AkashicSolver v2 [EXPERIMENTAL]

Read the node's full name before you judge it: "AkashicSolver v2 [EXPERIMENTAL]". This is the pack's most ambitious and most fiddly sampler, built specifically for EQ-VAE models - the family of VAEs (KohakuBlueleaf's EQ-SDXL-VAE is the one you'll actually see) that adds equivariant transformations to the latent space and gets noticeably better detail and color out of SDXL checkpoints. Those models sample differently from the old VAEs, and this sampler is a toolkit of patches for exactly that.

Under the hood it's an SA-Solver - a stochastic multi-step Adams integrator with a tau function that controls how much stochasticity gets mixed into the deterministic trajectory. On top of that it layers three optional extras: SMEA (a sinusoidal multipass interpolation for high-res coherence), a native detail boost, and "Combat CFG Drift," which corrects the latent mean drift that high CFG causes. That last one is borrowed from ComfyUI-Latent-Modifiers and it does what it says - it pulls the latent back toward zero-mean when CFG has pushed it off, which is the classic cause of color-shift artifacts.

Inputs worth touching, in order:

• tau (0–1, default 0.5) - the SA-Solver stochasticity knob. The README recommends 0.5–0.6 for EQ-VAE models. This is the one that changes the character of the output.
• order (1–3, default 2) - same story as any multistep solver: 2 is the recommended default.
• eta, s_noise - ancestral noise controls. adaptive_eta is on by default here, which is a hint that the author considers auto-phasing the default way to run it.
• smea_strength, ndb_strength - both default to 0, i.e. off. Push smea_strength up for high-res coherency on big images; ndb_strength is a detail boost.
• combat_cfg_drift + combat_drift_intensity - off by default; enable for EQ-VAE at high CFG.
• adaptive_noise, use_detail_enhancement - the pack-wide shared options.

Output is a SAMPLER socket for SamplerCustom. The README's recipe: AkashicAOS, AkashicAOS Alt, or AkashicEQFlow scheduler + this sampler at tau 0.5–0.6, order 2, plus an external rescaleCFG node at 0.7 for EQ-VAE models. That last bit is not optional decoration - the sampler literally prints a warning to the console telling you to use rescaleCFG 0.7. This is the same rescale recipe the v-prediction crowd uses, and it's the difference between great EQ-VAE output and washed-out output.

The EXPERIMENTAL label is earned. Most optional knobs default to 0 or off because they're still being tuned; running with everything on is a good way to stack effects you can't attribute. Start with the README recipe and change one knob at a time.

Install is the pack-wide one, and like the rest of the pack it's pure Python with no requirements.txt and no model downloads:

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

Restart ComfyUI, or use ComfyUI Manager and search "ComfyUI-Adept-Sampler". The console will greet you with a 🌀 line every run.

The honest positioning: if you're not on an EQ-VAE model, you don't need this node - SA-Solver on a normal SDXL is fine but not transformative. If you are on EQ-VAE, this is one of very few samplers actually tuned for it, and it shows. Just respect the rescaleCFG requirement and don't treat it as a general-purpose sampler.