Sampler Ideas

ComfyUI-IdeaForge

Standalone ComfyUI sampler nodes built from published mathematical concepts. The ideas are referenced; every implementation here is an original clean-room implementation written for this pack.

All samplers follow the ComfyUI 0.44 sampler contract and use a built-in Anima-style flow linear-shift sigma grid (an optional sigmas input overrides it). Warp Sigma Scheduler is model-agnostic.

Nodes

| Node | Concept source | Cost | Notes | | --- | --- | --- | --- | | Overshoot Sampler | AMO Sampler (arXiv:2411.19415), base overshooting mechanism | 1 NFE/step | Oversimulates each step by (1+overshoot)×, then re-noises back to the target sigma's marginal. Attention modulation from the paper is not included. | | ERK Guidance Sampler | "Error as Signal" (ERK-Guid) | 2 NFE/step | Heun step with an embedded Euler estimate; the pair difference is the solver's own local error estimate and is added back as a guidance direction scaled by erk_strength (0 = pure Heun). | | Flow Taylor Sampler | RF-Solver (arXiv:2411.04746), Taylor-expansion idea | 1 NFE/step | Second-order step: velocity derivative estimated by finite difference from the previous step. | | Warp Sigma Scheduler | Align Your Steps (arXiv:2404.14507), schedule-shape spirit | – | SIGMAS node: takes any built-in scheduler grid for the connected model and warps point placement (detail = denser low-sigma, smooth = denser ends). No per-model optimisation is performed. | | Density Sampler | Diffusion-density / mode-tracking writeup (Karczewski) | 2 NFE/step | Experimental approximation: a probe evaluation measures how the prediction field bends away from the current data estimate, and the step is corrected against that bend. |

KSampler registration

On load, the pack also registers four of the samplers into the stock KSampler sampler list — overshoot, erk_guidance, flow_taylor, density_probe — with fixed default parameters (overshoot 0.5, erk_strength 0.5, density 0.3 / probe 0.05), so they can be picked from any ordinary KSampler (scheduler simple recommended). The standalone nodes remain the place for adjustable knobs. Registration is wrapped so a failure can never break node loading.

Verified on real hardware — and one bug worth writing down

Tested on an RTX PRO 6000 (MIG 24 GB) pod, mac studio m3 ultra 512gb, ComfyUI 0.44, Chiarezza 2.9B hybrid: all four samplers generate cleanly, via both the standalone nodes and the KSampler registration.

Overshoot's first version diverged into neon blow-out even at overshoot 0.2. The cause: its re-noise step restored the marginal with the VP-diffusion formula, which over-weights the signal (x0) component under flow interpolation — a small, systematic per-step bias that compounds. The flow-correct step subtracts (sigma_next − s_over) · denoised after re-noising; a numpy check of the one-step statistics (signal coefficient 0.348 before, 0.298 after, target 0.300) confirmed the fix before the pod re-test passed at overshoot 0.5.(same mac studio)

Install

cd ComfyUI/custom_nodes
git clone https://github.com/seren-alegro/ComfyUI-IdeaForge
# restart ComfyUI

Or copy this folder as ComfyUI-IdeaForge into ComfyUI/custom_nodes manually, then restart ComfyUI.

Tests

tests/ — math tests for the stepping rules (run where torch exists).

License

MIT

---

한국어 요약

논문·글에서 본 수학 구상 다섯 개를 깨끗하게 다시 구현한 실험 샘플러 팩이에요. Overshoot, ERK Guidance, Flow Taylor, Density Probe 샘플러와 Warp Sigma 스케줄러가 들어 있고, 앞의 네 샘플러는 일반 KSampler 목록에도 등록돼요. 전 샘플러 포드와 mac studio m3 ultra 512gb으로 실측을 마쳤고, Overshoot의 flow용 재노이즈 보정 과정도 위에 기록해뒀어요.