🔧 Sigma Debug (Boyo)

Here's a quiet frustration of ComfyUI: you set a scheduler, and ComfyUI goes and computes a list of sigmas - noise levels, one per step - and you never get to look at them. When an image comes out soft, or banded, or the sampler is producing NaNs, the first question is usually "what did that scheduler actually give me?" BoyoDebugSigmas is the answer: a pure passthrough for the SIGMAS tensor that prints the whole schedule to the console.

Sigmas are the backbone of sampling. The scheduler decides how much noise to remove at each step, and it hands the sampler a descending list - from a big starting sigma down to near zero, plus a terminal zero. BoyoDebugSigmas shows you every value: the count (including that terminal zero, which trips people up), the max, the min before the terminal, the full value list, the tensor dtype, and whether it's living on CPU or CUDA. Wire it in, run one step, and you'll see something like this:

============================================================
BOYO SIGMA DEBUG
============================================================
  Count  : 9 (including terminal zero)
  Max    : 1.000000
  Min    : 0.125000 (before terminal)
  Terminal: 0.000000
  Values : [1.0, 0.875, 0.75, 0.625, 0.5, 0.375, 0.25, 0.125, 0.0]
  dtype  : torch.float32
  device : cuda
============================================================

That one print tells you a lot. If the values stay suspiciously high, your schedule ends too hot and details never resolve. If there are NaNs in there, that's your real bug - no sampler can recover from a NaN schedule. And if you're debugging a flow model and the numbers aren't in the [0, 1] range, that alone explains why your Wan or LTX output looks wrong; flow models want sigmas in [0, 1], full stop.

Inputs are one field: sigmas, type SIGMAS. Output is sigmas, the exact same tensor, so you can drop the node anywhere a scheduler output flows - after BasicScheduler, after BatchSchedule, after a custom sigma generator - and feed the passthrough into SamplerCustomAdvanced or whatever sampler consumes it. It's the sibling of BoyoDebugSampler, same philosophy: inspect, pass through, remove when you're done.

Installing it

Same pack as everything else: search Boyonodes in ComfyUI Manager, or clone into custom_nodes and restart:

cd ComfyUI/custom_nodes
git clone https://github.com/DragonDiffusionbyBoyo/Boyonodes

The debug nodes pull nothing extra - plain torch tensors and print(). The pack's heavy requirements (chatterbox-tts, funasr, seed-vc, and friends) belong to its audio and video corners; none of them matter for this node.

Where people get burned

• The terminal zero is normal and expected - every sigma tensor ends in 0.0. Don't "fix" it.
• Output goes to the terminal, not to the canvas. If you're running in a bundled/portable install, that's the console window you launched ComfyUI from.
• If the whole pack fails to appear after install, read the first console traceback - a single heavy module in this kitchen-sink pack can fail to import and take every Boyo node down with it. Install whatever the traceback names, restart, move on.

The use case is narrow but real: when a render misbehaves and you're about to blame the sampler, spend one run proving whether the schedule was sane first.