ComfyUI Node
Krea2 Apply Regional
A ComfyUI node in conditioning/krea2_regional with 15 inputs and 2 outputs.
Krea2 Apply Regional
- model
- conditioning
- regions
- base_loras
- MODEL
- CONDITIONING
◄restrict_img_attnfalse►
◄exclusive_maskstrue►
◄adaptive_masksoff►
◄adaptive_steps2►
◄adaptive_threshold0.45►
◄base_loras_exclude_regionsfalse►
◄region_lock_strength0.00►
◄region_lock_start0.35►
◄region_lock_end0.85►
◄restrict_end_percent1.00►
◄unmaskable_layersskip►
Categoryconditioning/krea2_regional
Inputs (15)
| Name | Type | Default | Description |
|---|---|---|---|
| model | MODEL | — | |
| conditioning | CONDITIONING | — | |
| regions | KREA2_REGIONS | — | |
| restrict_img_attn | BOOLEAN | false | Block image<->image attention across regions. Strongest anti-bleed lever; can look collaged. |
| exclusive_masks | BOOLEAN | true | Where grown/feathered region masks overlap, each token keeps only its strongest region (no double LoRA in the gap). |
| adaptive_masks | COMBO | off | FreeFuse-style adaptive routing: watch where the model actually places each subject during the first steps (prompt-token/image-token attention), then snap the region masks to those silhouettes. 'refine boxes' keeps discovery inside your (slightly grown) boxes; 'free' lets regions land anywhere. |
| adaptive_steps | INT | 21–8 | How many model calls to observe before snapping the masks. |
| adaptive_threshold | FLOAT | 0.450.05–0.95 | Affinity cutoff (fraction of each region's peak). Lower = larger discovered regions. |
| base_loras_exclude_regions | BOOLEAN | false | Style/base LoRAs apply everywhere EXCEPT inside the subject regions — keeps a style LoRA from contaminating character LoRAs. |
| region_lock_strength | FLOAT | 0.000–1 | Latent-only Region Lock: snapshot each region's predicted-clean latent once identity has formed and pull it back toward that snapshot each step. Suppresses late-step drift/mutation inside regions. 0 = off; 0.2-0.4 anchors, 0.7+ freezes hard. |
| region_lock_start | FLOAT | 0.350–0.95 | When (as a fraction of the schedule) to take the snapshot. |
| region_lock_end | FLOAT | 0.850.05–1 | When to release the lock so the model can integrate seams and lighting. |
| restrict_end_percent | FLOAT | 1.000.05–1 | When restrict_img_attn is on: release the image<->image isolation at this fraction of the schedule. Identity locks in during the restricted window; the open steps afterwards integrate seams and lighting. 1.0 = restricted for the whole run (classic). 0.4-0.6 is a good cohesion sweet spot. |
| base_lorasopt | KREA2_LORAS | — | |
| unmaskable_layersopt | COMBO | skip | Some LoRA layers (timestep/modulation embedders from newer trainers) have no spatial token axis and can't be region-masked. 'skip' drops them from REGION LoRAs (safest for isolation); 'apply globally' keeps their full effect but it touches the whole image. |
Outputs (2)
| Name | Type | Description |
|---|---|---|
| MODEL | MODEL | — |
| CONDITIONING | CONDITIONING | — |