# ⚡ [<a target="_blank" rel="ugc" href="http://ZEUS.ONL">ZEUS.ONL</a>] Jigsaw Core NodesAn uncompromised, high-performance custom node infrastructure for ComfyUI (v0.28.0+). Built specifically to handle modern large Diffusion Transformer (DiT) architectures, eliminate VRAM-bottlenecks on mid-tier GPUs like the RTX 4070, and unlock unprecedented texture realism without destructive Turbo-LoRA quality degradation.Official Network: [https://zeus.onl](https://zeus.onl)<hr />## 🍳 The Core Arsenal (Module Overview)### 1. 📡 Jigsaw Adaptive Auto-Sigma Weaver <code>jigsaw_auto_sigma.py</code>)An intelligent, non-destructive Dynamic Frequency &amp; Noise Estimator operating entirely on the latent level.* How it works: Hooks directly into the model inference loop and fires a live 2D Fast Fourier Transformation <code>torch.fft.fft2</code>) to measure microscopic noise frequency peaks in real-time.* The fix: Isolates ultra-high frequency zones to dynamically damp and compress remaining sigmas. It completely eliminates mathematical grid/tile artifacts common in large Qwen/Wan backbones and fixes harsh "desert skin" textures at extreme resolutions, "matte-powdering" skin while keeping fabrics and backgrounds razor-sharp.### 2. 📐 Jigsaw LoRA Conflict Harmonizer <code>jigsaw_lora_harmonizer.py</code>)A mathematical matrix savior node that prevents identity distortion when stacking multi-purpose LoRAs.* How it works: Materializes full delta matrices in RAM and processes them through an advanced Gram-Schmidt Orthogonalization algorithm <code>_orthogonalize</code>).* The fix: Cleans overlapping vector weights so they never cancel each other out or cause unnatural contrast burns. Forces your character LoRAs to maintain 100% geometric facial DNA accuracy even when operating alongside heavy compression weights. Includes an automated <code>harmonizer_cache</code> folder setup to prevent VRAM spikes during consecutive iterations.### 3. 🔮 Universal VAE Merge v2 <code>UniversalVAEMerge.py</code>)A highly optimized, layer-agnostic VAE fusion node engineered to combine the strengths of different generative spaces.* How it works: Abandons old, rigid SDXL block maps and operates purely on the dynamic intersection of state-dict keys <code>sd_a.keys() &amp; sd_b.keys()</code>).* The fix: Flawlessly merges cutting-edge 3D-causal video VAEs (like WAN 2.1 FP32) with architectural base model VAEs without causing color-space collapse. Features dedicated canvas bias injection fields to pump crystal-clear, analog contrast directly into the final layer weights.### 4. 🏎️ Jigsaw TCD Vector Accelerator <code>jigsaw_tcd_accelerator.py</code>)⚠️ STATUS: EXPERIMENTAL ALPHA CORE ⚠️* Description: An architectural trajectory core designed to simulate lightning-fast step reductions using procedural Flow-Matching time-stretching.* **Current Status:** NOT FUNCTIONAL / UNDER DEVELOPMENT. This module is currently undergoing heavy core-level rewriting. The mathematical time-dependent boundaries are causing vector clipping under specific DiT sampling conditions (resulting in mosaic noise or gray veils). Use for alpha-testing only; do not include in stable low-step workflows yet.---## 🔌 Wiring &amp; Interface Guide### Universal VAE Setup* Connect your base model VAE into <code>vae_a</code>.* Connect your high-fidelity uncompressed video VAE (e.g., WAN 2.1 FP32) into <code>vae_b</code>.* Set <code>merge_mode</code> to <code>weighted_sum</code>, <code>alpha</code> to <code>0.50</code>, and inject <code>0.15</code> into <code>contrast</code> to burn away the remaining mid-step fog.* Route the final output <code>vae</code> straight into your terminal <code>VAE Decode</code> node.### Harmonizer Setup* Loop your sequential standard LoRA loaders (e.g., character, style, clothing) as usual.* Intercept the final <code>MODEL</code> wire right after your last standard LoRA loader and route it through the <code>JigsawLoraConflictHarmonizer</code>.