Head-swap models force a choice: facial fidelity or proportional accuracy
A ComfyUI practitioner reports that existing head-swap models force a choice between preserving facial recognition and matching target head proportions—commercial tools excel at faces but ignore scaling, while open-source checkpoints like BFS handle geometry but weaken identity.
A ComfyUI user building a head-swap pipeline has run into a stubborn trade-off: commercial models preserve facial features and keep the source person recognizable, but ignore target head proportions. Open-source checkpoints like the BFS (Best Face Swap) collection on HuggingFace resize heads correctly but weaken facial expressions and identity strength. After testing every checkpoint in the Alissonerdx BFS repository paired with Qwen, the user confirms that head-size matching works well—but facial fidelity drops off.
The workflow requirements are concrete: strong facial feature preservation so the source person remains recognizable, head resizing to match the target image's proportions, and optionally some color or style adaptation so the swapped head doesn't look pasted on. Facial emotion transfer and head-position alignment are nice-to-haves. The user is open to large model sizes if they deliver on the first two priorities.
No single model or workflow has solved the problem yet. A hybrid approach—running a geometry-aware swap first, then a face-restoration pass with a strong identity model—might close the gap, but no packaged ComfyUI workflow for that sequence has surfaced. The next step is likely a custom node that chains BFS output into a face-enhancer like CodeFormer or GFPGAN with identity loss tuned high, or a new checkpoint trained explicitly on head-proportion datasets. Until then, practitioners are stuck choosing which priority to drop.