Difforum · Live Sampler

The Feedback Sampler runs to completion and hands you a clip. The Live Sampler does the same loop in real time and shows you the frames while it runs - which changes what the tool is for. It's not a rendering node. It's a performance instrument.

The setup

Pair it with a 1–4 step model - SD-Turbo, SDXL-Turbo, LCM, DMD2 - at around 512 px. The defaults are already that: steps 2, cfg 1.2, sampler_name lcm, scheduler sgm_uniform. Distilled models collapse a 20-step trajectory into two, which is the only reason a per-frame feedback loop can keep up with reality; if you point this at a regular SDXL checkpoint at 20 steps you'll get a slideshow, not a live render.

Required inputs are the usual four (model, positive, negative, vae) plus init_image. run_frames (480) is how many frames to run, and loop_camera (on) repeats the camera when the run is longer than the clip - that's what makes a short move into an endless one for VJ use.

The inputs that make it a live tool

• live_source - '' off, '0' for webcam, or a video path. It's mixed in before each frame, so the model is re-imagining the live feed continuously. Point it at a webcam and you have a magic mirror; point it at a looping video and you have a per-frame restyle engine.
• source_blend (0.9) - how much of the live source enters each frame. This is the feedback-vs-live dial. High means the source dominates; lower means the model's own history accumulates and the image drifts away from the input, which is where the interesting artifacts live.
• stream_dir and spout_name - where the frames go. stream_dir writes live PNGs to a folder, which any other app can pick up. spout_name is a Spout sender for Resolume, TouchDesigner or OBS - Windows only, and it needs pip install SpoutGL; the node prints a "Spout output disabled" line and carries on if the import fails, which is the right behaviour and also the reason people sometimes think Spout is broken when it's just not installed.
• target_fps - caps the speed. 0 means as fast as the hardware goes, which on a 4090 with a turbo model is faster than your eyes and warmer than your room.
• keep_frames (240) and live_preview - a ring buffer, so a long session doesn't fill RAM with 10,000 frames you'll never look at. frames returns only the tail.

Outputs

frames (the kept portion - wire it to Save or Video Combine after you stop) and report, which is worth reading once: it tells you the achieved frame rate.

Install

ComfyUI Manager → search "Difforum", or:

cd ComfyUI/custom_nodes && git clone https://github.com/chillithebillis/Difforum.git difforum

Restart. Base install is numpy only; webcam and video-file input need OpenCV (pip install opencv-python) and Spout output needs SpoutGL on Windows. Both are imported lazily, so a machine without them still loads the pack fine - only the features that need them go quiet.

Where people get burned

The model, first and always. A 4-step LCM model at 512 px is the difference between an instrument and a coffee break. Second: run_frames of 480 at 24 fps is 20 seconds, and the node will happily run until it finishes or you interrupt - there's no "keep going until I say stop", so for a long session you re-queue with a large run_frames rather than expecting it to stay open.

And the occasional reminder: a live source mixed at high blend with a strong camera move produces feedback that eats itself into mush within a few seconds. That's the loop, not a bug. Pull source_blend up or dial the energy down when it happens.