FisherKing-ZiT-[Detailer]-I2I-ReferenceWorkflow-v1.0 [Low VRAM]

Overview

A general-purpose, mask-driven Detailer workflow for ComfyUI.

Unlike detailer workflows tied to a predefined target, this workflow allows the region being detailed to be selected independently from the generation pipeline.

The supplied implementation uses Ultralytics + SAM for target detection and segmentation. By changing the Ultralytics detection model, the same workflow can be used to detail eyes, faces, heads, hands, characters, body regions, objects, or other detectable targets without restructuring the pipeline.

The selected region can also be inverted, allowing the detected subject to remain protected while the surrounding region is regenerated.

Generic architecture:

Select → Mask → Crop → Detail → Stitch

The selection mechanism and detailing engine remain independent, allowing each stage to be configured and experimented with separately.

Pipeline

The included implementation uses:

Detect → Segment → Crop → Detail → Stitch

Detect — Ultralytics identifies the target region using the selected detection model.

Segment — SAM refines the detected region into a more accurate mask.

Crop — Inpaint Crop extracts the localized region while retaining the information required to place it back into the source image.

Detail — The cropped region is regenerated through a standard VAE → KSampler → VAE pipeline.

Stitch — The detailed crop is composited back into the original image.

The Detailer Diagnostics section exposes intermediate results for inspection and troubleshooting.

Why a Modular Detailer?

The objective is not to replace FaceDetailer. FaceDetailer remains an excellent all-in-one solution for conventional detailing workflows.

This workflow instead exposes the individual stages of the detailing process and removes the assumption that the target must always be a face.

The modular architecture provides:

• Transparent detection and sampling stages

• Selectable detailing targets

• Independent control over the generation model and LoRAs

• Independent prompt and sampling configuration

• Easy inspection of intermediate results

• Reusable mask-driven detailing architecture

• Simple experimentation with detection and segmentation

• Target switching without rebuilding the pipeline

• Support for both detailing the selected region and preserving it

The key difference is simple:

FaceDetailer selects the face.
This workflow allows you to select what gets detailed.

Target Selection

The included workflow uses an Ultralytics detector to determine the target region.

The selected detector model effectively defines what the Detailer operates on.

For example:

Eye detector → Eye Detailer

Face detector → Face Detailer

Head detector → Head Detailer

Hand detector → Hand Detailer

Person detector → Person / Character Detailer

Other suitable detection models can be used for other targets.

Changing the target only requires changing the appropriate detector model.

The downstream:

Segment → Crop → Detail → Stitch

pipeline remains unchanged.

No pipeline restructuring is required when switching detection targets.

Segmentation

Detection determines the approximate target location, while SAM refines the detected region into the mask used by the detailing pipeline.

This separation allows the detector and segmentation stages to be tuned independently.

Detection quality is therefore important: SAM can refine a successfully detected region, but the detector must first provide a suitable target for segmentation.

The exposed mask preview makes it easy to verify the selected region before evaluating the detailing result.

Detailing Engine

The included example uses ZiT Redcraft, but the detailing engine is intentionally independent of the detection and segmentation stages.

Other compatible image-generation models can be substituted by replacing the appropriate:

Model / CLIP / VAE / Conditioning

components and configuring any model-specific sampling requirements.

The exposed sampling stage also provides direct control over:

• Prompt

• LoRAs

• Sampler

• Scheduler

• Steps

• CFG

• Denoise

• Model-specific sampling configuration

This makes the generation stage independently configurable without changing how the target region is detected, segmented, cropped or stitched.

LoRA Support

Because the detailing sampler is exposed as a standard generation pipeline, LoRAs can be applied directly to the detailing model.

This allows the workflow to perform more than conventional restoration or enhancement.

Depending on the model, LoRA and prompt, the selected region can be refined, restyled or substantially regenerated while the surrounding image remains preserved.

Multiple LoRAs may also be combined as supported by the selected generation model.

LoRAs are optional and are not required for the detection, segmentation or stitching stages.

Mask Inversion

The same target selection can be used in two different ways.

Standard Mask

The detected region becomes the detailing target:

Detect Target → Detail Target → Preserve Surroundings

For example:

Person detector → regenerate/detail the person while preserving the background

Inverted Mask

By enabling mask inversion in Inpaint Crop, the detected region instead becomes the protected region:

Detect Target → Preserve Target → Detail Surroundings

For example:

Person detector → preserve the character while regenerating the background

This allows the same detection and detailing pipeline to operate either inside or outside the selected region without restructuring the workflow.

Detailer Diagnostics

The workflow exposes several intermediate stages through the Detailer Diagnostics section.

These previews allow the user to inspect:

• Detected / segmented mask

• Source crop

• Active detailing region

• Detailed crop

• Final stitched result

This makes it easier to identify where an unexpected result originates.

For example:

Incorrect target → inspect detection

Poor mask → inspect segmentation

Incorrect crop/context → inspect Inpaint Crop

Weak or excessive modification → inspect prompt, LoRA, denoise and sampling

Stitching artifacts → inspect mask and stitch configuration

The diagnostic nodes are provided for visibility and troubleshooting and do not define the detailing target themselves.

Modular Architecture

The workflow can be understood as three independent layers.

1. Selection Layer

What should be detailed?

Ultralytics or another compatible selection mechanism identifies the target.

2. Mask Layer

Where is regeneration allowed?

SAM and the crop/mask configuration determine the active region.

Mask inversion can reverse this relationship so that the selected target is protected instead.

3. Generation Layer

How should the region be regenerated?

The model, LoRAs, prompt and sampling configuration determine the resulting detail.

Because these layers are separated, each can be modified without unnecessarily changing the others.

Extending the Workflow

The exposed crop and generation stages also provide convenient entry and exit points for additional processing.

The cropped region can be processed before sampling, and the generated crop can be processed again before stitching.

This allows additional processing stages to be inserted or chained while retaining the original detection, mask and stitch information.

The generation stage can therefore be expanded, replaced or chained depending on the intended detailing task.

Important

The workflow is only as reliable as the target selection used for a particular image.

Different detector models may perform differently depending on:

• Image resolution

• Target size

• Viewing angle

• Occlusion

• Stylization

• Unusual or heavily modified anatomy

If the intended region is not detected correctly, first inspect the Detailer Diagnostics and adjust the detector, detection settings or preprocessing resolution before changing the generation stage.

Once the intended region is correctly detected and segmented, the downstream detailing pipeline can remain unchanged.

Summary

This workflow generalizes the familiar Detailer pattern by separating what gets detailed from how it gets detailed.

Select → Mask → Crop → Detail → Stitch

Change the detector to change the target.

Change the mask polarity to detail or preserve that target.

Change the model, prompt, LoRAs or sampling configuration to change how it is regenerated.

The result is a reusable general-purpose Detailer architecture that can be adapted to different targets without rebuilding the workflow.