Difforum · Live Sampler (realtime loop)

Stock ComfyUI re-runs your whole graph every time you hit Queue Prompt - fine for a still, useless for "I want to see this Deforum loop happen." Difforum · Live Sampler (realtime loop) is the pack's native realtime engine: it keeps the model resident and runs its own internal loop, each tick warping the camera, folding the symmetry, re-diffusing 1–2 steps, colour-matching, and streaming a live preview into the node as it goes. You queue once and watch it generate. No external runtime, no auto-queue hack, no frames piling up in a temp folder. This is the node that takes Difforum from "render a clip" to "perform a clip."

The catch is honest and unavoidable: realtime means a realtime model. The recipe the README spells out is few-step distillation - SD-Turbo, SDXL-Turbo, LCM, or DMD2 at 1–2 steps, cfg ~1, at 512px. On a Turbo model it genuinely runs live; on a 20-step SDXL render it just runs fast-ish with a nice preview. Same steps/cfg/sampler_name/scheduler family as the Feedback Sampler, but with live-grade defaults baked in (steps=2, cfg=1.2, sampler_name=lcm, scheduler=sgm_uniform).

The inputs worth knowing beyond the sampler basics:

• duration_frames - how long the loop runs (default 240). It's a render length, not an endless stream.
• live_preview (default on) - the in-node preview. Turn it off if you're running headless or exporting instead of watching.
• live_source - the "magic mirror": set it to a webcam device index (0) or a video path and the loop stylizes the live feed - camera → kaleidoscope → diffusion, realtime. source_blend (default 0.9) mixes the live source over the warped feedback.
• stream_dir - writes frames to a folder OBS/Resolume/ffmpeg can read live; spout_name - Spout video out for VJ apps (needs the SpoutGL extra). This is the "map it to a VJ show" half of the node.
• loop_camera (default on) - wraps the camera track so the loop keeps cycling past the end of the schedule.
• cadence - diffuse every Nth tick only; the biggest raw FPS lever after the model choice. target_fps throttles the loop to a rate, which is how you keep a live set from running away.

Output: frames (IMAGE batch) of the whole session - you get the live show and the rendered clip out of one run.

Install

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

Restart ComfyUI ([Difforum] loaded N nodes) or ComfyUI Manager → "Difforum". numpy-only core; optional extras: webcam needs opencv-python, Spout needs SpoutGL (the README flags both). The difforum_realtime_live.json template wires a Turbo checkpoint in with a kaleidoscope ready to go.

Common issues

• "It's not realtime." You're on a full-strength model, or at a big resolution. Drop to a Turbo/LCM/DMD2 model, 1–2 steps, 512px, and add cadence=2–3. The math is relentless; meet it halfway.
• No preview. live_preview is off, or the node didn't reload - restart ComfyUI fully (console: [Difforum] loaded N nodes).
• Webcam input does nothing. opencv-python isn't installed, or the device index is wrong. Try 0, then 1.
• MPS users: fp8 checkpoints fail on Apple Silicon (no fp8 kernels). Use fp16/bf16 or GGUF - the Model Profile node auto-adjusts its recommendation on Macs.

The honest framing: this is the most fun node in the pack, and the one with the most moving parts. The first time you feed it a webcam and watch your face get kaleidoscoped and re-diffused in realtime, you'll understand why the author wrote REALTIME.md. Start with the template, get a loop happening, then start pulling the VJ levers.