Krea 2 - Regional Detailer:
Krea2RegionalDetailer makes a focused second pass over subjects or regions in an existing image. It finds the area to improve, takes a crop with surrounding context, resamples that crop using the region’s own conditioning and LoRAs, then blends the result back into the image. When subject detections are available, it matches them to regions so the detail pass follows detected silhouettes rather than relying only on the original region masks. Its automatic targeting can limit large regions to their head area or skip very large regions, helping you refine details without unnecessarily redrawing the whole composition. It also anchors the surrounding crop during sampling, which helps preserve nearby content.
Krea 2 - Regional Detailer Input Parameters:
image
The IMAGE you want to detail. The node uses it as the original source for every crop, even when detail passes overlap, and composites the finished crops onto a copy. This helps prevent one subject’s detail pass from becoming the starting point for another. No minimum, maximum, or default is specified.
model
The BASE MODEL used to resample each crop. Connect the base model, not the output of Apply Regional: the node applies each region’s LoRAs to a model clone for that region’s crop. Using an already region-modified model could apply effects you did not intend. No minimum, maximum, or default is specified.
vae
The VAE that encodes image crops for sampling and decodes the sampled results back to images. It should be appropriate for the model and image workflow you are using. No minimum, maximum, or default is specified.
regions
The KREA2_REGIONS data from the Builder, carrying region masks, conditioning, and any region LoRAs. The node uses these to decide where to work and how to sample each crop; regions without a mask or conditioning are not detailed. The first region in the Builder list is composited last, so it takes precedence where blended regions overlap. No minimum, maximum, or default is specified.
seed
The starting random seed for sampling. Each region uses this value plus its index in the regions list, giving regions distinct sampling seeds; changing it can change the detail result. Default: 0; minimum: 0; maximum: 18,446,744,073,709,551,615.
steps
The number of sampling steps for each detailed crop. More steps mean more sampling work per crop and can increase processing time, especially when several regions are detailed. Default: 8; minimum: 1; maximum: 100.
cfg
The classifier-free guidance value passed to the sampler. It controls how strongly sampling follows the supplied conditioning; the most useful value depends on your model and workflow. Default: 1.0; minimum: 0.0; maximum: 30.0; adjustment step: 0.1.
sampler_name
The sampling method used for each crop. Available options are the sampler names exposed by your installed ComfyUI KSampler; the source does not define a fixed option list or default for this input. Choose a sampler appropriate for your model and the look you want to preserve.
scheduler
The schedule used with the selected sampler for each crop. Available options are the schedulers exposed by your installed ComfyUI KSampler; the source does not define a fixed option list or default for this input. Changing it changes how the sampling steps are distributed.
denoise
How much of each selected area the detail pass may redraw. Lower values favor preservation: the node’s guidance describes 0.12–0.18 as a conservative clean-up pass, 0.20–0.25 as a stronger likeness pass, and 0.35 or higher as more likely to recompose content inside the mask. At 0, the node returns the original image without detailing. Default: 0.18; minimum: 0.0; maximum: 1.0; adjustment step: 0.01.
detection_masks
Optional MASK input containing one mask per detected subject. When it contains usable masks, the node matches detections to regions, crops around the matched subjects, and repaints their silhouettes rather than simply using the Builder region masks. This input takes precedence over segs. If it is connected but yields no usable masks, the node does not silently fall back to region-only detailing. No minimum, maximum, or default is specified.
segs
Optional SEGS detections, such as person or face detections in the format described by the source’s Impact Pack compatibility code. The node converts their cropped masks and crop locations into image-sized masks and uses them only when detection_masks is not connected. A segment missing a cropped mask or crop location is ignored. If connected detections produce no usable masks, no detailing is performed. No minimum, maximum, or default is specified.
base_conditioning
Optional CONDITIONING used as a template for conditioning extras associated with each region’s positive conditioning. If negative is not supplied, the node also uses these extras when constructing its zeroed-negative fallback, zeroing a tensor pooled_output when present. This input does not replace each region’s own conditioning. No minimum, maximum, or default is specified.
