Krea 2 Best Samplers & Schedulers (Tested)

Sampler choice changes Krea 2 output more than most people expect — not just speed, but how faithfully it holds a subject’s identity across seeds. We compared samplers on Krea 2 Turbo at matched seeds. Here’s what actually held up.

The short answer

• Best quality (and identity fidelity): res_2s + bong_tangent scheduler. It’s the most faithful, at roughly 2× the generation time.
• Best balance (recommended default): res_multistep + simple. It lands very close to res_2s in quality and is actually faster than euler — a rare case of better and quicker.
• Soft baseline: euler + simple. Fine, a touch softer, nothing special.
• Avoid on Turbo: er_sde. Identity tends to drift between seeds on the distilled Turbo model (it behaves better on un-distilled setups at very low steps, but that’s a different regime).

If you only remember one thing: res_multistep + simple is the sensible default for Krea 2 Turbo, and you reach for res_2s + bong_tangent when a specific image needs the extra fidelity.

Where these samplers come from

res_2s, res_multistep, res_3s and the bong_tangent scheduler are part of the RES4LYF sampler family. In ComfyUI they can be injected into the standard KSampler, so you don’t need a special sampler node to use them — just select them where you’d normally pick euler/dpmpp.

Two Krea-2 gotchas worth knowing

1. torch.compile doesn’t work with Krea 2. Compiling the model currently throws a dtype error (bf16 vs float) in the inductor backend, so there’s no output-identical speed-up from torch.compile right now. Worth re-testing after ComfyUI updates, but don’t count on it today.
2. Negative prompts barely apply on Turbo. Because Turbo runs at CFG 1, sampler choice won’t rescue a prompt that relies on negatives — you have to phrase things positively. (See how to prompt Krea 2.)

Skip the tuning — run a tuned pipeline online

Dialing in samplers, schedulers, and step counts is exactly the kind of fiddly work a hosted pipeline handles for you. On AIFLUX, Krea 2 Turbo runs on a tuned graph, so you just prompt and generate.

👉 Generate with Krea 2 online

Frequently asked questions

What is the best sampler for Krea 2 Turbo? res_2s + bong_tangent for maximum quality; res_multistep + simple for the best speed/quality balance (and it’s faster than euler).

Which sampler should I avoid on Krea 2? er_sde on Turbo — identity tends to drift across seeds.

Does torch.compile speed up Krea 2? Not currently — it fails with a dtype error on this architecture. There’s no output-identical accelerator available today.

Do I need a special node for res_2s / bong_tangent? No — the RES4LYF samplers and schedulers can be injected into the standard ComfyUI KSampler.

Prefer results over tuning? Run Krea 2 on AIFLUX →