Skip to content

Detailer

Detailer is an optional workflow step that can be enabled from the menu for text, image, and control generation.

What Detailer does

Detect objects

Using selected model(s) and settings, Detailer detects objects in the image.

Settings: - Minimum confidence of detected object - Maximum number of detected objects - Max overlap between detected objects - Minimum and Maximum size of detected objects

Detail object

  1. Crop each detected object and resize it to a resolution suitable for the loaded model.
  2. Run inpaint for each cropped object using configured rules.
  3. Resize the result back to original size and place it into the original image.

Settings: - Prompt and Negative prompt
- Steps and Strength of inpaint operation
- Edge padding and Edge blur

Notes

Prompt

If detailer prompt is specified, it will be used, otherwise primary prompt will be used. In case you want to add detailer-specific prompts but also keep the primary prompt, you can use [prompt] in the detailer prompt text box to refer to it.

For example, [prompt]. blue eyes

Advanced Usage

Separating prompts

If Detailer prompt contains multiple lines, each line is used separately: - For each detailer model (in order) - For each detected object (in order)

Example with single model and 3 detected objects:

Base Prompt:
  three woman walking down the city street  
Detailer prompt:  
  pink makeup and red lipstick  
  green eyeshadow  
  blue eye contacts

If you have 2 models and 3 detected objects, the first model uses the first 3 lines (one per object), and the second model uses the next 3 lines (one per object).

The same rule applies to negative prompts.

Per-class prompts

When using a multi-class model (or several models with different classes), prompt lines can be assigned to detections by class name instead of by line position
Prefix a line with [CLASS=name] to route that line's prompt to any detection whose class matches name:

[CLASS=face] detailed eyes, sharp iris
[CLASS=hand] five fingers, correct anatomy

Class names are matched case-insensitively against the classes listed in logs when the model loads
Multiple classes can share one line: [CLASS=head,face] ....
Lines without a [CLASS=...] tag are used as before — assigned by position to any detection with no matching class tag

If a [CLASS=name] tag never matches a detection, a warning is logged in case it's a typo

The same syntax applies to negative prompts.

Note: Per-class prompts are not compatible with Merge detailers — merging combines all detections from a model into a single region, so per-class targeting no longer applies.

Segmentation

Some Detailer models can output segmentation masks, not just bounding boxes. To use segmentation masks, enable Use segmentation. If the selected model does not support segmentation, SD.Next falls back to normal bounding boxes.

Tip

Models that have segmentation capabilities often have seg in the name

Multi-class models

Most object detection models are trained on one class, for example face or hands. Some models support multiple classes, and you can choose which classes to detect. yolo11m is one example of a multi-class model. When using a multi-class model, available classes are listed in logs at runtime.

Models

YOLO

Typical Detailer models are from the YOLO model family, which is a popular object detection framework. SD.Next includes several pre-defined single-class Detailer models: face-yolo8n, hand_yolov8n, person_yolov8n-seg, eyes-v1, eyes-full-v1
And one multi-class model: yolo11m

Additional models can be added by downloading a pretrained model and placing it in the folder specified in Settings -> System paths -> Yolo models (default: models/yolo).

Transformers

In addition to YOLO models, Detailer can also use as several transformers-based models, some specific to vision-language tasks, and some for open-vocabulary detection and segmentation.:
- Facebook-SAM3 hybrid promptable concept segmentation and detection network - Qwen3-VL vision-language autoregressive foundation models, in 2B, 4B, and 8B variants - Florence-2 lightweight multi-task vision sequence-to-sequence models, in base and large variants - Grounding-DINO open-vocabulary object detection models, in tiny and base variants

Select any of the above models in the detailer model selection dropdown
and enter your human-readable target descriptions as detailer instructions

For example:

glasses, optional hat, left hand, largest tree in the background

Notes: - VL models are much larger, so use them with caution
- LLMs process data sequentially, meaning if a model cannot find one specific item, it may skip subsequent items in the prompt
- SAM3 is a gated model

Expert Mode

The model list is usually shown as a dropdown where one or more models can be selected. You can also use the button next to the list to switch to a text box for manual model input and optional parameters. Text is parsed as a model list separated by comma ,, newline \n, or semicolon ;. Each model can include parameters separated by colon :.

Example:

face-yolo8n:steps=5:denoising_strength=0.3,
eyes-v1:mask_blur=0.5
hand_yolov8n:sampler_name=DDIM

For full list of parameters, see Parameters page

Warning

Expert mode is intended for advanced use and SD.Next does not validate the input. Overriding some parameter values may result in unexpected behavior or errors