SamplerEulerAncestralDancing_Experimental
Euler ancestral where eta itself gets to dance
- SAMPLER
SamplerEulerAncestralDancing_Experimental is exactly what its name promises: Euler ancestral with a twist, and marked experimental so nobody can say they weren't warned. It's from the ComfyUI Extra Samplers pack by Clybius, it appears in the sampler dropdown after install, and it outputs a SAMPLER for your custom sampling chain.
Let's get the ancestral-sampler context right first, because it decides whether this node is for you at all. Euler ancestral (Euler a) is a non-converging sampler: it injects fresh noise at each step, so it never settles on a stable image the way Euler or DPM++ 2M do. Same seed and prompt, it's reproducible - but add more steps and the image changes rather than refining. That's not a bug, it's the character of ancestral sampling: a bit more creative variety, a bit less predictability. The knowledge base's sampler essay calls this the one piece of sampler theory that hasn't changed across the flow-matching era, so it applies to whichever model family you're on.
The "dancing" part is the pack's addition. Normally eta (the strength of that per-step noise injection) is fixed for the whole run. This sampler lets the eta itself vary: there's a eta input and a separate eta_dance input, and the sampler oscillates or "dances" between them across the steps. leap (default 2) controls the cadence - how many steps between the two eta values, with a range of 1 to 16. Lower leap = faster oscillation, more aggressive texture changes; higher = smoother.
The inputs that matter:
eta- the base noise strength (ancestral energy). Default 1.eta_dance- the other end of the oscillation. Default 1 means no dancing at all; set it different frometato actually use the feature.leap- steps between oscillations. Default 2.s_noise- how much of that per-step noise is applied (ancestral samplers scale their noise injection by this). Default 1.noise_sampler_type- the usual pack selection: gaussian, uniform, brownian, pyramid, perlin, laplacian, etc.
Honest take: this is a toy in the best sense - an experimental sampler for people who want to feel out how eta oscillation changes texture and variety. It's not going to beat a well-chosen converging sampler for reliability, and the "experimental" tag in the class name is the author's own disclaimer. If you're a beginner, don't start here; if you've already got a workflow you like and want to shake up the noise character on a fixed seed, it's a fun hour.
Install is the shared pack routine - ComfyUI Manager → "ComfyUI Extra Samplers", or:
cd ComfyUI/custom_nodes
git clone https://github.com/Clybius/ComfyUI-Extra-Samplers
then restart. No model downloads; the dependency is kornia. Keep eta_dance equal to eta if you just want plain Euler ancestral from this pack, and only then start making it dance.
Inputs (5)
| Name | Type | Default | Description |
|---|---|---|---|
| noise_sampler_type | COMBO | 7 options: gaussian, uniform, brownian, highres-pyramid, pyramid, perlin, +1 | |
| eta | FLOAT | 1.000–100 | — |
| eta_dance | FLOAT | 1.000–100 | — |
| s_noise | FLOAT | 1.000–100 | — |
| leap | INT | 21–16 | — |
Outputs (1)
| Name | Type | Description |
|---|---|---|
| SAMPLER | SAMPLER | — |