nawka12andClaude Sonnet 4.6
b8cf2fccb3
All schedulers now work correctly on both EDM-style (SDXL, sigma_max > 1) and flow-matching models (Cosmos/Anima, sigma_max = 1.0): - Flow-model fix: Karras rho=7 collapses at sigma_max=1 (max_inv_rho=1), causing the first step to drop to ~0.5 and producing white images. All Karras and log-SNR schedulers now fall back to linear sigma-space interpolation when sigma_max <= 1.0, giving gradual first steps (~0.9+). Affected: AOS-V, AOS-e, AkashicAOS, AkashicAOS Alt, Entropic, SNR-Optimized, LogSNR-Uniform, Constant-Rate, Adaptive-Optimized, Cosine, Tanh Mid-Boost, Exponential Tail, Jittered-Karras, Stochastic, JYS, Hybrid JYS-Karras. - AOS-V / AOS-e: fix duplicate sigmas at linspace phase boundaries (both flow and standard paths) — division-by-zero in er_sde stage 2. - Stochastic: add monotonicity cleanup after sort to handle edge case where random perturbation produces equal consecutive sigmas. - JYS: fix length bug — returned steps+2 elements instead of steps+1. _compute_jys_timesteps now returns exactly num_steps values (no trailing 0 sentinel). Hybrid JYS-Karras shape mismatch crash fixed. - AYS-SDXL: fix length bug for exact step counts (10/15/20/25/30) — returned steps elements instead of steps+1. Now appends terminal 0 separately so users get the requested number of denoising steps. Interpolation path ported from numpy to pure torch. - adept_samplers.py: move inline imports (get_phase_correction, compute_adaptive_noise_scale, compute_binned_corrections) to top-level. Bump version: 1.4.0 → 1.5.0 Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
ComfyUI-Adept-Sampler
Advanced custom samplers and schedulers for ComfyUI, ported from the Stable Diffusion WebUI reForge extension.
Also available for SD WebUI reForge: nawka12/adept-sampler
Features
Samplers (4)
| Sampler | Description |
|---|---|
| Adept Solver | Hybrid Predictor-Corrector pipeline using Adams-Bashforth integration (DEIS) with UniPC correction steps and dynamic thresholding |
| Adept Ancestral Solver | Enhanced Euler Ancestral with phase-dependent step sizing, adaptive noise injection (Eta), context-aware derivative corrections, and adaptive noise scale |
| AkashicSolver v2 | SA-Solver (Stochastic Adams) implementation integrated with SMEA (Sinusoidal Multipass) interpolation for high-res coherence, Combat CFG Drift, and adaptive noise scale. Use external rescaleCFG (0.7) for EQ-VAE models |
| Mirror Correction Euler | Euler Ancestral with semantic reflection probe. Uses a 3-call Heun correction (x_probe = 2·D(x) − x) in the first correction_phase fraction of steps for improved curvature estimation. Probe norm limiting for stability. Optional smooth phase decay and adaptive noise scale |
CFG Fix Nodes (1)
| Node | Description |
|---|---|
| Adept Spectral Modulation (CFG) | Patches the model's CFG function to apply Clybius Spectral Modulation (frequency-domain correction) to the noise prediction. Connect MODEL → MODEL before your sampler. Based on ComfyUI-Latent-Modifiers |
Schedulers (19+)
| Category | Schedulers |
|---|---|
| Anime-Optimized | AOS-V (v-prediction), AOS-ε (epsilon) |
| EQ-VAE / Akashic | AkashicAOS (Continuous Power-Function), AkashicAOS Alt (stronger detail bias), AkashicEQFlow (pure CDF-based crossover-focused log-SNR) |
| Research-Based | AYS-SDXL (Align Your Steps), JYS (Jump Your Steps), SNR-Optimized |
| General Purpose | Entropic, Cosine-Annealed, LogSNR-Uniform, Constant-Rate, Adaptive-Optimized |
| Experimental | Stochastic, Jittered-Karras, Hybrid JYS-Karras, Tanh Mid-Boost, Exponential Tail |
Installation
-
Navigate to your ComfyUI custom nodes folder:
cd ComfyUI/custom_nodes/ -
Clone or copy this repository:
git clone https://github.com/nawka12/ComfyUI-Adept-Sampler.git # OR copy the folder directly -
Restart ComfyUI
Usage
Using Schedulers
Connect a scheduler node to generate custom sigma schedules:
[Load Checkpoint] → [Adept Scheduler (AOS-V)] → [SamplerCustom] → [VAE Decode]
↑
[Sampler Node]
Using Samplers
Connect a sampler node to use custom sampling algorithms:
[Load Checkpoint] → [BasicScheduler] → [SamplerCustom] → [VAE Decode]
↑
[Adept Solver Sampler]
Using Spectral Modulation
Patch the model before passing it to SamplerCustom:
[Load Checkpoint] → [Adept Spectral Modulation (CFG)] → [SamplerCustom] → [VAE Decode]
↑
[Sampler Node]
Node Reference
Scheduler Nodes
All scheduler nodes take a MODEL input and output SIGMAS.
