The biggest time sink in text-to-image is rarely the first frame—it’s when the shot is already 80% there, but the light is flat, the headline is cramped, or one bottle clutters the set… and a full re-roll changes the face, breaks the layout, and scrambles the type.
The old path is “regenerate + Photoshop.” One of Image 2’s strengths is multi-turn natural-language editing: point at what to change and what to keep—like briefing a retoucher. This post shares 5 tested image-2 commands for editing one thing without wrecking the whole frame.
Use it as a practical playbook: lock invariants, change one goal per turn, then proofread before export.
📌 Edit commands below are in English (Image 2 handles structured English well). Every turn should include
keep … unchanged. Before shipping, manually proof on-image text, trademarks, and compliance wording.
1. Why Image 2 fits “chat edits” instead of re-rolls
Versus rewriting the whole prompt and rolling a new image, Image 2 wins because:
-
Stronger local intent — say “only brighten the key light from the left, keep pose and face identical” and it prefers lighting moves over reinventing the person.
-
Explicit locks for invariants — logo, face, product silhouette, already-correct headline lines—pin them with
verbatim/unchangedto reduce “fix A, break B.” -
Iteration cost close to chat — 3–5 polish turns usually beat 10 blind re-rolls + manual composites, especially for marketing heroes and e-commerce detail shots.
For growth and design teams: AI image editing shifts from lottery pulls to a controllable retouch conversation.
2. Standard workflow (five steps)
-
Generate a structurally usable base — composition, subject, and overall style can be “good enough”; don’t chase perfect micro-detail in round one.
-
List locks vs this turn’s single change — locks: face / logo / product shape / correct copy; this turn: light OR background OR one text line—never five asks at once.
-
Use a three-part command: goal + locks + bans —
Change: …/Keep unchanged: …/Do not: …. -
Accept one goal per turn — finish light before type; finish type before cleanup. Mixing goals causes interference.
-
Full proof before export — especially in-image text, finger/edge artifacts, and brand-color drift.
3. Five tested commands (copy-paste)
Command blocks stay in English across locales; narration below explains each case.
Case 1: Change lighting only—keep face and pose
Scenario: a flat portrait hero needs left rim light without a face swap.
Multi-turn edit on the current image with Image 2.
Change ONLY lighting: add soft cinematic rim light from camera-left,
slightly deepen shadow on the right cheek for depth.
Keep unchanged: face identity, expression, hairstyle, pose, wardrobe, framing, background layout.
Do not regenerate the person. Do not change skin tone dramatically.
Photorealistic, natural contact shadows, Image2 commercial portrait quality.
Result: mood “stands up” while identity stays recognizable—classic conversational edit usage.
Case 2: Create headline negative space without redesigning the subject
Scenario: the subject is too large; no room for a top Headline.
Edit the current poster layout with Image 2.
Change: push the main subject slightly downward and scale to ~85%,
create clean negative space in the top 20% for a future headline.
Keep unchanged: subject identity, outfit colors, overall art style, color grade.
Do not add new objects. Do not mirror the subject.
Maintain high-end marketing poster look, Image 2 quality.
Result: safe space for later type—avoids text covering the face in natural-language retouch workflows.
Case 3: Fix one poster line; leave other copy alone
Scenario: the promo percentage is wrong; price and button are already correct.
In-image text edit using Image 2 multi-turn editing.
Replace the headline text ONLY with exactly: "SUMMER SALE 40% OFF"
Keep unchanged: all other on-image text (price, button, footer) verbatim,
font style hierarchy, poster artwork, subject, colors.
No gibberish, no mirrored letters, no missing strokes.
Crisp legibility, commercial poster typography, Image2 text accuracy.
Result: no full AI poster re-roll for one word—far higher success than rewriting the whole prompt.
Case 4: Local cleanup / remove a distracting prop
Scenario: an extra bottle on the table steals focus from the SKU.
Local cleanup edit with Image 2.
Remove the small bottle on the far right of the table only.
Fill the area with consistent table texture and lighting.
Keep unchanged: main product shape, logo, label text, camera angle, overall lifestyle mood.
Do not restyle the product. Do not add new props.
Photorealistic e-commerce secondary image, Image 2 high detail.
Result: closer to “native scene” than heavy PS healing—fast cleanup for e-commerce secondaries.
Case 5: Three-round progression (light → background → CTA copy)
Scenario: a decent app launch frame that needs chat polish to go live.
Round 1：
Keep UI mockup and headline unchanged. Increase contrast slightly and add cool blue ambient fill from the right. Do not alter icon shapes.
Round 2：
Keep subject and text unchanged. Simplify the background: reduce grid noise, keep a clean dark navy gradient only.
Round 3：
Change the CTA button label ONLY to exactly: "Try Free". Keep button shape, color, and all other text verbatim. No extra characters.
Result: shows the right multi-turn editing rhythm—one goal per turn; after three rounds, shippable quality.
4. Universal command template
Save this block; change only the bracketed fields per turn:
Multi-turn natural-language edit with Image 2 on the CURRENT image.
Change (ONE goal only): [lighting / layout / one text line / remove object / color grade].
Keep unchanged: [face / product / logo / other text lines / pose / style].
Do not: [full regenerate / new people / invent logos / change aspect ratio].
Success criteria: [e.g. readable headline, same identity, clean background].
Output: commercial-ready, Image 2 quality.
5. Re-roll vs multi-turn edit comparison
| Step | Full re-roll | Manual PS fix | Image 2 multi-turn chat edit |
|---|---|---|---|
| Keep subject identity | Unstable | High | High (if you write keep) |
| Change one text line | Often wrecks layout too | Precise but slow | Fast for iteration |
| Time per pass | Repeated lottery draws | 0.5–2 hours | About 5–20 minutes / turn set |
| Learning curve | Low | Design skill | Learn “one change per turn” |
| Best when | Structure is fully wrong | Pixel-perfect finish | 80→95 sprint |
6. Who benefits most?
✅ Marketing / growth teammates
Use Image 2 to A/B light and copy fast—without asking design to rebuild every time.
✅ E-commerce ops & DTC sellers
Detail/main-image nits (clutter, color cast, one wrong line) fixed on the spot with conversational edits.
✅ UI / brand designers
Treat image-2 as a sketch session: lock layout, multi-turn toward final, then polish in Figma.
✅ Content creators
When a cover is almost right, nudge mood light and title in natural language instead of starting over.
7. Three pitfalls
- ❌ Don’t pack five asks into one turn — “fix light + swap background + rewrite three lines + change pose” makes Image 2 drop balls; strictly one goal.
- ❌ Don’t omit Keep unchanged — without locks, the model “helpfully” redraws faces or reflows all type.
- ❌ Don’t treat multi-turn as proof-free — text edits can still miss strokes; human-check every AI edit before export.
8. Closing thoughts
Image 2 multi-turn natural-language editing turns “re-roll and pray” into “brief a retoucher.”
Next time a frame is already mostly usable—just light, layout, or one copy line—don’t hit regenerate first. Run three image-2 chat turns with the templates here.
Change one thing, lock one thing, verify one thing—you’ll see what Image2 really saves: the good compositions destroyed by blind re-rolls.
Hands-on notes; commands free to copy and adapt. Before shipping, verify copy compliance and brand visual rules.
