<h1 id="adaptive-reactor-face-swapper">Adaptive ReActor Face Swapper</h1>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:<ul><li>Easy frames: ReActor face swap only</li><li>Hard frames: ReActor face swap with ReActor Mask Helper</li></ul>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.<h2 id="main-features">Main features</h2><ul><li>Adaptive per-frame quality routing</li><li>ReActor face swapping</li><li>ReActor Mask Helper for difficult frames</li><li>Florence-2 frame descriptions</li><li>Frame-similarity caching</li><li>Reuses decisions for visually similar frames</li><li>True lazy branch execution</li><li>Sequential video processing</li><li>Support for long videos through VHS meta-batching</li><li>Optional Flux FaceDetailer section</li><li>Low-VRAM guidance</li><li>Built-in instructions and troubleshooting notes</li><li>H.264 MP4 video output</li></ul><h2 id="how-it-works">How it works</h2>The original video frame is sent to the AI Quality Controller.The controller compares the current frame with the previous frame:<ul><li>Similar frames reuse the previous Florence description and hard/easy decision.</li><li>Significantly changed frames are described again and receive a new decision.</li></ul>The decision is sent to a lazy image selector:<ul><li><code>False / Easy</code> selects the ReActor-only result.</li><li><code>True / Hard</code> selects the ReActor + Mask Helper result.</li></ul>Only the selected processing branch should execute.<h2 id="required-companion-workflow">Required companion workflow</h2>This workflow uses a saved ReActor face model.A separate Face Model Compiler is provided for building a compatible <code>.safetensors</code> face model from either:<ul><li>image folders</li><li>a source video with sampled frames</li></ul>Place the resulting face model in: "ComfyUI/models/reactor/faces/"Then select it through the included <code>ReActor Load Face Model</code> node.<h2 id="required-custom-nodes">Required custom nodes</h2>The workflow uses nodes from packages including:<ul><li> ComfyUI-ReActor </li><li> ComfyUI Video Helper Suite </li><li> Florence-2 nodes </li><li> ComfyUI Impact Pack / Impact Subpack </li><li> ComfyUI-GGUF for the optional FaceDetailer path </li><li> Frame Similarity Cache and Lazy Image Selector companion nodes </li></ul>Use ComfyUI Manager to install missing nodes where available.<h2 id="required-models">Required models</h2>Depending on the enabled sections, you may need:<ul><li><code>inswapper_128.onnx</code> </li><li> RetinaFace face detector </li><li> GFPGAN or another supported face-restoration model </li><li> Florence-2 model </li><li> SAM model for ReActor Mask Helper </li><li> saved ReActor face model </li><li> optional Flux, CLIP, VAE, and YOLO face detector for FaceDetailer </li></ul><h2 id="important-settings">Important settings</h2>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 <code>512–640</code>.<h2 id="important-preview-warning">Important preview warning</h2>ComfyUI <code>Preview Image</code> 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:<ul><li> ReActor-only result </li><li> Mask Helper result </li><li> swapped-face preview </li><li> FaceDetailer preview </li></ul>The final lazy selector should be the only processed-image output connected to the active Video Combine node.<h2 id="starting-a-different-video">Starting a different video</h2>The frame-similarity cache may persist while ComfyUI remains open.Before processing a different source video:<ul><li> restart ComfyUI, or </li><li> reset the cache once and return the reset control to its normal state </li></ul>This prevents the first frames of a new video from reusing a decision from the previous video.<h2 id="performance-expectations">Performance expectations</h2>Processing speed depends heavily on the selected branch.<ul><li> Easy frames should remain close to normal ReActor processing time. </li><li> Hard frames will take longer because Mask Helper performs face detection, segmentation, and blending. </li><li> FaceDetailer adds substantial additional processing time. </li></ul><h2 id="optional-facedetailer">Optional FaceDetailer</h2>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: 4-8Denoise: 0.10-0.15Seed: Fixed for videoGuide size: 384-512<h2 id="recommended-use">Recommended use</h2>This workflow is best suited for videos containing a mixture of:<ul><li> normal front-facing frames </li><li> mouth movement </li><li> profile changes </li><li> hands near the face </li><li> partial facial obstruction </li><li> varying expressions </li></ul><h2 id="limitations">Limitations</h2>The AI controller relies on Florence-2 descriptions and keyword rules. It may occasionally misclassify a frame.Results also depend on:<ul><li> the quality of the saved face model </li><li> source-video resolution </li><li> face size and visibility </li><li> motion blur </li><li> lighting </li><li> facial obstruction </li><li> selected restoration and masking settings </li></ul><h2 id="responsible-use">Responsible use </h2>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.

