![Banner](./banner.png)

ComfyDL

DeepLearning is just a few clicks away!

> **GUI version available:** [ComfyDL_UI](https://github.com/Cynthia-lxx/ComfyDL_UI) ships > ComfyDL as a built-in node pack on a ready-to-run ComfyUI fork — clone it, launch `main.py`, > and the nodes are already there. See [Installation](#installation). --- ## What's This? **ComfyDL** lets you build deep learning workflows — from CNNs to BERT and beyond — by connecting nodes in ComfyUI, not by writing code. Built on a foundation inspired by the `d2l` codebase, it keeps evolving as we develop more useful nodes — visual, educational, and great for rapid prototyping. Drag, connect, and see results instantly. **ComfyDL has its own GUI version:** [ComfyDL_UI](https://github.com/Cynthia-lxx/ComfyDL_UI) ships ComfyDL as a built-in node pack on top of a ComfyUI runtime fork — see [Installation](#installation). ## Examples in Action > **Note:** The example workflows that ship with ComfyDL are served as **workflow templates** in the UI > (Templates → ComfyDL Examples): language-model, regression, classification and App-mode > playground workflows. `formula_playground_app` opens in **ComfyUI APP mode (beta)** by default. > Older examples that predate the node refactor are kept for reference under > [`example_workflows/_archived/`](./example_workflows/_archived). ### The Mysterious "?" A tiny node with a toggle that does... something. Try it and see. ![The Mysterious ?](./assets/example_what.png) ### Training a Language Model, Live A small language model trained end to end inside the graph — Vocab Build → Text Encode → Sliding Window → the Language Model pipeline → Generate → Save. Since the live-preview update, the `Language Model Train` node streams its cross-entropy curve **under itself while it trains** (the preview card below the node). If the run is interrupted, the trainer saves its full state and a crash-site **Resume** continues from the breakpoint step instead of starting over. ![Language Model training workflow](./assets/languange_model_train_workflow.png)

Language Model Train node with its live loss-curve preview

--- ## Installation There are two ways to run ComfyDL: the GUI build, or the classic manual install into an existing ComfyUI. ### Option 1 — the GUI build: [ComfyDL_UI](https://github.com/Cynthia-lxx/ComfyDL_UI) [ComfyDL_UI](https://github.com/Cynthia-lxx/ComfyDL_UI) is a runtime fork of ComfyUI that already ships ComfyDL as a built-in node pack. No `custom_nodes` step, no manual wiring — clone it, create a virtual environment (recommended, to avoid package conflicts), install, and launch: ```bash git clone https://github.com/Cynthia-lxx/ComfyDL_UI cd ComfyDL_UI python -m venv .venv .venv\Scripts\activate # Windows (.venv/bin/activate on Linux / macOS) pip install -r requirements.txt python main.py ``` Its README covers the details: . ### Option 2 — the classic manual install into an existing ComfyUI > **This option requires ComfyUI.** If you don't have it, download from: [https://github.com/Comfy-Org/ComfyUI](https://github.com/Comfy-Org/ComfyUI) 1. Navigate to your `custom_nodes` folder: custom_nodes folder location

