A Face-Swap and Identity-Lock Workflow

When you run multiple generation steps without locking a face reference, the AI model treats each node as a stateless request. The face it produces in shot one may share broad features with shot two, but the exact bone structure, skin tone, and eye spacing will subtly shift. Over a sequence of five or more shots this drift accumulates into a character that looks like several different people. The result destroys continuity for any content that needs a recognizable protagonist — ads, short drama, educational explainers, or social media series. Identity lock is not just a cosmetic concern; it is the difference between a usable pipeline and one that requires manual correction after every run.

The workflow follows a three-stage topology. Stage one is the identity extraction node, which receives a single high-quality portrait image and outputs a stable reference artifact — either a face embedding, a processed reference image, or a ControlNet conditioning map depending on the model family you are targeting. Stage two is a fanout: the reference artifact is wired as a shared upstream input to every generation node in the graph. Because Floniks nodes can accept multiple inputs and the editor draws explicit dependency edges, the reference is guaranteed to arrive before any downstream node executes. Stage three is a validation node that runs a similarity score between the reference portrait and each generated output, flagging any frame whose score falls below a configurable threshold. Frames that fail validation are routed to a retry branch rather than passed to the final output collector.

This three-stage structure keeps the reference management logic in a single upstream node. If you swap the reference portrait — for example, to generate a variant campaign with a different talent — you update only the first node's input and re-run the entire pipeline without touching downstream configuration.

In the Floniks workflow editor at /editor, add an Image Input node and connect it to a Face Reference Processor node. The processor node accepts parameters including face crop padding (how much background to retain around the detected face), detection confidence threshold (discard low-confidence detections before they pollute the embedding), and output format (embedding vector, masked crop, or full-resolution crop). Set the output format based on what your downstream generation model accepts — face-conditioned image models typically accept a cropped reference image, while avatar and video models may require an embedding vector.

Wire the processor's output port to the reference image input port on every generation node in the canvas. In the editor, you can hold Shift and click multiple destination ports to draw batch connections from a single source output. Once wired, those edges appear as a distinct color in the canvas, making it visually obvious which nodes are identity-locked and which are free-form. Save this wiring as a reusable template so future campaigns only need a new portrait to launch a full locked sequence.

Identity lock works well when the reference portrait is a clean, front-facing shot with neutral lighting. It degrades when the generated shot requires a steep three-quarter angle, heavy character makeup, or significant occlusion by props or clothing. To handle these cases, build a secondary reference input node that holds an alternate portrait shot at the target angle. Use a conditional branch node — available in the Floniks editor — to detect whether the generation prompt contains angle keywords like "profile", "rear", or "dutch angle", and if so, route the angled reference into the generation node instead of the primary front-facing portrait. This branch-based reference swapping keeps identity fidelity high without requiring a separate template for each camera angle.

For makeup and costume variants, the recommended approach is to retain the underlying face structure reference but reduce its influence weight, allowing the model more latitude to render cosmetic changes while still preserving bone structure and eye spacing. Most face-conditioning models expose a reference strength or influence parameter; values between 0.6 and 0.8 typically balance identity stability with costume flexibility.

Before the workflow routes outputs to storage or downstream editing steps, insert a Face Similarity Scorer node. This node compares each generated frame against the original reference portrait using a lightweight embedding comparison and returns a scalar similarity score between 0 and 1. Configure the node's pass threshold — 0.75 is a reasonable starting point for most use cases — and connect the pass port to the output collector and the fail port to a retry or discard branch. In the retry branch, you can add a prompt augmentation node that appends identity-reinforcing language before sending the frame back for regeneration. Cap retries at two or three to avoid infinite loops, and route persistent failures to a human review queue or a flagged output folder so they are not silently discarded.

Log the similarity scores in the workflow task record so you can analyze which prompt patterns or camera angles produce the lowest fidelity and tune your prompts accordingly over time.

Once the face-swap and identity-lock workflow is validated, save it as a named template in Floniks. The template captures the full node graph, all parameter defaults, and all wiring. When a new campaign starts, open the template, replace the portrait in the Image Input node, update the generation prompts, and run. Because the identity lock logic is embedded in the graph structure rather than in prompt text, new team members do not need to remember to include reference images manually — the workflow enforces it structurally. This is the principal advantage of the node-based editor over single-prompt approaches: constraints become architectural rather than procedural.