AuraFlow模型配置器

AuraFlow is one of the most interesting "also-rans" in the diffusion world, and this node is how you train on it from ComfyUI. Falcon's 6.8B rectified-flow model was the answer to the SD3 license collapse - Apache 2.0, genuinely good - and its v0.2 beat Flux Dev on prompt adherence for a while. Then Flux landed, the trainers dropped it, and AuraFlow mostly faded. Its one big second act: AstraliteHeart picked it as the base for Pony V7, which is why the tooltip's example path points at a pony-v7-base.safetensors. If you're training a LoRA on a Pony V7 derivative, this node is how you wire the base model in.

What it does

Like every model node in this pack, AuraFlowModelNode is a path collector with opinions. It takes full filesystem paths to the three pieces of the model and packages them into a config dict - the model_path output - that plugs into the model_config input on GeneralConfig. The node doesn't load weights and can't test them; it just records where they are so diffusion-pipe can find them when the Train node kicks off.

The required inputs are the three files that make up the model:

• transformer_path - the main AuraFlow model file (.safetensors). This one's non-negotiable; the source bails with an error if it's empty.
• text_encoder_path - the text encoder file.
• vae_path - the VAE file.

Then there's max_sequence_length (default 768, range 128–8192 in steps of 128), and the tooltip carries the exact guidance you need: Pony V7 wants 768, base AuraFlow wants 256. That's the author flagging the real difference between the two lineages - if your prompt adherence or training quality looks off, check this value first. Set it to whatever your checkpoint's README says, not the default.

How it fits the workflow

The chain is always the same in this pack: model node → GeneralConfig → Train. You'll also need an AdapterConfigNode set to lora (AuraFlow is LoRA-only here - the README's compatibility table shows LoRA ✅, full fine-tune ❌), a dataset config, and an optimizer config. If a path is wrong, the failure shows up at training start, not in this node, so double-check the full paths - including the WSL2 drive-letter quirk (Z:/... rather than /mnt/z/...) if your models live on a Windows drive.

Installing the pack

This is the standard install for TianDongL/Diffusion_pipe_in_ComfyUI - Linux/WSL2 only (there's a separate Windows repo), and the git submodules are mandatory, not optional:

cd ComfyUI/custom_nodes/
git clone --recurse-submodules https://github.com/TianDongL/Diffusion_pipe_in_ComfyUI.git
git submodule update    # without this, training will not work
pip install -r Diffusion_pipe_in_ComfyUI/requirements.txt

The requirements include deepspeed, so expect a chunky install and a PyTorch/CUDA version that matches your GPU - the README tells you to pick your own torch combination rather than trusting the defaults.

Common issues

Mostly path errors, because there's nothing else to this node. "Model not found" at training time means a path is wrong, a file wasn't downloaded, or you used a Linux-style path where the pack expects a drive letter. And if you grabbed a non-Pony AuraFlow checkpoint, remember to drop max_sequence_length to 256 - it's the single most likely cause of subtly bad training on this model and the easiest one to miss.