| Node | Parameters |
|---|---|
| Adept Scheduler (AOS-V) | steps |
| Adept Scheduler (AOS-ε) | steps |
| Adept Scheduler (AkashicAOS) | steps |
| Adept Scheduler (AkashicAOS Alt) | steps |
| Adept Scheduler (AkashicEQFlow) | steps |
| Adept Scheduler (Entropic) | steps, power |
| Adept Scheduler (JYS) | steps |
| Adept Scheduler (AYS-SDXL) | steps |
| Adept Scheduler (Stochastic) | steps, noise_type, noise_scale, base_schedule |
| Adept Scheduler (Advanced) | steps, scheduler (dropdown), entropic_power |
Sampler Nodes
All sampler nodes output SAMPLER for use with SamplerCustom.
| Node | Key Parameters |
|---|---|
| Adept Solver Sampler | order (1-3), use_corrector, detail enhancement options |
| Adept Ancestral Sampler | eta, s_noise, adaptive_eta, phase_noise, enhanced_derivative, adaptive_noise |
| AkashicSolver v2 | tau (0-1), eta, s_noise, order, adaptive_eta, smea_strength, ndb_strength, combat_cfg_drift, combat_drift_intensity, adaptive_noise |
| Mirror Correction Euler Sampler | eta, s_noise, correction_phase (0-1), smooth_phase, adaptive_noise |
CFG Fix Nodes
| Node | Key Parameters | Input → Output |
|---|---|---|
| Adept Spectral Modulation (CFG) | strength (0-2), percentile (1-15) | MODEL → MODEL |
Recommended Settings
For v-prediction models (e.g., SDXL)
- Scheduler: AOS-V or AYS-SDXL
- Sampler: Adept Solver (order=2, corrector=on)
For epsilon-prediction models
- Scheduler: AOS-ε
- Sampler: Adept Ancestral (eta=1.0, adaptive_eta=on)
For EQ-VAE models (e.g., AkashicPulse)
- Scheduler: AkashicAOS, AkashicAOS Alt, or AkashicEQFlow
- Sampler: AkashicSolver v2 (tau=0.5-0.6, order=2) + external rescaleCFG node (0.7)
- Optional: add Adept Spectral Modulation and/or enable Combat CFG Drift
Mirror Correction Euler
- Scheduler: Any (Karras, AOS-ε, AkashicAOS)
- correction_phase=0.3–0.5 is a good starting point
- Enable smooth_phase for EQ-VAE smooth latents
Adaptive Noise Scale
Available on Mirror Correction Euler, Adept Ancestral, and AkashicSolver. When enabled:
- Collects noise excess metrics during a warmup phase (texture sigma region 0.5–5.0)
- After calibration, restarts generation with a per-phase correction factor applied
- Useful when the default s_noise is over- or under-shooting for a given model
Credits
Original reForge extension and algorithms developed for Stable Diffusion WebUI. Ported to ComfyUI as custom nodes.
License
This project is licensed under the GNU General Public License v3.0 (GPL-3.0).
License Summary
Permitted: Commercial use, Modification, Distribution, Patent use, Private use
Requirements: License and copyright notice, State changes, Disclose source, Same license
Limitations: No Liability or Warranty
Why GPL-3.0?
This license ensures compatibility with Stable Diffusion WebUI reForge and its ecosystem, while protecting the open-source nature of the project.
Full license text: https://www.gnu.org/licenses/gpl-3.0.html