Difforum · Detail Guard (anti-mush)

Every Deforum-style feedback render has an enemy: mush. Each frame goes through a warp, a VAE encode/decode, and a re-diffusion, and every one of those steps softens the image a little. Run it for 120 frames and what started as crisp detail has melted into smooth, featureless goo - the classic feedback-loop blur. Difforum · Detail Guard (anti-mush) is the counter-punch: an unsharp mask, a contrast pass, and a controlled hit of noise, all in one IMAGE-to-IMAGE node. The noise is the clever part - the diffusion sampler picks up fresh high-frequency detail and resolves it into structure rather than just amplifying grain. It's the same trick the Feedback Sampler's built-in sharpen/noise knobs use; this is the standalone version you drop on footage, a finished batch, or any image that needs its detail back.

Inputs are refreshingly few:

• sharpen (default 0.3, 0–2) - the unsharp amount. 0.2–0.5 is the README's working range; too high and you get crunchy halos.
• noise (default 0.02, 0–0.5) - how much high-frequency grain to inject. The sweet spot is small: 0.02–0.05. This is the "food for the diffusion" knob, not a texture slider.
• contrast (default 1.0, 0.5–2) - a gentle S-curve nudge. Punch it up when the loop is flattening everything toward grey.
• noise_mode - gaussian or plasma. Gaussian is the standard detail-reseed; plasma is spatially correlated noise, which can read as more "organic" texture on some subjects.
• seed (optional) - makes the noise deterministic so the same frame always gets the same grain, which matters for anti-flicker. Leave unset and the noise re-rolls each run - usually not what you want on video.

Output: one image (IMAGE), same shape in, same shape out, ready to chain or save.

Why you'd reach for it specifically: it's the post-hoc version of the sampler's built-in detail guard. Two scenarios. First, you rendered a clip with sharpen=0 and it's mushy - run the batch through this and save it instead of re-rendering. Second, you're processing footage in the VJ pipeline (Load Video → grade → trails) and you want a detail pass that doesn't depend on the sampler at all. It works on a single still too, which makes it a fine "rescue this soft image" node.

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; the effect chain runs on CUDA/MPS/CPU automatically.

Common issues

• "It looks crunchy." Sharpen too high, or noise too high - halos and visible grain respectively. Back both off; 0.2 and 0.02 are there for a reason.
• Flicker got worse. If you applied it per-frame with random noise, the re-rolled grain flickers. Fix seed (or better, let the Feedback Sampler's in-loop sharpen/noise handle it with its fixed-seed mode).
• No visible change. At defaults the effect is intentionally subtle - it's a preservative, not a filter. Crank sharpen to 1+ if you want to see it move.

Detail Guard won't win any awards on its own, but it's the difference between "Deforum render, frame 90" and "Deforum render that still looks good at frame 90." Keep it on the clipboard for every long clip.