AetherScale • Diagnostics

If the pack has a "read me before posting a bug report" button, this is it. AetherScale's Diagnostics node produces a one-shot snapshot of your whole environment - the kind of thing support threads die without. It probes the private NVIDIA runtime, lists your GPUs, reports your Python/torch/CUDA versions, and dumps the DLSS5/neural-rendering state into a single JSON string. One input (refresh, a boolean that controls whether it re-probes rather than reusing cached state - leave it true), one output (diagnostics). That's the whole node, and that's fine; it's a tool, not a toy.

What the JSON tells you

Run it and you get, in one blob: the AetherScale version, whether the NVIDIA runtime is ready and which version is installed versus required, Python/torch/CUDA versions, a per-GPU breakdown with name, total VRAM in GiB and compute capability, effect-cache entry count, available runtime capabilities, and the neural-rendering state. It even lists every node the pack ships and a set of plain-English notes explaining the pack's design choices - like why Restoration groups three tasks into one node and why the HDR binding is adaptive. The author wrote it as documentation you can query.

When to run it

Three moments. Right after first install - confirm runtime_ready: true before you blame a VFX node for a missing runtime. When a node errors - the diagnostics will tell you at a glance whether the runtime bootstrapped at all, whether your card's compute capability is in range, and whether the effect cache is accumulating entries. And before you ask for help anywhere - the Neural Rendering section of the README explicitly asks for your GPU model, driver version, AetherScale version, the traceback, and the node's stats output when you report an issue. This node is how you produce the first three in one copy-paste. There's no troubleshooting equivalent of "works on my machine" harder than a diagnostics dump.

Installing it

Same routine as the rest of the pack:

cd ComfyUI/custom_nodes
git clone https://github.com/vizart-vj/ComfyUI-AetherScale.git

Restart ComfyUI. Manager users can search AetherScale instead. The environment contract applies - Windows 10/11, an NVIDIA RTX card, Python 3.10+ - and note that Diagnostics itself needs nothing special to run: it probes the runtime state without downloading anything, so it works even when the runtime install has failed, which is precisely when you need it most.

The honest note

Diagnostics can't fix anything - it reports. When the JSON shows runtime_ready: false, the actual remedy lives in the sibling Runtime node (install_or_update, or repair after a driver update). And don't leave one wired into a production workflow queueing on every run; it's a checkup, not a processing step. Drop it on the canvas when something's wrong, read the dump, act, delete. Two minutes of that beats an hour of guessing, and if you do end up posting the output, you've already done the thing that makes people actually want to help you.