* Select <code>orthogonal</code> mode to trigger the Gram-Schmidt calculation or <code>priority_last</code> to lock character priority. ⚠️ *CRITICAL SIGNAL FLOW RULE:** If your workflow utilizes a separate Turbo-LoRA (such as Hyper, Lightning, or TCD weights), you MUST connect it SEPARATELY BEHIND the Harmonizer node (between the Harmonizer output and the KSampler input). Loading the Turbo-LoRA before the Harmonizer will cause the Gram-Schmidt orthogonalization to strip away the velocity vectors, destroying the low-step acceleration trajectory!---## 🧪 Installation &amp; System Requirements1. Clone this repository directly into your ComfyUI custom nodes directory: ```bash cd ComfyUI/custom_nodes git clone <a target="_blank" rel="ugc" href="https://github.com">https://github.com</a> ```2. Ensure your embedded environment is up to date with Python 3.13+ compatible deep learning modules: ```bash python -m pip install -U torch huggingface-hub transformers ```---Here's a concise summary you can drop into a changelog or PR description:---Jigsaw Adaptive Sigma Weaver — v13 changes1. Fixed: sigma range was silently discardedThe node previously ignored the actual magnitude of the incoming <code>sigmas</code> tensor — it only read the step count, then generated a synthetic curve always scaled to roughly <code>1.25 → 0.0</code>, regardless of what sigma range the sampler/model actually used (e.g. <code>14.6 → 0.0</code>). This has been fixed: the real <code>sigma_maxsigma_min</code> endpoints are now preserved, and only the shape of the curve between them is reshaped via the flow-matching mu-shift. This also means the node no longer silently breaks partial-denoise / non-full-strength workflows.2. New: the live FFT damping described in the docs is now actually implementedThe node's description always claimed to hook the inference loop and use a live 2D FFT to detect and damp high-frequency artifacts — the shipped code never did this. v13 adds a real implementation:- Hooks the diffusion model's forward pass via ComfyUI's standard composable <code>add_wrapper_with_key</code> API (same mechanism used by well-behaved nodes elsewhere in the ecosystem — not the single-slot <code>model_options["model_function_wrapper"]</code>, which silently clobbers other nodes using the same slot).- At each sampling step, runs <code>torch.fft.fft2</code> on the model's output, measures how much energy sits above a configurable frequency cutoff relative to what a flat spectrum would predict there, and damps only that excess — legitimate sharp detail and normal noise characteristics are left untouched.- Verified on synthetic test data: pure white noise (flat spectrum) is correctly left alone; a simulated periodic grid/tile artifact has its excess high-frequency energy reduced by &gt;90%, while low-frequency image structure stays at ~100% correlation (untouched).- Gated to only activate from ~15–50% into the sampling schedule (progress-based smoothstep) so early, coarse-structure steps aren't touched.- Two new user-facing parameters: <code>hf_damping_strength</code> (0 = off, default 1.0) and <code>hf_cutoff</code> (default 0.62), plus an optional <code>debug</code> toggle that prints per-step measurements.Backward compatible: setting <code>hf_damping_strength</code> to 0 restores v12's sigma-only behavior (now with the scaling fix applied).---## 🪐 Credits &amp; License* Sponsorship &amp; Network Branding: Powered by [<a target="_blank" rel="ugc" href="http://ZEUS.ONL">ZEUS.ONL</a>](<a target="_blank" rel="ugc" href="https://zeus.onl">https://zeus.onl</a>)* Core Engineering: Jigsaw, Zeus &amp; AI Collaboration Units* Target Environment: Tailored for automated 9.3MP uncompromised render setups.

Files with SHA256: 793b28664b923127052122bb99b0f329b9f5d0465610129d4ba0b617d2d14659