Skip to content

Add Halo-Lipsy: AMD unified memory lip sync node#2514

Merged
ltdrdata merged 1 commit intoComfy-Org:mainfrom
bkpaine1:add-halo-lipsy
Jan 20, 2026
Merged

Add Halo-Lipsy: AMD unified memory lip sync node#2514
ltdrdata merged 1 commit intoComfy-Org:mainfrom
bkpaine1:add-halo-lipsy

Conversation

@bkpaine1
Copy link
Copy Markdown
Contributor

Summary

Adding Halo-Lipsy - a native Wav2Lip lip sync node that finally works on AMD APUs with unified memory.

Why This Node Exists

Existing lip sync nodes fail on AMD unified memory systems because they:

  • Use subprocess.run() which escapes the venv → library mismatches → crash
  • Write temporary files that end up as 0-byte "ghost files"
  • Try to cast vfloat16 tensors directly to NumPy → unified memory conflict
  • Auto-detect CUDA and fail without NVIDIA drivers

Halo-Lipsy fixes all of this:

  • Native Wav2Lip inference in the ComfyUI process (no subprocesses)
  • Face detection forced to CPU (prevents memory fighting)
  • Safe tensor casting: always .float().cpu().numpy()
  • Works with ROCm's CUDA translation layer

Features

  • Sync tuning (frame offset, mel timing multiplier)
  • Edge blending for clean face compositing
  • Adjustable face detection and smoothing
  • Force CPU mode for zero VRAM usage

Tested On

  • Ryzen AI Max+ 395 (Strix Halo) - 64GB unified memory
  • ROCm 7.11
  • ComfyUI with HunyuanVideo (15GB+) loaded simultaneously
  • Also works on NVIDIA/CUDA systems

Repository

https://github.com/bkpaine1/Halo-Lipsy

Credits

  • bkpaine1 - Testing, AMD hardware, bug hunting
  • Claude Code (Anthropic Opus 4.5) - Architecture & implementation
  • Wav2Lip - Original model by Rudrabha Mukhopadhyay

Halo-Lipsy is a native Wav2Lip implementation for ComfyUI that finally works
on AMD APUs with unified memory (Strix Halo, etc.).

Key features:
- No subprocess hacks - runs natively in ComfyUI process
- Face detection on CPU to prevent memory conflicts
- Safe tensor casting for unified memory compatibility
- Sync tuning and edge blending options
- Works on ROCm, AMD APUs, and NVIDIA

Tested on Ryzen AI Max+ 395 (Strix Halo) with 64GB unified memory and ROCm 7.11.

Repository: https://github.com/bkpaine1/Halo-Lipsy
@ltdrdata ltdrdata merged commit ca8c925 into Comfy-Org:main Jan 20, 2026
2 checks passed
@bkpaine1 bkpaine1 deleted the add-halo-lipsy branch January 20, 2026 21:15
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants