Explore, pick, and vary¶
Explore a deliberate set¶
The default scout contains four candidates. This is large enough to expose meaningful direction changes while remaining readable as a contact sheet and reasonable for local hardware.
Advanced workflows may use 1, 4, or 8 outputs. Each image receives an independent generator and a recorded actual seed so batching changes do not silently reuse one random stream.
Read recommendation badges correctly¶
- Closest structure compares control and output structure.
- Cleanest edges looks for excessive or missing edge behavior.
- Most distinct identifies a candidate least redundant with the set.
Badges never reorder candidates and never say “best.” A candidate can have more than one technical badge, and a designer can select one with none.
Pick without deleting alternatives¶
The candidate set remains intact. The study records the selection without deleting or reordering alternatives, and the selected candidate becomes the parent record for future work.
Vary with an explicit radius¶
Use subtle to refine one direction, balanced to reopen composition choices,
and bold to explore a wider neighborhood. Locks are model-independent requests;
check CapabilityReport for the backend’s exact support or approximation.
FLUX.2 Klein does not expose a native numeric denoise-strength parameter.
AIsketcher therefore maps the three public levels to deterministic, versioned
edit instructions, adds the active structure-lock instruction to the model
prompt, and records variation_strength_requested,
variation_strength_applied, and variation_strength_method in backend
metadata. Legacy SDXL image-to-image backends continue to apply the numeric
strength directly.
For seed-plan choices, see Seeds and output counts.