ReActor Face Model Compiler — Video or Folder Input

The ReActor Face Model Compiler is a ComfyUI workflow designed to create reusable ReActor .safetensors face models from either a reference-image collection or a directly loaded video.

It is intended to serve as the companion workflow for my adaptive ReActor face-swap system. Build the identity model once, save it to ReActor, and reuse it across future image and video face-swap projects.

Two Input Modes

This workflow supports two different ways to create a face model:

📁 Folder Mode
Build a model from up to six folders of curated reference images.

🎬 Video Mode
Load a video and automatically sample frames from it to create the reference dataset.

The input selector controls which source is used:

• False - Folder Mode

• True - Video Mode

The unselected input path is skipped.

Folder Mode

Folder Mode is designed for curated datasets where you already have individual reference images.

The six folder inputs can be used to organize different types of references, such as:

• front-facing images

• left-side profiles

• right-side profiles

• smiles and talking expressions

• open-mouth or tongue-visible expressions

• different lighting conditions

• close-up and medium-distance images

All enabled folders are normalized, combined into one image batch, and passed to the ReActor face-model builder.

You do not need to use all six folders. Unused folder loaders can point to an empty folder.

Video Mode

Video Mode allows you to create a face model directly from footage without manually extracting every frame first.

The workflow samples frames using select_every_nth.

Suggested starting values:

• Every 5th frame - larger dataset and more repeated frames

• Every 10th frame - balanced starting point

• Every 15th frame - smaller and faster dataset

For a 30 FPS video:

• Every 5th frame produces approximately 6 references per second

• Every 10th frame produces approximately 3 references per second

• Every 15th frame produces approximately 2 references per second

Avoid using every frame from a long video. Hundreds of nearly identical images usually do not improve the model and can increase processing time or weaken the average identity.

A useful source video should contain varied:

• head angles

• facial expressions

• camera distances

• lighting conditions

• left and right profiles

• neutral, smiling, and talking frames

Reference Quality

The quality of the finished face model depends heavily on the quality and variety of the references.

Good references generally have:

• a clear and reasonably large face

• sharp eyes, nose, and mouth

• limited motion blur

• natural facial proportions

• one dominant person

• varied angles and expressions

• minimal compression or filtering

Avoid:

• duplicates and near-duplicate frames

• heavily blurred images

• very small or distant faces

• strong beauty filters

• sunglasses or severe obstruction

• extreme compression

• images containing multiple equally prominent faces

• images of different people

More images are not automatically better. A varied set of 20 clear references will often outperform a set of 100 repetitive or low-quality images.

Recommended Dataset Size

Suggested ranges:

• 8–12 images - minimum practical test set

• 15–40 images - recommended starting range

• 40–75 images - useful for broader variation

• 75+ images - only worthwhile when they add genuinely new angles, expressions, or lighting

For video input, adjust select_every_nth until the sampled frame count falls into a reasonable range.

How to Use Folder Mode

1. Place the reference images into one or more folders.

2. Enter the full directory path in each Reference Folder node.

3. Keep image_load_cap at 0 to load every image.

4. Set the input selector to False.

5. Enter a unique face-model name.

6. Queue the workflow.

7. Confirm the saved model appears in the ReActor faces directory.

How to Use Video Mode

1. Load the reference video in Optional Video Reference Input.

2. Set select_every_nth, normally between 5 and 15.

3. Set the input selector to True.

4. Enter a unique face-model name.

5. Queue the workflow.

6. Confirm the saved model appears in the ReActor faces directory.

Output

The completed model is saved to:

"ComfyUI/models/reactor/faces/"

Example:

"person_name_v01.safetensors"

The saved model can then be selected in a ReActor Load Face Model node.

Recommended Build Method

Start with:

"Mean"

Mean generally provides the most stable identity average when the dataset contains varied angles, expressions, lighting, and camera distances.

Versioning

Do not immediately overwrite a model that already works well.

Use versioned filenames such as:

person_name_v01

person_name_v02_more_profiles

person_name_v03_mouth_expression

sperson_name_v04_cleaner_dataset

Test each version on the same collection of difficult frames before deciding which one performs best.

Testing the Finished Model

A face model should be tested on more than a single front-facing image.

Recommended test cases:

• neutral front view

• smile

• open mouth

• tongue visible

• left profile

• right profile

• hand close to the face

• partially obstructed face

• low-resolution frame

• bright and dark lighting

A model may perform very well in front-facing frames while still failing on profiles or complex mouth movement. Add references that specifically cover the weak cases, rebuild the model, and compare the new version.

Troubleshooting

Only one reference folder affects the model

Confirm the final builder is connected to:

"All Reference Images — Folders 1–6"

and not directly to a single normalization node.

The identity looks weak or averaged

Remove:

• blurred references

• filtered images

• distant faces

• duplicates

• images with inconsistent appearance

Add sharper front, profile, and expressive references.

Profile swaps fail

Add clear left- and right-profile images. Slight three-quarter angles are useful, but they do not fully replace true profile references.

Open-mouth or tongue frames fail

Add references showing:

• natural smiles

• talking

• partially open mouths

• widely open mouths

• tongue-visible expressions

The wrong person appears in the model

Remove images containing another prominent face. Group photos and screenshots with multiple similarly sized faces are poor references.

Video Mode creates too many frames

Increase select_every_nth.

For example:

5 → 10

10 → 15

15 → 20

Video Mode produces repetitive references

Use a shorter or more varied source clip, or increase select_every_nth.

No model file is created

Confirm:

• ReActor is installed correctly

• model saving is enabled

• the filename is valid

• the output directory is writable

• all required models and dependencies are installed

Required Custom Nodes

This workflow requires:

• ComfyUI-ReActor

• Video Helper Suite

• the included lazy input selector custom node

ComfyUI Manager can be used to identify and install missing nodes.

Companion Workflow

This compiler was created as the companion tool for my adaptive ReActor face-swap workflow.

The intended process is:

1. Build the identity with this compiler.

2. Load the resulting .safetensors model in the face-swap workflow.

3. Process easy frames through ReActor.

4. Route difficult frames through ReActor Mask Helper.

Personal Notes

A face model cannot recover facial detail that is absent from the source material. Small, noisy, compressed, or blurred faces provide less useful identity information.

A video is convenient, but it is not always better than a curated folder. Video frames often contain repeated poses, motion blur, compression, and inconsistent sharpness. For the strongest results, manually remove poor or repetitive frames before compiling.

The best reference set is not necessarily the largest one. It is the set that covers the widest range of useful facial information while remaining clean and consistent.

Important Notice (LegalMumboJumbo)

Use this workflow only with reference material that you are authorized to use.

Do not create or distribute deceptive face models involving people who have not consented. Clearly label synthetic or altered media whenever it could reasonably be mistaken for authentic footage.

This workflow is intended for authorized creative work, fictional characters, private experimentation, education, and responsible AI development.