Skip to main content

Seedream 5.0 Pro

BytePlus's controllable image model — generation and interactive editing with up to 10 reference images and region- / sketch-guided edits. The image-side sibling of the Seedance video family.

Capabilities

FeatureSupport
Text-to-ImageYes
Image-to-Image (editing)Yes
Reference ImagesUp to 10
Quality Tiers1.5K, 2K
Aspect Ratios1:1, 2:3, 3:2, 3:4, 4:3, 9:16, 16:9, 21:9
Output FormatPNG
Outputs per RunUp to 4
Sketch / region-guided editsYes (doodle editor + coordinate markers)
Layer WorkbenchYes (composite rail images as layers)
SeedNo
Negative PromptNo

Sketch-Guided Editing

Seedream understands drawn guidance on top of the source image. Open the doodle editor on an image node and:

  • Scribble a region and describe what goes there — "replace the marked area with a glass skylight".
  • Drop coordinate markers (the ⊕ crosshair tool) to pin exact positions: paired same-color markers tell the model "move this here" with pixel-level intent.

This is the most precise edit control of any image model in NGMC — use it when "left of the lamp" prose isn't landing.

Reference Images

Up to 10 references in one run. Like Seedance on the video side, references guide identity, style, and composition; use @-mentions to give each one a role. See Referencing Images in Prompts.

Layer Workbench

Image nodes running Seedream 5.0 Pro (with 2+ images in the output rail) unlock the Layer Workbench — a Photoshop-style compositor where each rail image becomes a layer: drag / scale / rotate, reorder, adjust opacity, then bake the composite back to the rail at the background's native resolution. Combine it with cutout outputs for quick collage-style comps.

Prompting Tips

  • Direct the camera like Seedance. The Seed-family models share prompt DNA — composition and lighting language ("low-angle, golden hour, 35mm") transfers directly from your video prompts.
  • Use markers for spatial edits. Coordinates beat prose for "put it exactly there" tasks.
  • Batch up to 4 outputs when exploring; keep 1 for controlled edits.

Limitations

  • No 4K tier (1.5K / 2K only)
  • No seed, negative prompt, or masked inpainting
  • No streaming preview — outputs arrive complete
  • Native multi-image set generation (one prompt → coordinated image series) is not available

See Also