Adaptive ReActor Face Swapper – AI Quality Routing

v1.0 - Initial Public ReleaseFirst public release of the Adaptive ReActor Face Swapper workflow.Included features<ul><li>Adaptive hard/easy frame classification</li><li>ReActor-only processing for normal frames</li><li>ReActor Mask Helper processing for difficult frames</li><li>Florence-2 frame descriptions</li><li>Frame-similarity decision caching</li><li>True lazy branch execution</li><li>Sequential VHS meta-batch processing</li><li>Long-video support</li><li>Optional Flux FaceDetailer section</li><li>Built-in setup guide and troubleshooting notes</li><li>Low-VRAM recommendations</li><li>H.264 MP4 output configuration</li></ul>Important requirements<ul><li>Keep <code>frames_per_batch</code> set to <code>1</code>.</li><li>Keep processed branch preview nodes disabled during normal rendering.</li><li>Reset the frame-similarity cache or restart ComfyUI before processing a different video.</li><li>A compatible ReActor <code>.safetensors</code> face model is required.</li><li>The companion Face Model Compiler can create a face model from image folders or sampled video frames.</li></ul>Recommended first testUse a short video clip before attempting a full-length render. Verify that:<ul><li>ordinary frames take the ReActor-only path;</li><li>difficult mouth, profile, hand, or obstruction frames use Mask Helper;</li><li>the final video contains different sequential frames;</li><li>the output frame rate matches the source video.</li></ul>Known limitationsFrame classification relies on Florence-2 captions and keyword rules, so occasional misclassification is possible. Results depend heavily on the source resolution, face visibility, reference-model quality, motion blur, lighting, and obstruction.

