Negative momentum cuts sampler error 11-21% vs dpmpp_2m at same step count

YakovSava · reddit · 2026-11-14

Reddit user YakovSava dissects why dpmppdualsdemomentumized from ComfyUI-Extra-Samplers produces blue blobs on flow models like Anima, and ships fixed samplers.

The bug isn't momentum: the sampler was written for variance-exploding models; its noise step has wrong variance on flow-matching models (latent std 3-4.5 with eta=1 vs 1.2 expected), plus a 3M correction stacked on an already second-order 2S step and an overwritten r parameter. Verified via an exact port (1e-10 match) with per-component ablations.
What momentum should be: in the DPM-Solver++ frame it's an EMA of denoised output — positive coefficients act as a low-pass filter (error up to +61% worse), negative coefficients become extrapolation and reduce error, stable only for |m|<1.
Measurements (Anima, bf16, 15 images, paired bootstrap): deterministic m=-0.6 gives 11-21% lower error than dpmpp2m at equal steps (1.2x fewer steps); damped extrapolation (beta=0.5) beats dpmpp2msde by 7-10% with no color cast; a one-eval-per-step sampler beats dpmppsde by 16-29% at equal NFE.
Artifacts: node pack with ddimmomentum and dpmpp2msdedamped, plus momentumized variants of 19 samplers (unvalidated; lms/ipndmv versions blow up latents at default strength).

Original post →

Negative momentum cuts sampler error 11-21% vs dpmpp_2m at same step count

More from Multimodal