Inference_Core_ModelSamplerTonemapNoiseTest

Worth reading the category and the class name before anything else: this node lives under custom_node_experiments, and "Test" is literally in its name. This is the pack's own scratch space, not a polished production feature - treat it as something to try, not something to expect documented, stable behavior from across every checkpoint.

What it does

Diffusion sampling can run into a familiar failure mode: push CFG high enough for strong prompt adherence and the predicted noise starts running hot, compounding step over step into oversaturated colors, blown-out contrast, and deformed faces - the classic "burn" everyone eventually hits chasing better prompt following. This node patches the model's sampler to apply a tonemapping operation to the predicted noise during sampling, which is the same underlying idea as HDR tonemapping in photography or rendering: compress extreme outlier values back toward a sane range instead of letting them clip or compound unchecked. Applied to noise prediction, the goal is to tame that burn without simply lowering CFG and giving up the prompt adherence you raised it for in the first place.

The one input

multiplier (0–100, default 1) scales how strongly the tonemap operator is applied. At its default, it should sit close to a no-op; push it higher to compress the noise range more aggressively. There's no documented target value here - this is a "turn the dial and look at the result" node rather than one with a known sweet spot, which is consistent with its experimental status.

model in, patched MODEL out - wire it into your KSampler exactly like a normal model patch node, nothing else in the graph changes.

When it's worth trying

This is a different lever aimed at the same symptom CFG-burn troubleshooting usually addresses by just lowering CFG. If you specifically want to keep CFG high for prompt adherence and are looking for an alternative to backing it off, this is worth experimenting with. It's not something to add to every workflow by default - most checkpoints look fine within their normal CFG range and don't need any of this.

Installing it

Comes bundled with the LykosAI Inference Core Nodes pack, no separate install. Via ComfyUI Manager, search "ComfyUI-Inference-Core-Nodes". Manually:

cd ComfyUI/custom_nodes
git clone https://github.com/LykosAI/ComfyUI-Inference-Core-Nodes

then install.py, or pip install -e . and restart - no acceleration extras needed for this one, it's a sampler-level model patch, not a separate inference model.

Troubleshooting

If output looks identical whether the node is in your graph or not, multiplier is probably sitting too close to its neutral default - push it up meaningfully and compare before concluding it's not doing anything. Given its experimental status, don't be surprised if results vary unpredictably between checkpoints or don't hold up as well as the mainline, non-experimental nodes in this pack; that's the tradeoff of reaching into the pack's test bucket rather than its supported feature set.