2. Clone this repository: ```bash git clone https://github.com/Cynthia-lxx/ComfyDL ./ComfyDL ``` 3. Install dependencies: ```bash pip install -r ./ComfyDL/requirements.txt ``` 4. Restart ComfyUI. **You should see the new ComfyDL nodes appear in the node menu.** --- ## Function Overview The built-in node library ships **231 nodes** across 35 categories — **133 provided by ComfyDL** plus 98 ComfyUI core nodes (`Network & Layers` → `Activation` 14 + `Basic` 8 + `Attention` 8 + `Normalization` 7 + `Regularization` 1 + `Training` 23 + `Pooling` 2 + `Convolution` 5 + `Recurrent` 3 + `Text` 4; `model` → `loaders` 7 + `merging` 11 + `latent` 2 + `conditioning` 2; `3d` → `Preview 3D` 1) that were added on top of it: | Category | Count | Description | |---|---|---| | **CV Models** | 5 | CNN fundamentals & model construction | | **DataOps** | 2 | | | **Datasets** | 30 | Dataset download, load, preview & stats, synthetic regression / classification data generators | | **Device Utils** | 3 | GPU/CPU device utilities | | **GAN** | 2 | GAN training updates | | **Model Utils** | 7 | Model info, mode, forward, layers, params, clone & persistence | | **NLP Models** | 9 | RNNLM & attention/Seq2Seq building blocks (RNN/GRU builders soft-archived) | | **NLP Utils** | 5 | Text tokenization & vocabularies | | **Notebook** | 1 | | | **ObjectDetection** | 10 | Anchor boxes, IoU, NMS | | **Segmentation** | 4 | VOC semantic segmentation tools | | **Tensor Basic** | 5 | Tensor I/O, conv, transpose, broadcast, activation | | **TorchOps** | 11 | Loss, optimization, metrics & the confusion matrix | | **Training** | 6 | From-scratch linear-regression & softmax-regression trainers + one-box production regression / classification trainers | | **Visualization** | 14 | Plots, charts & bounding box visualization | | **d2l/_Legacy/Model Utils** | 1 | Deprecated (soft-archived): Model Mode; use the core Training Mode | | **d2l/_Legacy/NLP Models** | 7 | Deprecated (soft-archived): Multi-Head Attention, Add & Norm, Transformer Encoder Block/Encoder; use the core Attention/Recurrent nodes | | **d2l/_Legacy/Tensor Basic** | 3 | Deprecated (soft-archived): Broadcast, Reshape, Activation | | **image (Comfy core)** | 1 | Per-channel image batch statistics | | **image/color (Comfy core)** | 3 | Grayscale, normalize & brightness/contrast/saturation | | **image/transform (Comfy core)** | 1 | Arbitrary-angle rotation + canvas expand | | **utilities (Comfy core)** | 4 | MessageBox, NoOp pass-through & benchmark timer | | **conversion** | 8 | | > The 20 categories above are the ones ComfyDL itself provides; the other 14 categories of the > shipped library (`Network & Layers/*`, `model/*` and `3d`, 98 nodes in total) are pure ComfyUI > core categories. > > For the complete node reference, see **[FUNCTIONS.md](./FUNCTIONS.md)** (English) or **[FUNCTIONS_zh.md](./FUNCTIONS_zh.md)** (中文). --- ## Repository Layout ``` ComfyDL/ ├── src/d2lcore/ # D2L-inspired core implementation — a reference layer (torch.py, ...) ├── nodes/ # ComfyUI node definitions (thin mapping layer) │ # incl. self-developed model_utils.py, image_tools.py & │ # nlp model wrappers (model_nlp.py, model_attention.py, model_seq2seq.py) └── example_workflows/ # Example workflows served as UI templates (+ _archived/ legacy set) ``` > **Note:** A mirror copy of `src/d2lcore/` also exists at the repository root as `d2lcore/` (outside the plugin folder). They must be kept in sync whenever the D2L core logic is modified. --- ## License This project is licensed under the **GNU General Public License v3.0** (or any later version) — see the [LICENSE](./LICENSE) file for details. ComfyDL is distributed as the node pack of the [ComfyDL_UI](https://github.com/Cynthia-lxx/ComfyDL_UI) project and follows the same license; upstream ComfyUI keeps its own copyright and maintainers. --- ## Special Thanks ComfyDL stands on the shoulders of the incredible **`d2l`** (Dive into Deep Learning) community. - **Codebase:** We heavily reference and adapt implementations from the [`d2l-pytorch`](https://github.com/dsgiitr/d2l-pytorch) repository. Its clear, textbook-grade code serves as a high-quality reference and starting point for many of our nodes — and continues to guide the development of new ones. We are deeply grateful to all contributors who made this resource available under the permissive **MIT-0** license. - **Inspiration:** The design and pedagogical philosophy behind this project are fundamentally inspired by the book **《Dive into Deep Learning》** (《动手学深度学习》PyTorch版), which offers one of the most accessible and practical paths to mastering deep learning. This project would not exist without their vision and generosity. We encourage everyone to explore the original book and repository: - **Book (Chinese):** https://zh.d2l.ai/ - **GitHub Repository:** https://github.com/dsgiitr/d2l-pytorch --- ![Meow~](./assets/neko.jpg)

Name	Type	Default	Description
target_type	COMBO	match link	9 options: match link, STRING, INT, FLOAT, BOOLEAN, TENSOR, +3
valueopt	*		—

Convert

Inputs (2)

Outputs (1)