DDRK Omega Unified KSampler (advanced)

If you've been around ComfyUI long enough to have opinions, you know the sampler rules flipped on you. DPM++ 2M Karras was the safe default on SD 1.5/SDXL, and then flow-matching models arrived and Karras went from safe to actively wrong on them. Suddenly you're running one Euler workflow for Flux and another sampler entirely for SDXL, relearning parameters every time you swap a checkpoint. DDRK Omega Unified KSampler is this pack's answer to exactly that whiplash: a drop-in replacement for the stock KSampler that figures out which model family it's talking to and recalibrates itself.

What it actually does

The node looks at your model's sigma schedule - sigma_max above ~5 means EDM (SDXL, SD 1.5, SD 2), otherwise flow matching (Flux, SD3, Qwen, Krea, HiDream, Chroma, Lumina) - and routes sampling through three phases, each with its own integrator policy and post-processing. The headline feature is the HC2 integrator, an exponential multistep method in the DPM-Solver++/UniPC family: second-order accuracy at one model call per step, which is roughly Heun quality at half the compute on flow-matching models. There's also ancestral SDE noise (FM only), Karras-style churn (EDM only), SABER spatial stabilization, and a final-step perceptual sharpen. The author measured all of this and, refreshingly, published the negative results too: HC2 shows no advantage on EDM, where auto or heun stays the sensible pick.

The inputs that matter

Beyond the usual model, positive, negative, latent_image, seed, steps and cfg, you mostly touch three things:

• integrator - auto, hc2, rk4, heun, euler. Leave it on auto unless you specifically want HC2.
• smart_defaults - auto-detects the family and sets scheduler, shift, integrator and enhancer levels. It never touches steps or CFG, which stay checkpoint-specific and are yours to set. It defaults to off; the README's quick start says turn it on.
• auto_optimize - caps or disables the enhancer stack by compute budget on FM few-step runs. On by default.

The EDM knobs (s_churn, s_tmin, s_tmax, s_noise) and the HC2 knobs (limiter_kappa, hc2_max_order, and the experimental hc2_corrector/sigma_adapt) live in the optional inputs and can stay at defaults until you know you need them. Output is a LATENT, wired straight into the VAE decode like any KSampler.

Installing it

No dependencies, no model downloads - the pyproject.toml ships with an empty dependency list. Either grab it from ComfyUI Manager (search "DDRK Omega Sampler") or:

cd ComfyUI/custom_nodes/
git clone https://github.com/HVOSTOVSKY/DDRK-Omega-Sampler.git

Restart ComfyUI and it's in the node menu.

Where people get burned

• At 6 steps or fewer, the node forces euler unless you've explicitly selected hc2. The console reports the override, so check there if your low-step run suddenly isn't doing what you picked.
• Compare at equal model calls, not equal steps. Heun costs 2 calls per step, RK4 costs 4, HC2 and Euler cost 1. An HC2 run at 20 steps is not "like" a Heun run at 20 steps; it's cheaper.
• sharpness does nothing on EDM. The parameter is shared across families; the feature isn't.
• denoise can't go to 0.0 - it scales steps via steps/denoise, so exactly zero is a division by zero. The minimum is 0.01, which is also just "do nothing."
• If a run goes sideways, flip debug_mode on and it writes per-step JSON/CSV/summary logs to your ComfyUI output folder. The README is honest that most of the 1.6.0 fixes were found reading those logs, not the code.

Fair warning up front: this is a young pack with essentially no community history yet, so you won't find a Reddit thread full of settings for it. What you get instead is an unusually well-documented project - read the README's "What is actually verified" section before you trust the numbers.