ComfyDL

It says Accuracy, but it actually hands you a count

Nine activation functions in one node — only two of the knobs do anything

Additive Attention

The residual + layer-norm layer hiding inside every Transformer block

Plot a training curve as a filled area chart, straight from a tensor

The anchor-matching step behind object detection, exposed as a node

The friendliest object-detection dataset in existence

Turn a tensor into a bar chart with the numbers written on the bars

BLEU in one node, so you can score a translation by hand

Bounding boxes in (cx, cy, w, h)? Convert them to corners in one step

Bounding boxes in (x1, y1, x2, y2)? Convert them to center-width-height

Pairwise IoU between two sets of boxes, computed in one shot

Stretch a tensor to a bigger shape without copying the data

See exactly where your classifier gets confused

Convolve a tensor with a kernel — stride and padding included

Cross-correlation with the loops on display, straight from the textbook

Num_batches, batch_size, dataset_size

DataLoader Preview

The same preview grid, but it shows up in your UI automatically

Is your dataset actually balanced? Dataset Stats will tell you

Which GPU does ComfyUI actually see? This node answers in two ints

The scaled dot-product attention behind every transformer, as one node

ComfyDL's Download node

One dropdown, download + extract included

Fashion-MNIST as a ComfyUI dataset

[CLS] … [SEP] … [SEP], with segment IDs

Clamp runaway gradients before they nuke your training

The LSTM's cheaper, lighter cousin — as a node you can stack

Extrude a tensor into a mesh you can actually orbit

See the distribution inside a tensor in one glance

Brightness, contrast, and saturation in one pass — 1.0 means leave it alone

Grayscale that keeps all three channels — the sneaky part

Normalize an image to z-scores — and know it won't come back to [0,1]

Rotate by any angle, counter-clockwise — and expand the canvas or don't

What's actually inside that image tensor? Just ask it

Xavier weight init in one node

The 1998 CNN that still teaches everyone deep learning — now a node

Y = Xw + b as three wires — linear regression without a single line of code

The bridge from raw tensors to a trainable dataset

The softmax that knows where to look — and where to look away

A Windows popup that taps you when the graph finishes — if you're on Windows

A copy of a model that shares nothing — deepcopy, done properly

Run your model on a tensor — eval mode, no gradients, done

How big is that model? Total params, trainable params, and a summary

Read your network's mind — every layer, indented, as text

Load weights into a model you already built — the architecture must match

Train vs eval — the tiny toggle that changes what your model does

Every parameter, its shape, and whether it actually trains

Save a model's weights — only the weights, and only to a .pt

Turn anchor guesses into actual detections — that's non-max suppression

Generate every anchor box your detector will ever need

The node that turns anchor boxes into something a loss can chew on

How one attention becomes eight

How your detector stops drawing fifty boxes on one dog

The ComfyDL node that does literally nothing, and that's the point

The regression half of training a single-shot detector

Turning a detector's offset predictions back into real boxes

When you need to actually see your label distribution

The MATLAB-style line plot hiding inside ComfyUI

Attention can't see order — this node gives it a position map

The two-layer MLP that every Transformer block repeats

ComfyDL's noise generator for testing, init, and inputs

The node that fixes 90% of your shape errors in ComfyDL

The Lego brick of every modern CNN, as a node

A real deep network, assembled in one node

Grouped convolutions without leaving the graph

One line of history, no hand-written math

RNN Language Model (high-level)

RNN Language Model (from scratch)

The payoff node of the LM demo

No hidden torch layer, just weights

Seeing 10,000 points instead of 10,000 numbers

Half a machine-translation model, built honestly

The polite way to tell a network to ignore padding

The node that actually trains your model

The node that makes your detector's output real

Turning attention and weight matrices into something you can see

Attention, rendered right in the graph

A batch of tensors, finally looking like a grid of pictures

Watching an optimizer wander its way to the minimum

The loss function that taught a generation of regression — half, on purpose

Type a matrix by hand and get a tensor — the pack's manual data entry

A fake dataset generator that makes your first training loop honest

The node that lets you actually see what a tensor is doing

Is the GPU actually faster? This node settles the argument with a stopwatch

From raw text to tokens — the first step in every NLP workflow here

A BERT-class encoder you build (and see into) yourself

One Transformer encoder block, the Lego brick of BERT

Swap any two dimensions — and learn why order matters so much

Force every sequence to the same length, the boring way that just works

One node, every GPU on your machine, comma-separated

Ask for GPU 0, get told 'cpu' instead — the failover node

One training step for the critic — where GAN loss actually comes from

The generator fights back — one step at fooling the critic

Turn a pile of tokens into a real vocabulary — the pack's index-maker

Numbers back to words — the node that makes model output readable

Words in, numbers out — the node that makes text computable

The 21 names behind every PASCAL VOC segmentation label

The lookup table that turns VOC's colors into class numbers

VOC's color-coded masks, converted to class indices you can train on

Crop the image and its label together, so the segmentation data stays honest

A 2 GB dataset that appears at your doorstep (and where it hides)

The '?' node has one toggle and it opens a browser tab — don't flip it at work