adaptiveReactorFace_faceSwapV10.zip

adaptiveReactorFace_faceSwapV10.json

Face Swap v1.0

<h1 id="reactor-face-model-compiler-video-or-folder-input">ReActor Face Model Compiler — Video or Folder Input</h1>The ReActor Face Model Compiler is a ComfyUI workflow designed to create reusable ReActor <code>.safetensors</code> 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.<h2 id="two-input-modes">Two Input Modes</h2>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:<ul><li><code>False</code> - Folder Mode</li><li><code>True</code> - Video Mode</li></ul>The unselected input path is skipped.<h2 id="folder-mode">Folder Mode</h2>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:<ul><li>front-facing images</li><li>left-side profiles</li><li>right-side profiles</li><li>smiles and talking expressions</li><li>open-mouth or tongue-visible expressions</li><li>different lighting conditions</li><li>close-up and medium-distance images</li></ul>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.<h2 id="video-mode">Video Mode</h2>Video Mode allows you to create a face model directly from footage without manually extracting every frame first.The workflow samples frames using <code>select_every_nth</code>.Suggested starting values:<ul><li>Every 5th frame - larger dataset and more repeated frames</li><li>Every 10th frame - balanced starting point</li><li>Every 15th frame - smaller and faster dataset</li></ul>For a 30 FPS video:<ul><li>Every 5th frame produces approximately 6 references per second</li><li>Every 10th frame produces approximately 3 references per second</li><li>Every 15th frame produces approximately 2 references per second</li></ul>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:<ul><li>head angles</li><li>facial expressions</li><li>camera distances</li><li>lighting conditions</li><li>left and right profiles</li><li>neutral, smiling, and talking frames</li></ul><h2 id="reference-quality">Reference Quality</h2>The quality of the finished face model depends heavily on the quality and variety of the references.Good references generally have:<ul><li>a clear and reasonably large face</li><li>sharp eyes, nose, and mouth</li><li>limited motion blur</li><li>natural facial proportions</li><li>one dominant person</li><li>varied angles and expressions</li><li>minimal compression or filtering</li></ul>Avoid:<ul><li>duplicates and near-duplicate frames</li><li>heavily blurred images</li><li>very small or distant faces</li><li>strong beauty filters</li><li>sunglasses or severe obstruction</li><li>extreme compression</li><li>images containing multiple equally prominent faces</li><li>images of different people</li></ul>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.<h2 id="recommended-dataset-size">Recommended Dataset Size</h2>Suggested ranges:<ul><li>8–12 images - minimum practical test set</li><li>15–40 images - recommended starting range</li><li>40–75 images - useful for broader variation</li><li>75+ images - only worthwhile when they add genuinely new angles, expressions, or lighting</li></ul>For video input, adjust <code>select_every_nth</code> until the sampled frame count falls into a reasonable range.<h2 id="how-to-use-folder-mode">How to Use Folder Mode</h2><ol><li>Place the reference images into one or more folders.</li><li>Enter the full directory path in each Reference Folder node.</li><li>Keep <code>image_load_cap</code> at <code>0</code> to load every image.</li><li>Set the input selector to <code>False</code>.</li><li>Enter a unique face-model name.</li><li>Queue the workflow.</li><li>Confirm the saved model appears in the ReActor faces directory.</li></ol><h2 id="how-to-use-video-mode">How to Use Video Mode</h2><ol><li>Load the reference video in Optional Video Reference Input.</li><li>Set <code>select_every_nth</code>, normally between <code>5</code> and <code>15</code>.</li><li>Set the input selector to <code>True</code>.</li><li>Enter a unique face-model name.</li><li>Queue the workflow.</li><li>Confirm the saved model appears in the ReActor faces directory.</li></ol><h2 id="output">Output</h2>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.<h2 id="recommended-build-method">Recommended Build Method</h2>Start with:"Mean"Mean generally provides the most stable identity average when the dataset contains varied angles, expressions, lighting, and camera distances.<h2 id="versioning">Versioning</h2>Do not immediately overwrite a model that already works well.Use versioned filenames such as:person_name_v01person_name_v02_more_profilesperson_name_v03_mouth_expressionsperson_name_v04_cleaner_datasetTest each version on the same collection of difficult frames before deciding which one performs best.<h2 id="testing-the-finished-model">Testing the Finished Model</h2>A face model should be tested on more than a single front-facing image.Recommended test cases:<ul><li> neutral front view </li><li> smile </li><li> open mouth </li><li> tongue visible </li><li> left profile </li><li> right profile </li><li> hand close to the face </li><li> partially obstructed face </li><li> low-resolution frame </li><li> bright and dark lighting </li></ul>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.<h2 id="troubleshooting">Troubleshooting</h2><h3 id="only-one-reference-folder-affects-the-model">Only one reference folder affects the model</h3>Confirm the final builder is connected to:"All Reference Images — Folders 1–6"and not directly to a single normalization node.<h3 id="the-identity-looks-weak-or-averaged">The identity looks weak or averaged</h3>Remove:<ul><li> blurred references </li><li> filtered images </li><li> distant faces </li><li> duplicates </li><li> images with inconsistent appearance </li></ul>Add sharper front, profile, and expressive references.<h3 id="profile-swaps-fail">Profile swaps fail</h3>Add clear left- and right-profile images. Slight three-quarter angles are useful, but they do not fully replace true profile references.<h3 id="open-mouth-or-tongue-frames-fail">Open-mouth or tongue frames fail</h3>Add references showing:<ul><li> natural smiles </li><li> talking </li><li> partially open mouths </li><li> widely open mouths </li><li> tongue-visible expressions </li></ul><h3 id="the-wrong-person-appears-in-the-model">The wrong person appears in the model</h3>Remove images containing another prominent face. Group photos and screenshots with multiple similarly sized faces are poor references.<h3 id="video-mode-creates-too-many-frames">Video Mode creates too many frames</h3>Increase <code>select_every_nth</code>.For example:5 → 1010 → 1515 → 20<h3 id="video-mode-produces-repetitive-references">Video Mode produces repetitive references</h3>Use a shorter or more varied source clip, or increase <code>select_every_nth</code>.<h3 id="no-model-file-is-created">No model file is created</h3>Confirm:<ul><li> ReActor is installed correctly </li><li> model saving is enabled </li><li> the filename is valid </li><li> the output directory is writable </li><li> all required models and dependencies are installed </li></ul><h2 id="required-custom-nodes">Required Custom Nodes</h2>This workflow requires:<ul><li>ComfyUI-ReActor </li><li>Video Helper Suite </li><li> the included lazy input selector custom node </li></ul>ComfyUI Manager can be used to identify and install missing nodes.<h2 id="companion-workflow">Companion Workflow</h2>This compiler was created as the companion tool for my adaptive ReActor face-swap workflow.The intended process is:<ol><li> Build the identity with this compiler. </li><li> Load the resulting <code>.safetensors</code> model in the face-swap workflow. </li><li> Process easy frames through ReActor. </li><li> Route difficult frames through ReActor Mask Helper. </li></ol><h2 id="personal-notes">Personal Notes</h2>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.<h2 id="important-notice-(legalmumbojumbo)">Important Notice (LegalMumboJumbo)</h2>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.

