diff --git a/LICENSE b/LICENSE
new file mode 100644
index 0000000..2f1cb60
--- /dev/null
+++ b/LICENSE
@@ -0,0 +1,27 @@
+MIT License
+
+Copyright (c) 2026 Akash (akashzeno)
+
+Portions of this software derive from "colored-noise-sampling"
+(https://github.com/hadardavidson/colored-noise-sampling), Copyright (c) 2026
+Hadar Davidson, Roy Issachar, and Sagie Benaim, also under the MIT License:
+- the bundled gamma matrices in gamma_matrices/ (verbatim), and
+- the frequency-domain noise-shaping algorithm reimplemented in src/engine/.
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.
diff --git a/README.md b/README.md
new file mode 100644
index 0000000..55a2ce0
--- /dev/null
+++ b/README.md
@@ -0,0 +1,122 @@
+# ComfyUI-ColoredNoiseSampling
+
+Frequency-shaped (**colored**) noise for ComfyUI's stochastic diffusion sampling — a clean,
+native port of the *transferable mechanism* behind
+[Colored Noise Diffusion Sampling (CNS)](https://github.com/hadardavidson/colored-noise-sampling)
+(Davidson, Issachar & Benaim, 2026, arXiv:2605.30332).
+
+Instead of injecting plain white Gaussian noise at each stochastic step, this pack injects noise
+whose **power spectrum is shaped** — and whose shape can **vary across the sampling trajectory**
+(broadband early → high-frequency detail late). It plugs into ComfyUI's standard
+`noise_sampler` seam, so it works with the existing custom-sampling graph and the all-in-one node.
+
+> **Honest scope.** The CNS *technique* ports cleanly and is genuinely useful as a quality/character
+> knob. The paper's quantitative results (FID 6.27 vs 8.26) are specific to **SiT on ImageNet-256**
+> and will **not** reproduce on SD/SDXL/Flux. The optional γ-matrix mode ships the paper's original
+> matrices, but on non-SiT models they are a documented heuristic, not a calibrated schedule. For
+> general use, prefer the **parametric** mode.
+
+---
+
+## What it does (in one paragraph)
+
+At each stochastic step it draws white noise, FFTs it, multiplies the spectrum by a
+radially-symmetric per-frequency profile, inverse-FFTs, and **renormalizes to unit variance** —
+then hands that to the sampler, which applies its own per-step amplitude (`sigma_up · s_noise`).
+The per-frequency profile is either a **power-law** `amplitude(f) ∝ f^(−α/2)` (α: 0 white, +1 pink,
++2 brown/red, −1 blue, −2 violet) with a time-varying exponent, or a **γ-matrix residual**
+`(1 − γ(progress, f))` reproducing the paper's algorithm.
+
+## Nodes
+
+All three live under the **`sampling/colored_noise`** category.
+
+### 1. Colored Noise Sampler — `ColoredNoise_SamplerSelect`  → `SAMPLER`
+The core node. Outputs a `SAMPLER` for `SamplerCustom` / `SamplerCustomAdvanced`. Wraps a chosen
+**stochastic** base sampler and colors its per-step noise.
+
+Key inputs: `base_sampler` (all stochastic ancestral/SDE samplers, incl. the `_RF` rectified-flow
+variants for Flux/SD3), `eta`, `s_noise`, `mode` (`parametric` | `gamma_matrix`),
+`alpha_start`/`alpha_end` + `interpolation` (the time-varying spectral tilt), the γ-matrix controls,
+and `energy_scale`.
+
+### 2. Colored Noise (Initial) — `ColoredNoise_Noise`  → `NOISE`
+Outputs a `NOISE` object producing **colored initial latent noise** (a single constant color:
+white/pink/brown/blue/violet or a custom α). Use it on the `noise` input of `SamplerCustomAdvanced`.
+Honors `batch_index` determinism and nested (video) latents.
+
+### 3. Colored Noise KSampler — `ColoredNoise_KSampler`  → `LATENT`
+All-in-one (model + conditioning + latent → latent), for when you don't want to wire the custom
+sampling graph. Picks a stochastic base sampler + scheduler + steps + cfg + the coloring controls,
+plus an optional `color_initial_noise` toggle.
+
+## Two modes
+
+| Mode | What it does | When to use |
+|---|---|---|
+| **parametric** (default) | Power-law spectrum `f^(−α/2)`, α interpolated `alpha_start → alpha_end` across the trajectory (linear or exponential). Model-agnostic. | Everything. Start here. |
+| **gamma_matrix** | Loads a `[steps, bins]` γ matrix and injects noise into the *unresolved* bands per the paper. Bundled SiT/ImageNet matrices included. | Reproducing/experimenting with the paper; SiT-like setups. |
+
+### Quick parametric recipes
+- **White** (baseline / sanity): `alpha_start = 0`, `alpha_end = 0`.
+- **CNS-like** (broadband → high-freq late): `alpha_start = 0`, `alpha_end = -1` … `-2`.
