SD3模型配置器

Stable Diffusion 3 isn't the headline act it was a couple of years ago, but it's still around, it still has a LoRA ecosystem, and if you've got SD3 weights on disk this node is how you train against them inside the pack. SD3ModelNode is about as simple as a model picker gets: one required path, one optional toggle, done.

The support table marks SD3 as LoRA yes, full fine-tune no - so this node is LoRA-focused, which for SD3 is the reasonable position. It's also one of the fp8-friendly models in the table, which matters if your SD3 training is VRAM-bound.

How it works

It builds a {"type": "sd3", "diffusers_path": ...} config and hands it to ModelConfig for the dtype/timestep merge. Nothing is loaded yet - the deepspeed subprocess launched by Train does the real work. The one optional field is flux_shift, default off: enabling it applies the resolution-dependent timestep shift that Flux uses, which some SD3 training recipes borrow to improve convergence at different resolutions. It's a "try it and see" knob, not a "must be on" one.

The diffusers_path requirement is worth reading closely: the tooltip says it needs a complete Diffusers folder (the example is stable-diffusion-3-medium-diffusers). Not a single checkpoint file, not a half-downloaded directory - the full structure with model_index, transformer, text encoders, VAE, and tokenizer pieces. This is the most common way people trip on this node.

The inputs that matter

• diffusers_path - required, the complete SD3 diffusers model folder.
• flux_shift - optional, off by default; the Flux-style resolution-dependent timestep shift.

Output: model_path → ModelConfig.

Install

Same pack-wide story - ComfyUI Manager (search "Diffusion_pipe_in_ComfyUI") or:

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

Linux/WSL2 only, heavy deepspeed/diffusers deps included.

Where people get burned

The "complete diffusers folder" requirement is the whole game. Point this at a single .safetensors or a partial download and the training subprocess fails with a load error that doesn't obviously point back at the path field. And don't sleep on the fp8 angle - the pack supports fp8/quantization for SD3, and combined with ModelConfig's float8 transformer dtype, that's the difference between fitting and not fitting on a modest card. On the flux_shift front: default off is a sensible default; switch it on only if you're chasing specific resolution behavior, not as a routine improvement.