Edit images with written instructions
Upload one or more reference images, state the exact change, and say what should remain intact. H3 Imagine keeps the browser workflow available without a node graph.
Start with a concrete subject
Keep each revision to one focused request, then use the previous output as the next reference.
Name the composition
Keep each revision to one focused request, then use the previous output as the next reference.
State the details to preserve
Keep each revision to one focused request, then use the previous output as the next reference.
Write an edit that preserves the right things
Reference-based editing is most reliable when the instruction contains two parts: what must change and what must remain unchanged. Start by identifying a single target, such as the background, a product material, a selected object, or the time of day. Then name the details that are already correct: the camera angle, subject identity, product geometry, lighting direction, typography, or framing. This gives the model a boundary instead of asking it to reinterpret the whole image.
Useful instruction patterns
Background replacement
Replace only the background with a quiet moonlit library. Keep the person, pose, clothing, lens perspective, and face unchanged.
Material change
Change the bottle from frosted glass to matte white ceramic. Preserve its silhouette, label placement, camera angle, and shadow.
Lighting change
Make this a late-afternoon scene with warm window light. Keep every object, layout, and product color unchanged.
Composition merge
Combine these product references into one gift set. Keep each product recognizable and place them on one travertine surface.
Use references deliberately
Add one reference image for each visual fact the result must carry forward. A product packshot can control shape and label placement; a portrait can control subject identity; a location photo can control the setting. Adding unrelated visual directions makes it harder to tell which input should win. For a first pass, upload only the images that matter to the requested change, generate one image, and make the next request from the strongest result.
Ready to make an image?
Start with a prompt or reference image in the browser generator.
Open generator