ComfyUI nodes for advanced prompt weighting control
Top 73.8% on sourcepulse
This repository provides advanced ComfyUI nodes for fine-grained control over CLIP text prompt weighting and interpretation, targeting users who need to precisely influence image generation based on textual prompts. It offers multiple methods for normalizing token weights and interpreting prompt weighting schemes, enabling more nuanced and predictable results compared to standard ComfyUI implementations.
How It Works
The core innovation lies in the BNK_CLIPTextEncodeAdvanced
node, which introduces token_normalization
(none, mean, length, length+mean) and weight_interpretation
(comfy, A1111, compel, comfy++, down_weight) parameters. These allow users to modify how prompt weights affect CLIP embeddings, moving beyond simple linear interpolation. The length
normalization, for instance, distributes weights across tokens of longer words to maintain consistent magnitude changes. The weight_interpretation
options mimic or extend established methods like Automatic1111's scaling and Compel's masked embedding mixing for down-weighting, offering greater control over concept emphasis and de-emphasis.
Quick Start & Requirements
custom_nodes
directory.Highlighted Details
BNK_CLIPTextEncodeSDXLAdvanced
with cross-CLIP model balancing and BNK_AddCLIPSDXLParams
for detailed conditioning.Maintenance & Community
Licensing & Compatibility
Limitations & Caveats
The Mix Clip Embeddings
node is marked as deprecated, with its functionality merged into core ComfyUI nodes. The SDXL-specific BNK_AddCLIPSDXLRefinerParams
node is primarily for experimentation.
1 year ago
1 day