negative
Optional CONDITIONING supplied as the negative conditioning for crop sampling. When omitted, the node constructs a zeroed-negative fallback from the region’s conditioning and available base_conditioning extras. Connect this when your workflow requires a particular negative conditioning. No minimum, maximum, or default is specified.
crop_padding
The amount of context added around the selected area before sampling, measured as a fraction of that area’s longer bounding-box side. For example, 0.25 adds padding calculated from one quarter of that longer side on each edge, subject to the image boundaries. More padding gives the sampler more surrounding context and can enlarge the crop. Default: 0.25; minimum: 0.0; maximum: 1.0; adjustment step: 0.05.
detail_size
The target length, in pixels, for the sampled crop’s longer side. Crops larger than this are downscaled for sampling and then resized back; smaller crops are enlarged toward this size. Sampling dimensions are rounded to 16-pixel increments. A larger value can retain more crop resolution but increases sampling work. Default: 1024; minimum: 256; maximum: 4096; adjustment step: 16.
feather_px
The feathering amount, in image-resolution pixels, used at the composite boundary. It also contributes to the mask used during sampling, helping the transition between detailed and original pixels. Increase it for a softer transition; reduce it for a tighter edge. Default: 24; minimum: 0; maximum: 256.
mask_grow_px
In detection mode, expands the detected silhouette before feathering so edges such as hair can be included in the repaint. It does not add this growth in region-only mode. Too much growth can include nearby content you meant to preserve. Default: 8; minimum: 0; maximum: 128.
shift
Controls the sampling shift used for each crop. Options: model default (default) keeps the model’s existing shift and is the recommended setting for a low-denoise pass; auto (RAW only, per-crop) calculates a resolution-based shift for a compatible Krea 2 RAW Flux sampling model. The source warns against using auto with Turbo, which is described as using a fixed shift of 1.15. If the model is not compatible or the per-crop adjustment fails, the node keeps its existing shift.
detail_target
Controls how much of a selected subject or region may be detailed. Options: auto (default), full region, and face/head only. Auto details small areas fully, limits larger areas to a head-focused mask, and skips very large areas according to the two bounding-box thresholds below. Full region uses the selected mask without those auto decisions; face/head only keeps the upper part of its bounding box.
auto_face_bbox_frac
In auto mode, the bounding-box area threshold at which the node switches from full-area detailing to face/head-only detailing. It is a fraction of the full image area, not a fraction of the subject’s painted-mask pixels. Lowering it makes more areas head-only. Default: 0.18; minimum: 0.02; maximum: 1.0; adjustment step: 0.01.
auto_skip_bbox_frac
In auto mode, the bounding-box area threshold at which the node skips detailing an area altogether. This helps avoid a second broad regeneration pass over a very large region. Lowering it causes more areas to be skipped; choosing full region bypasses the auto skip policy. Default: 0.60; minimum: 0.05; maximum: 1.0; adjustment step: 0.01.
head_height_frac
The approximate top fraction of the subject’s bounding box retained when face/head-only targeting is used, either directly or through auto mode. Adjust it to include more or less of the upper subject; it is a geometric head-focused selection, not a separate face detection. Default: 0.38; minimum: 0.10; maximum: 0.80; adjustment step: 0.01.
Krea 2 - Regional Detailer Output Parameters:
image
The resulting IMAGE after eligible crops have been sampled and blended into the original. Areas outside the applied masks remain based on the input image. If there are no regions, denoise is 0, connected detections yield no usable masks, or no region qualifies for detailing, this output remains unchanged from the input image.
applied_masks
A MASK showing where detailing was actually composited, with feathered or grown edges represented by intermediate mask values. For a batch, it contains a corresponding mask for each image; overlapping applied areas are combined into a union. An all-zero mask means no detail was applied, making this output useful for checking whether your settings and detections selected the areas you expected.
