Lumis in PixelForge · 1,000 prompts

PixelForge to Lumis prompt analysis

How real users phrase requests when bridging PixelForge and generative image workflows — classification, theme demand, sample dominance, and localization signals.

00Executive summary

Executive summary

Editing dominates; generation is a thin slice

  1. Background work (remove, replace, staging) accounts for a large share of themed demand — 983 users on removal alone.
  2. Enhancement is the most fragmented vocabulary (108 distinct prompts) — users know the outcome, not the tool name.
  3. Unique edits (292 prompts, 29%) show open-ended “natural language Photoshop” behavior.
  4. Sample prompts are only 7.7% of unique lines but drive ~70% of sessions — onboarding skews the funnel.
  5. Custom prompts skew 72% English; samples ship in 15+ languages — a gap worth validating with behavioral data.

Dataset breakdown

1,000 unique prompt strings
Unique prompts
1,000
Full coded set
Themes
495 49.5%
16 intent buckets
Unique edits
292
29.2% — uncategorized
Vague / sample
213
13.6% + 7.7%

How the thousand lines partition

Theme demand

Top 10 by unique users

Ranked themes (users)

Next step: Align homescreen sample prompts with the top five use cases so first-touch copy matches observed intent. Consider routing to sub-features (remove background, upscale) instead of dropping users only into free-text edit.

Top five themes — narrative

Volume leader

Background removal

983 users · 120 prompts · 2,951 generations

Users
983
Prompts
120
Gens
2,951

Stable vocabulary across languages — strongest candidate for a dedicated control instead of typed prompts.

Fragmentation

Enhancement & upscaling

385 users · 108 prompts · 1,421 generations

Prompts
108
Users
385
Gens
1,421

Many phrasings for “make it better” — guided presets (sharpen, upscale, denoise) would reduce guesswork.

Workflow

Background replacement

208 users · 73 prompts · 624 generations

Users
208
Prompts
73
Gens
624

Users specify environments and moods — pairs naturally with a two-step replace flow.

Precision

Object & element removal

184 users · 58 prompts · 517 generations

Users
184
Prompts
58
Gens
517

Targets are specific (text, watermark, person) — select-and-remove UX maps better than long prompts.

Expertise

Lighting, color & tone

68 users · 27 prompts · 207 generations

Users
68
Prompts
27
Gens
207

Photography vocabulary — presets and sliders align with how these users think.

Example phrasing

Top five themes — literal prompts

Background removal

“remove background” · “make the background transparent” · “delete background”

Enhancement

“upscale” · “sharpen” · “enhance the image quality and remove artifacts…”

Replacement

“change background” · “make the background white” · “beach background”

Object removal

“remove text” · “remove watermark” · “remove the people in the background”

Unique edits — verbs

292 prompts in the “unique edits” slice

Top verbs (frequency)

“Make” covers transformations; “move” signals spatial edits. The same verb hides different intents — these rows behave like mini creative briefs, not taxonomy noise.

Built-in sample prompts

3 prompts × 15+ locales

Users touched per sample (all languages)

Custom prompt language mix (illustrative)

English-heavy custom prompts vs localized onboarding — validate with product analytics.

Turtle (first sample) leads exposure. Non-English users often try samples but rarely return with custom text — worth a dedicated localization / confidence study.

Vague prompts

136 prompts — exploration & thin intent

Most repeated tokens

Objects & animals
~30%
cat, dog, car…
Testing
~25%
hi, test, “.”
Intent w/o context
~45%
edit, remove, 4k…