pi-Flow Sampler

pi-Flow Sampler is the engine of the whole pack - the node that turns a pi-Flow model into 4-step generation that keeps seed diversity. Where most speedup techniques (Lightning, Turbo, DMD2-style adapters) trade variety for speed and can make "every seed is the same image," pi-Flow frames the model as predicting a policy - a Gaussian mixture over possible velocity directions - and the sampler integrates that policy instead of following a single velocity. That's the whole difference, and it's why the pi-Flow release threads got 148+ upvotes in days: quality at 4 steps, but the seeds still vary like a full model.

You replace your KSampler with this node in a pi-Flow workflow. Don't try to run a pi-Flow model through a normal sampler - the author is blunt that pi-Flow doesn't use standard diffusion samplers, and the model outputs a policy a KSampler has no idea what to do with.

The inputs that matter

You'll set maybe four of these and never touch the rest.

• steps - default 4. The 4-step adapter works at 4 or more; the community finds 8 steps noticeably better for coherence on complex scenes while staying way faster than base. Anything ≥ 4 is fair game.
• seed - standard noise seed.
• conditioning - your positive prompt conditioning. Wire the FluxGuidance node in front of it and don't change its value: 3.5 for FLUX.1, 4.0 for FLUX.2. Wrong guidance = noisy output, and the workflow marks it "(DO NOT MODIFY)" for a reason.
• latent_image - the empty latent to denoise (EmptySD3LatentImage in the shipped workflows).
• denoise - default 1.0. Lower it for img2img-style work; it preserves the structure of the input latent.

Then the policy internals: substeps (128) is how many sub-integrations each network step gets along the policy rollout - leave it. final_step_size_scale (0.5) shrinks the last step relative to the others - leave it. diffusion_coefficient (0) controls stochasticity; 0 is fully deterministic, 1 is standard DDPM-style noise - most people leave it at 0. gm_temperature (auto) and manual_gm_temperature (1.0) shape the Gaussian mixture; auto picks a sensible value from your step count.

Output: a single LATENT that goes into VAE Decode.

Install and models

cd ComfyUI/custom_nodes
git clone https://github.com/Lakonik/ComfyUI-piFlow

Restart ComfyUI (or ComfyUI Manager → search "ComfyUI-piFlow"). Requires ComfyUI 0.17.0+; current releases pin a newer floor, so update ComfyUI first. Model files come from the Load pi-Flow Model node's README section - base model, text encoder, VAE, and the small GMFlow adapter per family. On an 8GB card people do run pi-Flow (Flux.2 at 4 steps in under a minute, per one 3060 Ti report), but expect offloading slowness - pi-Flow uses as much VRAM as the base model.

Troubleshooting

Noisy output at 4 steps: check your FluxGuidance value first - this is the #1 cause, not the sampler. Contrast looks wrong: you're probably running a 4-step adapter at 8 steps (or vice versa) - adjust adapter_strength on the loader (1.0–1.4) rather than the sampler. And if the image looks structurally broken at high resolution, that's the adapter's 1MP training resolution showing - bump to 8 steps or upscale progressively.