Krea 2 - Regional Detailer Usage Tips:
- Start with
detail_targetset toautoanddenoisearound 0.12–0.18 for a restrained clean-up pass. Inspectapplied_masksto confirm which areas were changed before increasing denoise. - Connect
detection_maskswhen you have reliable subject silhouettes and want detailing to follow the detected subjects. If you usesegsinstead, leavedetection_masksdisconnected because it takes precedence. - Use the BASE
modelwith the Builder’sregions. The node applies each region’s LoRAs for its own crop, so an Apply Regional model output is not the intended connection. - Keep
shiftatmodel defaultunless you specifically need per-crop shift for a compatible Krea 2 RAW model. Do not select the RAW-only auto option for Turbo. - If auto mode omits a wanted subject, check its bounding-box size against
auto_face_bbox_fracandauto_skip_bbox_frac. Chooseface/head onlyorfull regiondeliberately when you need to override auto targeting. - Adjust
detail_sizewith processing cost in mind: the node samples each eligible crop separately, and a larger target creates more sampling work. Usecrop_paddingto provide enough context without making every crop unnecessarily large. - In detection mode, increase
mask_grow_pxmodestly if hair or silhouette edges are missed. Adjustfeather_pxif the blended boundary looks too abrupt or too broad.
Krea 2 - Regional Detailer Common Errors and Solutions:
[Krea2Detailer] detector input connected but yielded no usable masks — nothing detailed.
- Explanation: A detector input was connected, but it produced no masks the node could use. The node leaves the image unchanged rather than silently switching to region-only detailing;
applied_maskswill be empty. - Solution: Check that your detector found subjects and that its masks or SEGS cover the current image. If you intend to detail Builder region masks without detections, disconnect both detector inputs.
[Krea2Detailer] skipped a SEG: …
- Explanation: The node could not convert one SEGS entry into an image-sized mask. The warning includes the underlying reason, and other usable segments may still be processed.
- Solution: Check that the SEGS source supplies a cropped mask and crop location for each detection, and that the detections correspond to the image being detailed.
[Krea2Detailer] region N skipped by auto target policy
- Explanation: Auto targeting determined that the selected area’s bounding box was large enough to skip, or the chosen target policy produced no usable area. The affected region is not composited.
- Solution: Inspect the region mask and
applied_masks. If the large area is intentional, chooseface/head onlyorfull region; otherwise correct the mask or adjust the auto thresholds.
[Krea2Detailer] matched 0/… detections to … regions
- Explanation: This is an informational log, not necessarily an exception. No detected subject was assigned to an eligible region, so detection-mode regions may remain unchanged. Matching considers overlap and a limited nearby-distance rescue for drifted detections.
- Solution: Verify that detections and Builder masks refer to the same image and subjects, and that the masks overlap or are reasonably close. Review
applied_masksto see whether anything was actually changed.
[Krea2Detailer] lora '…' failed to apply: …
- Explanation: A region’s LoRA could not be applied; the node logs the failure and continues processing that crop without that LoRA.
- Solution: Check that the LoRA is available and compatible with the base model, then review the reason appended to the warning. Confirm that you connected the base model rather than an already region-modified model.
[Krea2Detailer] per-crop shift unavailable (…) — using the model's shift as-is.
- Explanation: The optional RAW per-crop shift could not be applied. The node continues with the model’s existing shift.
- Solution: Leave
shiftatmodel defaultunless you specifically need RAW per-crop shifting. If you do need it, check the reason in the log and confirm the model uses compatible Flux sampling; do not use auto shift with Turbo.
No error appears, but the image is unchanged
- Explanation: The node returns an unchanged image when
denoiseis 0 or there are no regions. Individual regions are also left alone when they lack a usable mask or conditioning, have no matched detection in detection mode, are skipped by targeting, or produce a crop smaller than 16 pixels in either dimension. - Solution: Set
denoiseabove 0, confirm the Builder supplied masked regions with conditioning, and inspectapplied_masks. If you connected detections, check their matches; if you use auto targeting, check its size thresholds.
Sampling or VAE error during a crop
- Explanation: The source does not define a single fixed error message for failures raised by VAE encoding, the sampler, or VAE decoding. Such an error can interrupt detailing rather than producing a completed composite.
- Solution: Read the specific error shown by ComfyUI, then check model–VAE compatibility and your sampler settings. If the issue occurs on a large crop, try a smaller
detail_sizeor fewer eligible regions to reduce the work required per pass.
