SamplerER_SDE

If you've seen "just use er_sde" in a model's settings thread, this is the node they mean. SamplerER_SDE is ComfyUI's implementation of the Extended Reverse-Time SDE solver (ER-SDE, arXiv 2309.06169) - a high-order stochastic sampler that went from a research curiosity in early 2025 to the recommended default on a surprising number of models. Anima's settings table lists it as the neutral default. Chroma enthusiasts recommend it. People run er_sde + beta for character consistency on Qwen. It's real, it's current, and it's the one sampler node that shows up in workflow shares more than sampler discussions.

It ships with ComfyUI core (comfy_extras.nodes_custom_sampler); the solver landed in March 2025 and was migrated to the VP formulation in July 2025.

How it works

ER-SDE is a high-order solver - up to max_stage = 3 - that estimates the denoising trajectory more accurately per step than a first-order method like Euler. The "SDE" part means it's stochastic: it injects fresh noise along the way, which is what gives it the character consistency people cite (stochastic samplers tend to hold structure better than their deterministic cousins). That also means it's ancestral/non-converging - same seed gives the same image, but adding steps changes the image rather than refining it, so don't expect a clean convergence curve.

The knobs:

• solver_type - ER-SDE, Reverse-time SDE, or ODE. They differ in their noise scaler. Select ODE and the run goes deterministic.
• max_stage - solver order, 1–3. Higher is more accurate per step but slower. 3 is the default and usually the sweet spot.
• eta - the author's own tooltip says it plainly: "Stochastic strength of reverse-time SDE. When eta=0, it reduces to deterministic ODE. This setting doesn't apply to ER-SDE solver type." So for the ER-SDE solver type, eta is not your knob.
• s_noise - magnitude of the injected noise.

How people actually use it

The pattern that keeps recurring: er_sde on a conservative scheduler. Anima runs er_sde at 30–50 steps, CFG 4–5; the community settings pair it with beta57 (from the RES4LYF pack) for painterly output, and with simple for speed with a distilled LoRA. On Qwen, er_sde + beta is the go-to for better character consistency than Euler.

One honest caveat from the community: it's a neutral default, not a magical one. "euler_a" is softer, "dpmpp_2m_sde_gpu" more creative - people describe er_sde as the balanced middle. And it's a ComfyUI/k-diffusion thing, not something you can grab in every UI - there's a whole thread of Forge users asking "where do I get er_sde?" (thread 1nd9b5u). In ComfyUI it's just a node.

Where people get burned

• Large eta on the SDE solver types. The tooltip again: "Large eta may cause invalid outputs. If this occurs, try decreasing this value."
• Expecting it to behave like a converging sampler. It won't settle the way Euler or DPM++ 2M does; step-count changes reshuffle the image.
• Forgetting the scheduler. er_sde's reputation is tied to beta/beta57/simple. Stick it on karras and you get the flow-matching karras failure regardless of how good the solver is.

Wire the SAMPLER output into SamplerCustomAdvanced and you're set. If the model card doesn't mention er_sde, KSamplerSelect's dropdown is the safer default - this node earns its keep when the model was tuned on it.