
That's your call, not ours. A permissive provider policy is not a publicity-rights clearance. Terminal X routes the prompt and grounds the likeness against real reference photos; it does not verify likeness rights and does not review generated images by hand. Separately, named franchise characters (Pokemon, Marvel, Disney) are blocked by every provider including Grok, so those prompts get rewritten to an original character in that franchise's visual style.
Or skip the comparison shopping: Terminal X routes one prompt to the right model automatically — and runs several in parallel when a job needs more than one.