通用训练设置

If the model nodes are the engine and the dataset nodes are the fuel, GeneralConfig is the dashboard you sit at before you turn the key. It's the node that collects every other config - model, dataset, optimizer, adapter, eval - and turns them into the actual TOML file the diffusion-pipe trainer runs. The whole pack is shaped like a funnel, and this node is the neck. Understand it and you understand the pack.

What it takes

Four required inputs are wire-in, not type-in: optimizer_config (from OptimizerConfigNode), model_config (from any model node), and dataset_config (from GeneralDatasetConfig). Plus output_folder_name - the name of the folder created under ComfyUI's output/ directory for this run's artifacts.

Then the hyperparameters. The ones you'll actually touch as a beginner:

• epochs (default 50) - training rounds. Keep it but watch your eval curve; more isn't automatically better.
• micro_batch_size_per_gpu (default 2) - the smallest batch unit, per GPU. This is the knob you lower when you OOM. Modern architectures dislike tiny effective batches (the KB's Flux-2-era note: "Klein 9b hates small batch sizes"), so combine it with gradient accumulation rather than just dropping it to 1.
• gradient_accumulation_steps (default 1) - effective batch = micro batch × accumulation × GPUs. Raise this to grow effective batch without OOM.
• number_of_gpus and pipeline_stages (both default 1) - set GPUs to your hardware; pipeline_stages is how many GPUs the model is split across and should match your GPU count for pipeline parallelism.
• blocks_to_swap (default 20) - how many transformer blocks get offloaded to CPU RAM to save VRAM. Bigger swap = lower VRAM, slower training. Tune against your OOM errors.
• activation_checkpointing (default on) - recomputes activations to save memory; the tooltip says "usually enable." Leave it on.
• warmup_steps (default 500), gradient_clipping (1), save_dtype (bfloat16), partition_method (parameters) - set once, rarely revisit.

The optional inputs are where the rest of the pack hooks in: adapter_config (from AdapterConfigNode - required for any LoRA run!), advanced_config (from AdvancedTrainConfig), eval_dataset_config (from EvalDatasetConfig), sampler_config, and the eval-frequency and save-frequency controls (eval_every_n_epochs, save_every_n_epochs, checkpoint_every_n_minutes, caching_batch_size, video_clip_mode, and friends).

What comes out

Three outputs. train_config (type TRAIN_CONFIG) and config_path both feed the Train node - the config_path is the written trainconfig.toml the trainer actually reads. output_dir (a plain STRING) is where your run lands inside ComfyUI's output/ folder.

How it works

Under the hood it's a TOML serializer with an opinionated memory strategy. It creates the output folder, writes every setting into trainconfig.toml, folds in your optimizer/model/dataset/adapter/eval configs, and hands the Train node everything it needs to spawn the real diffusion-pipe subprocess (with TensorBoard monitoring on top). Nothing trains until Train runs - this node just makes sure the config file is coherent first.

Installing the pack

Linux/WSL2 only, submodules mandatory:

cd ComfyUI/custom_nodes/
git clone --recurse-submodules https://github.com/TianDongL/Diffusion_pipe_in_ComfyUI.git
git submodule update
pip install -r Diffusion_pipe_in_ComfyUI/requirements.txt

Common issues

The recurring one from the community: the trainer seems to ignore your settings and uses defaults. That's almost always the config-path wiring - make sure the config_path output actually reaches Train, and remember the pack's WSL2 path convention (Z:/..., not /mnt/z/...) for everything you type. And a missing adapter_config means no adapter - "LoRA not applying" is usually just this.