Adaptive ReActor Face Swapper

An adaptive ComfyUI video face-swapping workflow that uses ReActor, Florence-2, frame-similarity caching, and true lazy branch selection to process easy and difficult frames differently.

Instead of applying expensive facial masking to every frame, the workflow analyzes frame difficulty and selects one of two processing paths:

• Easy frames: ReActor face swap only

• Hard frames: ReActor face swap with ReActor Mask Helper

Hard frames may include open mouths, visible tongues, hands or fingers near the face, facial obstructions, strong profile angles, or other situations where a basic face swap may blend poorly.

Main features

• Adaptive per-frame quality routing

• ReActor face swapping

• ReActor Mask Helper for difficult frames

• Florence-2 frame descriptions

• Frame-similarity caching

• Reuses decisions for visually similar frames

• True lazy branch execution

• Sequential video processing

• Support for long videos through VHS meta-batching

• Optional Flux FaceDetailer section

• Low-VRAM guidance

• Built-in instructions and troubleshooting notes

• H.264 MP4 video output

How it works

The original video frame is sent to the AI Quality Controller.

The controller compares the current frame with the previous frame:

• Similar frames reuse the previous Florence description and hard/easy decision.

• Significantly changed frames are described again and receive a new decision.

The decision is sent to a lazy image selector:

• False / Easy selects the ReActor-only result.

• True / Hard selects the ReActor + Mask Helper result.

Only the selected processing branch should execute.

Required companion workflow

This workflow uses a saved ReActor face model.

A separate Face Model Compiler is provided for building a compatible .safetensors face model from either:

• image folders

• a source video with sampled frames

Place the resulting face model in:
"ComfyUI/models/reactor/faces/"

Then select it through the included ReActor Load Face Model node.

Required custom nodes

The workflow uses nodes from packages including:

• ComfyUI-ReActor

• ComfyUI Video Helper Suite

• Florence-2 nodes

• ComfyUI Impact Pack / Impact Subpack

• ComfyUI-GGUF for the optional FaceDetailer path

• Frame Similarity Cache and Lazy Image Selector companion nodes

Use ComfyUI Manager to install missing nodes where available.

Required models

Depending on the enabled sections, you may need:

• inswapper_128.onnx

• RetinaFace face detector

• GFPGAN or another supported face-restoration model

• Florence-2 model

• SAM model for ReActor Mask Helper

• saved ReActor face model

• optional Flux, CLIP, VAE, and YOLO face detector for FaceDetailer

Important settings

Keep:

"frames_per_batch = 1"

The similarity cache compares frames sequentially and is not designed for multiple frames per meta-batch.

Match the Video Combine frame rate to the source video.

For lower-VRAM GPUs, begin with an input width of approximately 512–640.

Important preview warning

ComfyUI Preview Image nodes are output nodes.

Enabling a preview connected directly to either processing branch can force that branch to execute, defeating lazy routing.

During normal rendering, keep these previews disabled:

• ReActor-only result

• Mask Helper result

• swapped-face preview

• FaceDetailer preview

The final lazy selector should be the only processed-image output connected to the active Video Combine node.

Starting a different video

The frame-similarity cache may persist while ComfyUI remains open.

Before processing a different source video:

• restart ComfyUI, or

• reset the cache once and return the reset control to its normal state

This prevents the first frames of a new video from reusing a decision from the previous video.

Performance expectations

Processing speed depends heavily on the selected branch.

• Easy frames should remain close to normal ReActor processing time.

• Hard frames will take longer because Mask Helper performs face detection, segmentation, and blending.

• FaceDetailer adds substantial additional processing time.

Optional FaceDetailer

The included FaceDetailer section is disabled by default.

It can improve small or compressed faces, but it may increase render time and cause identity drift if denoise is too high.

Recommended starting values:
Steps: 4-8

Denoise: 0.10-0.15

Seed: Fixed for video

Guide size: 384-512

Recommended use

This workflow is best suited for videos containing a mixture of:

• normal front-facing frames

• mouth movement

• profile changes

• hands near the face

• partial facial obstruction

• varying expressions

Limitations

The AI controller relies on Florence-2 descriptions and keyword rules. It may occasionally misclassify a frame.

Results also depend on:

• the quality of the saved face model

• source-video resolution

• face size and visibility

• motion blur

• lighting

• facial obstruction

• selected restoration and masking settings

Responsible use

Use this workflow only with content you are authorized to modify. Obtain appropriate consent and clearly disclose synthetic or altered media when it could reasonably be mistaken for authentic footage.