SoulGen Text-to-Image vs Fictional References: Which Starting Point Fits Your Character?
Current policy: SoulGen’s Terms §5.2(a) prohibit uploading, sharing or transmitting pornographic content. This comparison uses ordinary fictional character art; a search phrase is not permission. Read the provider’s current Terms.
Start with text when the character does not exist yet. Consider a reference only when you already have a suitable image and a clear reason to preserve its visual identity. A reference adds another input to manage; it does not prove that a series will stay consistent.
This comparison is based on public SoulGen pages checked on October 21, 2026. We did not upload an image or generate a result. If you arrived through a SoulGen AI Porn search, the examples here are ordinary, fully clothed fictional adults. The search phrase does not establish that a provider permits a particular use.
One brief, two different starting problems
Imagine a fictional 32-year-old cartographer named Mira. You need an editorial portrait showing a green field jacket, short silver hair, a red notebook, and warm light from a workshop window. There is no real person to reproduce. The first deliverable is a square profile image; a later project may need three related scenes.
In a text-first workflow, the design is still open. You are deciding what Mira looks like. In a reference-first workflow, an approved depiction already exists, and the question is what should remain recognizable. Confusing those stages can make an early draft feel compulsory before anyone has approved the character.
SoulGen’s About page lists text-to-image, and its public character screen exposes prompt, aspect-ratio, and image-count controls. The homepage also describes reference-based lookalikes. Those are feature descriptions, not our findings about fidelity.
| Decision | Text-first | Fictional-reference-first |
|---|---|---|
| Best starting material | A written character brief | An approved fictional image you may upload |
| Main design question | Which visual interpretation works? | Which traits must carry into the next depiction? |
| Input record to retain | Brief and prompt versions | Brief, source image, provenance and permission scope |
| First failure to inspect | Missing or conflicting requested details | Unwanted changes to identity, clothing or composition |
| When to stop | One design satisfies the brief | The new use preserves the required traits acceptably |
| Evidence still needed | Actual output at your selected settings | Actual similarity across your required scenes |
Build a brief that can be assessed
Write three separate lists before opening the generator:
- Fixed facts: adult age, hair, jacket, notebook, workshop setting.
- Flexible choices: exact chair, secondary objects, wall color.
- Delivery constraints: square crop, eyes visible at a small size, room around the head, no legible logos.
An original non-explicit prompt could be: “Editorial portrait of Mira, an original fictional 32-year-old cartographer wearing a green field jacket, short silver hair, holding a red notebook in a sunlit map workshop, warm natural light, clear facial detail, square composition.” This is a proposed brief, not a prompt we tested or a promise of exact compliance.
Treat the checklist as the decision-maker. An attractive output with the wrong notebook or an obstructed face is still a failed delivery. Conversely, a plain but readable portrait may satisfy a small profile-image requirement better than a complicated scene.
When a reference earns its extra work
A reference is useful when someone can name what must carry forward. “Keep the same hair silhouette and jacket colors while changing the setting” is assessable. “Make it more consistent” leaves the acceptance criteria undefined.
Use an original drawing, a commissioned image whose agreement permits this upload, or a generated fictional image you are entitled to reuse. Keep its creation record. Being able to view or download an image does not by itself supply permission to upload it to another service. An unknown gallery image is a weak foundation for a continuing character.
For the Mira example, keep a small reference sheet: the chosen portrait, a short trait list, and two details that may change. Compare later candidates against the sheet one trait at a time. If you need the face to match exactly for a production deliverable, do not assume the advertised lookalike feature meets that requirement without evaluating the actual result.
Budget attempts, not imaginary finished assets
The current public pricing table lists AI Character at one credit per image without Looks Like and two with it. A planning example of six candidates therefore starts at six versus twelve credits, before edits or other features. Confirm the actual selected mode before generating. These are listed input costs, not proof of six usable portraits.
Set a stop rule: after a small agreed batch, review the brief rather than continuing with the same ambiguous request. Record the chosen candidate and why the others failed. If all six miss the same requirement, clarify that requirement before spending again. If different candidates each solve half the problem, decide which base image needs the smallest justified edit.
Which route should you choose?
Choose text-first for early character exploration, a single independent portrait, or a project with no approved source image. Choose a fictional reference when a suitable original design exists and visual continuity has a defined role in the deliverable. If the only available reference has uncertain rights, return to the written brief.
For Mira’s first profile image, text-first is the sensible starting point because there is no approved identity to preserve. For a later workshop series, the selected fictional portrait may become a reference after its provenance and upload permissions are recorded. That recommendation comes from the project sequence, not an unperformed quality test.
Source and scope note: Public feature and rate pages linked above were checked October 21, 2026. Account availability, successful reference matching, free allowances, output resolution, and checkout prices were not tested. See the provider’s Terms before supplying any input.
Related decisions: Browse all five comparisons or use the homepage planner.