examples/README.md

Character Swap with H3 Relay 2.0

Import the canvas workflow, or use the API prompt. Install H3 Relay 2.0. The full model and runtime guide lists every required file.

Download h3_character_swap_v1.5.safetensors. Save it as ComfyUI/models/loras/h3-v2-challenge-eval-20261013/v2-1000.safetensors, matching the workflow and app catalog, or select its actual filename in the LoRA loader. SHA-256: 8f2c3c3eba161df8d0c6099598b4730f641f5ce01e04cd16d2427cff8a35218f. The separate FL2V Turbo8 adapter remains required.

Supply one source shot and a character reference. Prepare Source preserves the full duration when duration is zero, normalizes to 24 fps and aligns the canvas to 32 pixels. Describe the target person to SAM and review the expanded mask. The native reference node takes the character prompt and source performance/audio. The default canvas has no scene anchor; the application exposes optional reviewed anchors and saved masks.

Keep eight steps, normal scheduler, er_sde, CFG 1 and both LoRA strengths at 1. Source video and audio use the same H-1 history alignment. Final video retains the source soundtrack. Review raw decoded frames, identity, occlusion and continuation joins; this is still an experimental model.

Qualified runtime: ComfyUI 79be670e2d9be63e238785af307369d2b9039ed1, frontend 1.53.6, PyTorch 2.11.0 + CUDA 12.8, Linux and RTX 4090. Minimum schema version is ComfyUI 0.37.0. On the tested 24 GB GPU, full residency exhausted memory. Normal offloading with the included Text Stage Memory and Sampling Workspace nodes completed the same generation settings.

The existing V1 workflows and showcases remain unchanged for reproducibility.