Adept Solver Sampler

This is the sampler the README reaches for first, and for good reason: it's the pack's flagship. It's a multistep predictor-corrector that borrows the ideas that made DPM++ and UniPC popular and stacks them into one pipeline. You get converging, stable output - same seed, same image - which is what you want when you're trying to nail a prompt, not explore a seed.

Mechanically it's an Adams-Bashforth integrator wrapped in a UniPC-style corrector. The predictor step keeps a short history of past derivatives and uses higher-order interpolation to take a better step - order 2 averages the current and previous slopes, order 3 uses two previous points. The use_corrector toggle adds a UniPC-style correction pass that refines the prediction. On top of that, the code cribs from DEIS (an exponential-integrator trick that makes the step accounting cleaner) and DC-Solver (a compensation ratio that adjusts for uneven step spacing). If cond_scale is above 7, it also applies dynamic thresholding to keep extreme latents from running away at high CFG.

The inputs that matter:

• order (1–3) - how much history the predictor uses. Default 2 is the sweet spot the README recommends; 3 can help on long, smooth schedules but adds cost and can get twitchy on short ones.
• use_corrector - on by default. Turn it off to save time and you're basically running a plain multistep solver; most people leave it on.
• use_detail_enhancement, detail_strength, detail_radius - the optional detail-boost wrapper, only active when torchvision is available. Start with it off.

Output is a SAMPLER socket feeding SamplerCustom's sampler input. The README's recipe for v-prediction SDXL models (NoobAI vpred and friends) is AOS-V or AYS-SDXL scheduler plus this sampler at order 2 with the corrector on. It's also the pack's safe default for epsilon models if you want converging output rather than the ancestral line's variety.

Two things to know before you run it. First, it's opinionated about step counts - it's a multistep method, so give it at least 15–20 steps or the history buffer never fills out and you're paying higher-order complexity for Euler-level results. Second, the console prints a 🚀 status line plus warnings if it detects an extreme derivative and clamps it; the clamping is a safety feature, not a crash, but if you see it constantly it usually means your CFG is too high for the model.

Install is the pack-wide one and it's trivial - no dependencies beyond torch, no models to download:

cd ComfyUI/custom_nodes
git clone https://github.com/nawka12/ComfyUI-Adept-Sampler

Restart ComfyUI (or install via ComfyUI Manager by searching "ComfyUI-Adept-Sampler") and it's in the sampling/adept/samplers category. The pack is a port of nawka12's reForge extension, GPL-3.0.

Where it sits: if you're on SDXL and DPM++ 2M Karras has been doing the job, Adept Solver isn't a night-and-day upgrade - but it's a genuinely competent all-rounder that handles high CFG better than the 2M family (the dynamic thresholding kicks in exactly where DPM++ starts to oversaturate). If you're already in this pack for the EQ-VAE stuff, it's the safe converging complement to the ancestral samplers.