ReActor Face Model Compiler - Video or Folder Input

Face Model Compiler v2.1Initial public release of the ReActor Face Model Compiler.This ComfyUI workflow creates reusable ReActor <code>.safetensors</code> face models from either:<ul><li>a directly loaded video</li><li>up to six folders of reference images</li></ul>Main Features<ul><li>Video and folder input modes</li><li>Adjustable video-frame sampling</li><li>Automatic 512 × 512 reference normalization</li><li>Support for up to six organized image folders</li><li>Combined reference batching before model creation</li><li>Direct saving to the ReActor faces directory</li><li>Included setup guide, troubleshooting section, dataset recommendations, and pre-build checklist</li><li>Generic folder paths and filenames for easier redistribution</li></ul>Input Modes<ul><li><code>False</code> - Folder Mode</li><li><code>True</code> - Video Mode</li></ul>Recommended Starting PointFor folder input:<ul><li>use approximately 15-40 clear and varied images</li><li>include front, profile, smiling, talking, and open-mouth references</li><li>remove duplicates, blurry images, strong filters, and images containing other prominent faces</li></ul>For video input:<ul><li>start with <code>select_every_nth</code> between <code>5</code> and <code>15</code></li><li>use a video with varied angles, expressions, distances, and lighting</li><li>avoid extracting hundreds of nearly identical frames</li></ul>OutputFinished face models are saved to:"ComfyUI/models/reactor/faces/"The resulting <code>.safetensors</code> model can then be loaded with the ReActor Load Face Model node.Requirements<ul><li> ComfyUI-ReActor </li><li> Video Helper Suite </li><li> included Lazy Image Selector custom node </li><li> standard ReActor dependencies and models </li></ul>Use only reference material you are authorized to use. Obtain appropriate consent before creating or distributing models based on a real person’s likeness.

Files with SHA256: 2f423b633720908591d29e75e1784897bf52009bf44c1ce9b6af24f262da6fc7