Difforum · Feedback Sampler (Classic+)

This is the node the whole pack is built around. Difforum · Feedback Sampler (Classic+) is Deforum's famous feedback loop, rebuilt for modern models: each frame is born from the last - warp the previous frame along the camera path, re-diffuse it as img2img, match its colour back to the first frame, and repeat. The result is the continuous-morphing, psychedelic, living-image look Deforum was known for, except this version takes a plain ComfyUI MODEL, VAE, and CONDITIONING, so it runs SD1.5, SDXL, Flux, or SD3.5 - no 2023-era lock-in, no Python 3.12 refusal, no auto-queue hacks. If you want one sentence for what makes this pack worth installing, it's "Deforum that still runs on a current ComfyUI."

The required inputs read like a sampler plus a camera: model, positive, negative, vae, the params from Anim Setup, a camera track, an init_image for frame 0, and a strength_schedule (the per-frame denoise curve - lower strength = more stable, higher = more morph). Then the usual steps, cfg, sampler_name, scheduler, plus the two colour-coherence controls: color_mode (default lab - perceptual LAB matching, and yes, lab beats rgb for keeping a clip coherent) and color_coherence (0.7–0.9 is the sweet spot per the README).

The optional inputs are where it gets powerful, and the README's "Quality & coherence" section is basically a tour of them:

• depth (+ near/far/invert_depth/translation_scale) - unlocks real 3D camera. No depth map, no 3D motion; that's the pack's most common gotcha.
• positive_schedule - prompt travel, one conditioning per frame; the scene morphs its subject as the camera moves.
• control_net + control_image + control_strength - structure guidance each frame (keep it ~0.3–0.5 if it feels rigid). Must match your base model family (SD1.5↔SD1.5, SDXL↔SDXL).
• symmetry + symmetry_segments - fold the frame inside the loop; the diffusion heals the seams, giving the living-kaleidoscope look that the mesh of every frame compounds.
• sharpen (0.2–0.5) and noise (0.02–0.05) - re-inject detail every frame so the loop resolves it into structure instead of amplifying blur (the anti-mush trick, same as Detail Guard).
• border (default reflection) - stops black edges creeping in on zoom-out and rotation.
• cadence (2–3) - only diffuse every Nth frame; the in-between frames ride the camera warp. Roughly N× faster with smoother motion.
• seed_mode - fixed (default) reuses the same noise each frame, calming texture boil enormously; increment for livelier churn.

Output: a single frames IMAGE batch (frame 0 through max_frames), ready for VHS-style post, Flow Stabilize, or Save Video.

Install

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

Restart ComfyUI ([Difforum] loaded N nodes in console) or ComfyUI Manager → "Difforum". numpy only; the sampler uses ComfyUI's own common_ksampler underneath, so no extra deps.

Common issues & the fixes that actually work

• Flicker / texture boiling - lower per-frame strength to ~0.4–0.5, keep seed_mode=fixed, and run the frames through Flow Stabilize (~0.5). Naive frame blending ghosts; flow-aligned blending doesn't.
• Colour drift - color_mode=lab, color_coherence 0.7–0.9.
• All-black frames on SD1.5 - a real, pack-documented one: some ComfyUI builds make SD1.5 fp16 NaN. Launch with --force-fp32; it's a base-model precision issue, not Difforum.
• Mush after many frames - sharpen ~0.3, noise ~0.02–0.05.
• 3D camera does nothing - you're in 3d mode without a depth input.

This is the node people actually mean when they say "Deforum." It's also the one where the pack's 16 test suites and the difforum_feedback_classic.json template earn their keep - start from the template, swap in a modern checkpoint, and let the loop do its thing.