Model Enhancer (Tensor Prism)

"Enhancer" is doing a lot of heavy lifting here, so let's be clear about what this node does and doesn't do. It will not upscale your images. It will not fix your prompt. It doesn't retrain anything either. What Model Enhancer (Tensor Prism) does is take an already-loaded checkpoint and nudge its weights in place - a smoothing pass, a sharpening pass, a quality boost - then hand you back a MODEL that behaves slightly differently at sampling time. Same file, same disk, no training. It's a post-processing filter for the network's parameters.

That framing matters, because people reach for "enhance" nodes expecting magic. What you actually get is the kind of last-mile tuning you'd otherwise have to bake into a merge: gentle contrast in the weight distribution that can pull slightly flat or overbaked checkpoints toward more usable output. The author (Arctenox, first node pack, openly "vibe-coded" per the dev notes) shipped it in the Transform category, and 1.7.1's changelog specifically lists a fix for it - so make sure you're on the latest version before judging it.

How it works

Under the hood it's a weight-space operation. You feed in one checkpoint (a MODEL), and the node clones it, walks the parameters of whatever modules_to_enhance you picked - unet, vae, text_encoders, or all - and applies three knobs:

• smoothing (default 0.12): evens out weight noise; the "calm it down" knob.
• sharpening (default 0.12): the opposite - emphasizes weight structure; the "add some bite" knob.
• quality_boost (default 0.08): a magnitude lift aimed at richer output.

blend_strength (0.9) decides how much of that processed result survives against the original weights, and precision (fp16/fp32) controls the working dtype. method flips between linear and attention, the latter routing extra work into attention layers, and attention_iterations (default 2) sets how many passes those get. There's an adaptive_overbake_prevention toggle on by default - it's exactly what it sounds like: a brake that stops repeated passes from cooking the weights until the model goes mushy or oversaturated.

The only output is enhanced_checkpoint, a MODEL. Wire it into your KSampler's model input in place of the raw checkpoint.

The settings that matter

Start boring: leave modules_to_enhance on unet, keep all three strengths at 0.1 or below, and only raise one at a time. Bump smoothing when a model outputs noisy or grainy results, sharpening when it looks soft, quality_boost when it looks flat. If you push two strengths above 0.2 simultaneously you're basically merging a filter stack into your model and the results get unpredictable fast - which is also where adaptive_overbake_prevention earns its keep.

Installing it

The pack is ComfyUI-Tensor-Prism-Node-Pack. Easiest path: ComfyUI Manager → search "Tensor Prism" → Install, then restart. Manual install:

cd ComfyUI/custom_nodes
git clone https://github.com/AstrionX/ComfyUI-Tensor-Prism-Node-Pack

Restart ComfyUI and you'll see the welcome banner print your GPU's VRAM and node count. No model files to download; the only dependencies are torch, numpy and psutil, which any ComfyUI environment already has. One gotcha: the pack's own README lists a git clone URL that's typo'd and stale (Arctebix/Tensor_Prism.git), so use Manager or the comfy.icu URL above rather than copy-pasting from the README.

Where people get burned

The trap is expectation. This node tweaks weights; it won't rescue a bad checkpoint, and cranking quality_boost to 1 won't give you "better" images, just stranger ones. Also note this is a small, young pack from a first-time author - there's essentially no community corpus on it yet, so treat it as a tool to experiment with, not a settled technique. If a merge produced a model that's 90% there but slightly overcooked, this is a cheap last-mile lever. For anything else, you're probably reaching for a real merge instead.