Image & Video

I tried to make Vibe Transfer in ComfyUI — looking for feedback

New nodes add NovelAI's signature sliders to IPAdapter, reducing content leakage and enabling granular style control.

Deep Dive

A developer has successfully reverse-engineered and ported NovelAI's popular Vibe Transfer feature to the open-source ComfyUI workflow manager, creating a new tool for Stable Diffusion users. The custom nodes, built on the existing IPAdapter framework, directly address three key limitations of the original IPAdapter: lack of per-image control, content leakage where pose and composition bleed through, and no way to specify what type of information (textures vs. composition) is extracted from reference images. The implementation replicates NovelAI's two-slider interface—'Reference Strength' and 'Information Extracted'—bringing a more intuitive and controlled style transfer experience to the local workflow environment without requiring additional model downloads.

The technical implementation introduces two new nodes: 'VibeTransferRef' for chaining up to 16 reference images with individual controls, and 'VibeTransferApply' which processes them using block-selective injection. This method, inspired by the InstantStyle research paper, injects style information only into specific cross-attention blocks rather than all 44 in SDXL, significantly reducing unwanted content leakage. The system also includes a post-resampler for Information Extracted filtering, using a non-linear sqrt curve to match NovelAI's behavior. Initial single-image tests show the Strength slider functions similarly to IPAdapter's weight parameter, while the Information Extracted slider demonstrably changes what visual information is transferred, from just the 'vibe' at low values to full textures and composition at high values.

Key Points
  • Adds NovelAI's 'Strength' and 'Information Extracted' sliders to ComfyUI/IPAdapter for per-image control
  • Uses block-selective injection (based on InstantStyle paper) to reduce content leakage by targeting only style/composition blocks
  • Enables chaining of up to 16 reference images with individual control over what visual elements are transferred

Why It Matters

Brings commercial-grade, intuitive style transfer controls to open-source workflows, giving artists and developers more precise creative tools.