ComfyUI Extension: ComfyUI-geltz

Adaptive Refined Exponential Solver (ares)

Deterministic variation of the res_multistep sampler. Batches σ, auto-converts model outputs between ε/x₀/v, computes Δt, and applies a momentum aware Heun step to advance the latent and estimate x₀.

Clamps σ to [sigma_min, sigma_max], falling back to Euler when unavailable or with <2 sigmas, iterating _ARES_STEP across the schedule. Registers as ares within KSampler's sampler selection.

---

Cosine-Uniform Scheduler (csu)

Inspired from sgm_uniform. Computes a cosine-eased sigma schedule: it maps uniform u∈[0,1] through w=((1−cos(πu))/2)^γ to timesteps, converts to sigmas, enforces strict decrease, caps the first at σ_max, and ends with 0.

Flushes sigmas after generating to prevent NaNs. Registers as csu within KSampler's scheduler selection.

---

Dithered Isotropic Latent (dil)

Improves empty latents. Adds two latent initializers, DIL_EmptyLatent and DIL2_EmptyLatent, that start from noise and iteratively ascend a differentiable score based on edges, high-frequency energy, kurtosis, and orientation coherence, with normalization and dithering to return a LATENT.

The spectral variant adds per-channel seeds and frequency-domain shaping via beta and spectral_mix, and the file includes the gradient, blur, FFT, dtype/device, and node-registration utilities.

---

Quantile Match Scaling (qms)

Precise rescaling of CFG to prevent oversaturation. Does not affect original structure. Hooks pre-CFG and rescales the guidance g = cond − uncond by matching low, mid, and high frequency quantiles to the conditional.

Adapts cutoffs and quantiles each step, fits per-band linear maps with EMA clamps and a CFG-dependent rescale, applies them in FFT space, and returns cond_new = uncond + g_scaled.

---

Regional Split Sampler (rss)

Allows prompting by two regions, more can be defined in file. Splits the image width into left and right regions using a soft mask (center, feather), applies separate positive conditionings to each side, then calls nodes.common_ksampler with the chosen sampler/scheduler to generate a LATENT.

Registers as Regional Split Sampler with inputs for model, seed, steps, cfg, denoise, center, and feather.

---

Sigma-Weighted Shuffle (sws)

Improves image consistency. Hooks attention, derives a progress variable u from σ or step, scales Q by a temperature τ(u), and uses local Gaussian Sinkhorn transport to stochastically shuffle K and V while keeping the attention distribution within a KL cap.

Mixing is gated by attention entropy and u, uses EMA-smoothed transport, binary-searches α for K, schedules α for V, infers H×W from sequence length, installs via set_model_attn2_patch, and registers the node as SWS.

---

tokenteller

Useful to detect "prompt bleed". Parses conditioning to gather up to limit_streams token embeddings, derives a per-token value by norm/var/mean normalized to [0,1], and assigns word labels from prompt-like fields or indices.

Renders a 2D wave path displaced by those values into spikes, rasterizes a colored viridis-like curve and bitmap text plus a left list of word value pairs, and outputs a single IMAGE tensor.

---

vectorpusher

Improve adherence to prompts. Adds a conditioning node vectorpusher that tokenizes the prompt and, for each CLIP token, nudges its embedding toward a soft top-k neighbor blend using an entropy and attention-scaled trust-region step with a KL bound and angle cap.

Inspired from Vector Sculptor by Extraltodeus.