Replicant Character Lab — Wan2GP Plugin

Beta — usable end-to-end; expect rough edges.

Generation, swaps, posed body double, pose gen, prompt enhance and dataset building all work and have been GPU-run-proven. A full multi-hour LoRA train has not yet been run start-to-finish, so the Train step is experimental — see Status.

🧪 Testers wanted. Install it, build a character, and tell me what breaks — open an issue.

A Wan2GP plugin that ports the character-creator wizard from SupremeDiffusion into Wan2GP's Gradio UI.

What it does

Guides you through creating a reusable character — from name + description to a trained LoRA — as a single branded tab inside Wan2GP. Six steps:

1. Setup — name, description, style, reference image, and positive/negative prompt generation (abliterated Qwen3.5 enhancer)
2. Baseline — generate candidate base images and pick one (or reimagine/skip a supplied reference)
3. Human Clone — optional identity locking via face / body swap
4. Touch Up — optional inpaint + cohesion (img2img normalize) passes
5. Replicate — generate + approve pose variants
6. Train — build the LoRA datasets and train the character LoRA

Save/Load Character and Clear are session actions in the header (top-right), usable at any step.

Install

In Wan2GP, open the Plugins manager and "install from GitHub URL":

https://github.com/saintorphan/Replicant-CharLab-Wan2GP

Then enable Replicant Character Lab in the plugin list and restart WanGP. The installer clones the repo into plugins/, installs requirements.txt (InsightFace + onnxruntime — required for face swap / dataset crops), and the plugin creates its data directories on first run.

Not yet an official Wan2GP plugin — install via the GitHub URL above.

Data layout

Character-Lab data lives under <wan2gp_root>/character_lab/; shared model resources live under <wan2gp_root>/orphansuite/ (cross-plugin via .orphansuite.json). All roots are configurable and persisted from the wizard's OrphanSuite panel:

character_lab/
├── characters/<Name>/   character.json, base.png, poses/
└── datasets/<Name>/     video512/ highres/ full/ face/  (NNN.png + NNN.txt)
orphansuite/             shared across saintorphan plugins
├── models/face/         downloaded swap/enhancer models
├── sdxl_models/         SDXL/Pony/Illustrious checkpoints
└── sdxl_loras/          SDXL-family LoRAs

Models (downloaded from the OrphanSuite → Models panel)

The posed body double and dataset auto-captioning add their own optional models (BiRefNet, IP-Adapter / FaceID, ControlNet OpenPose + annotator, WD14 tagger), also offered in the panel. Nothing downloads without an explicit button press — generation is offline-gated, so a missing model tells you to fetch it here first.

Status

Beta — usable end-to-end. The data layer and wizard UI are complete, and the GPU-bound pipeline — base generation, face/body swap, posed body double (ControlNet OpenPose + FaceID), pose generation, prompt enhancement, and dataset building — is implemented and has been run on-GPU.

The one caveat is training: a full multi-hour LoRA train has not been run start-to-finish yet, so the Train step is experimental. Testers should start with the SDXL / kohya path, which is the most exercised; the Wan2.2 and LTX targets are the least proven and may need tuning.

Design notes

• Single top-level Wan2GP tab; logo banner header.
• 6-step wizard via visibility-toggled groups with a clickable step rail + Back/Next (Gradio 5.29 has no native stepper).
• Reuses portable logic from SupremeDiffusion (character schema, dataset builder/compositor, pose specs); SD/SDXL-family generation runs in-plugin via 🤗 diffusers (core/sd/) — no external checkout.
• Prompt enhancement via Wan2GP's abliterated Qwen3.5 enhancer.