+- **Constant pink/brown** (softer, low-freq emphasis): `alpha_start = alpha_end = 1` … `2`.
+
+`energy_scale` (default `1.0`) scales the noise std *after* unit-variance renorm — a deliberate
+"heat" knob; values ≠ 1 intentionally change the effective noise level.
+
+## γ-matrix files
+
+A model folder **`colored_noise_gamma`** is registered automatically. The two bundled matrices
+(`gamma_matrix_scaled.pt`, `gamma_matrix_scaled_cfg_1.5.pt`) appear in the dropdown. To add your
+own, drop a raw `[steps, bins]` float32 tensor (`torch.save`, `.pt`/`.pth`) into
+`models/colored_noise_gamma/`. *(New files may require a ComfyUI restart to appear.)* See
+[`gamma_matrices/SOURCE.md`](gamma_matrices/SOURCE.md) for provenance.
+
+## Compatibility
+
+- **Attention backends** — fully compatible with SageAttention 2/3 and FlashAttention 2/3 (and
+  xformers / pytorch / split / sub-quad) **by construction**. This pack only shapes the stochastic
+  noise term; it never touches the model forward pass or `transformer_options`, where attention is
+  selected. Your attention choice is untouched.
+- **Memory** — uses ComfyUI's native memory management only. Noise/FFT tensors are transient,
+  latent-sized, hold no model references, and do not impede offloading.
+- **Precision** — the FFT runs in float32 with a CPU fallback (half-precision FFT is unsupported on
+  most backends), then casts back; output matches the latent dtype.
+- **Latent ranks** — 4D image `[B,C,H,W]` and 5D video `[B,C,T,H,W]` get a 2D spatial FFT; 3D audio
+  `[B,C,L]` gets a 1D time FFT; nested (video) latents are colored per sub-tensor.
+- **Samplers** — only **stochastic** samplers are offered (deterministic ones never inject per-step
+  noise, so coloring them would be a silent no-op).
+
+## Install
+
+Clone into `ComfyUI/custom_nodes/` and restart ComfyUI. No extra dependencies (torch comes with
+ComfyUI). The nodes appear under `sampling/colored_noise`.
+
+## Usage
+
+**All-in-one:** `Load Checkpoint → Colored Noise KSampler → VAE Decode`. Set `base_sampler`
+(e.g. `dpmpp_2m_sde`), `mode = parametric`, `alpha_start = 0`, `alpha_end = -1`. Render.
+
+**Custom-sampling graph:**
+```
+Colored Noise (Initial) ─┐
+Colored Noise Sampler ───┤
+BasicScheduler ──────────┼─► SamplerCustomAdvanced ─► VAE Decode
+BasicGuider ─────────────┘
+```
+
+## Development / tests
+
+Pure-engine unit tests (no model weights), run in the `comfyenv` environment:
+```
+cd custom_nodes/ComfyUI-ColoredNoiseSampling
+python -m pytest tests/ -q          # 26 tests: spectral shaping, schedule, batch determinism, ...
+ruff check .
+```
+
+## Credits & license
+
+Method and bundled γ matrices: **Hadar Davidson, Roy Issachar, Sagie Benaim**,
+[colored-noise-sampling](https://github.com/hadardavidson/colored-noise-sampling) (MIT),
+arXiv:2605.30332. This pack is MIT-licensed; see [LICENSE](LICENSE).
diff --git a/gamma_matrices/SOURCE.md b/gamma_matrices/SOURCE.md
new file mode 100644
index 0000000..5e3897e
--- /dev/null
+++ b/gamma_matrices/SOURCE.md
@@ -0,0 +1,19 @@
+# Bundled gamma matrices — provenance
+
+`gamma_matrix_scaled.pt` and `gamma_matrix_scaled_cfg_1.5.pt` are copied verbatim from
+[hadardavidson/colored-noise-sampling](https://github.com/hadardavidson/colored-noise-sampling)
+(MIT License), directory `gamma_matrix/`.
+
+Each is a float32 tensor of shape **[250 steps, 32 radial frequency bins]**, values in `[0, 1]`,
+encoding empirical per-band *structural completion* `γ(step, freq)` for **SiT on ImageNet-256**.
+
+These values are **model- and dataset-specific**. The *shape* of the schedule (low frequencies
+resolve early, high frequencies late) is a general property of diffusion models, so the matrices
+behave sensibly as a heuristic on other models — but they are **not calibrated** for SD/SDXL/Flux
+and will not reproduce the paper's FID figures there. Prefer the **parametric** mode for general
+use; use these for experimentation or SiT-like setups.
+
+Paper: Davidson, Issachar & Benaim (2026), *Colored Noise Diffusion Sampling*, arXiv:2605.30332.
+
+To add your own matrices: drop a raw `[steps, bins]` float32 tensor saved with `torch.save`
+(`.pt`/`.pth`) into `models/colored_noise_gamma/`; it will appear in the node's dropdown.