CLIP NegPip+
Allows using negative weights in positive cond with orthogonal decomposition and smooth weight handling. Example: (sky,:-1.05)
Based on the implementation from ppm
Color Palette Extractor
Extracts N dominant colors via MiniBatchKMeans and outputs palette image plus CSV of hex codes.
Image Metadata Extractor
Reads PNG/TIFF info and outputs normalized prompt/settings summary as a single string.
Kuwahara Filter
Fast edge-preserving filter selecting mean color from the minimum-variance quadrant.
L₀ Smoothing Filter
Global edge-aware smoothing that flattens regions while preserving sharp boundaries.
Load LoRA (SDXL Blocks)
Allows setting block weights for an SDXL LoRA. Useful to avoid structural changes but keep the style i.e. only weighting input blocks.
LoRA Extract
Extracts the difference between two models as a LoRA with adjustable rank.
Uses the sd-mecha API
Load Image With Metadata
Loads image with embedded prompts/settings extraction, returns image, mask, and metadata text.
Local Laplacian Filter
Halo-free detail/tone manipulation via Laplacian pyramids with separable Gaussian blurs.
Kohya Lora Config
Parses a LoRA's header and extracts human-readable metadata as JSON.
Output compatible with Kohya's sd-scripts
ORBIT Merge
Direction-aware model merger decomposing deltas into parallel/orthogonal components with independent scaling.
Uses the sd-mecha API
Perturbed Attention Delta
Small edit of PAG with smart sigma-based scheduling.
Quantile Match Scaling
Stabilizes generation by matching frequency-band quantiles to conditional distribution.
SADA Model Acceleration
Skips redundant diffusion steps using trajectory stability analysis for faster sampling.
Based on Stability-guided Adaptive Diffusion Acceleration
Structured Latent
Generate seeded empty latents with various initialization methods (perlin, gaussian, fractal, etc.)
Temperature Adjust
LAB-space white-balance adjustment with HSV saturation compensation, range -1.0…+1.0.
Token Delta Perturbation
Shuffles attention tokens using a scaled delta, with a cosine-decayed perturbation scale.
Based on Token Perturbation Guidance with utilities from ppm
Token Visualizer
Visualizes token influence via 2D wave path with normalized spikes.
Token Sculptor
Strengthens prompt adherence by nudging CLIP embeddings toward soft top-k neighbor blends.
Inspired by Vector Sculptor
Velocity Scaling
Reduces over-brightening in v-prediction models via epsilon scaling adaptation.
Based on Elucidating the Exposure Bias in Diffusion Models