ComfyUI user finds refmod nodes lose clothing details vs plain image references
spacemidget75 · reddit · 2026-10-28
A ComfyUI user shares first-hand comparisons suggesting refmod nodes (Adudeguyman's, which bake a reference image into the safetensor) actually perform worse than plain image references: across several videos, clothing details went missing and faces drifted.
Swapping back to separate clothing and face image references instantly restored fidelity — belt buckles and buttons now match the reference. The poster isn't sure why refmods underperform and asks whether others see the same, a useful counter-intuitive data point for character-consistency workflows.
More from Multimodal
- Developer turns a few Airbnb photos into a full 3D walkthrough with Claude — CodeByPoonam · 2026-10-28
- MOSS-Image character LoRA trained on RX 7900 XT 20GB in 3.5 hours on native Windows — iBorgSimmer · 2026-10-28
- AI-generated Harry Kane Ballon d'Or photos look convincingly real — egeberkina · 2026-10-28
- 10M-Gaussian splat stuns on SuperSplat ahead of Blender 5.3 support — willeastcott · 2026-10-28
- Open-H3-Regenerate-2K: open 3-step diffusion refines video drafts up to native 4K, 5x faster than the API — linoy_tsaban · 2026-10-28
- Sub-10MB 3D Gaussian Splat of a Moth Shows Every Hair, Powered by WebGPU — willeastcott · 2026-10-28