SD3 Model Select

FluxTrainer isn't only for Flux - it also has an SD3 (Stable Diffusion 3 / 3.5) training path, and this is the node that loads the model files for it. SD3 splits its model across several separate files, and instead of one checkpoint dropdown you pick each piece: the main transformer and the three text encoders. This node gathers them into a single bundle the SD3 trainer can consume.

If you've only ever loaded a monolithic SDXL checkpoint, the multi-file layout can be a surprise. It's not this node being awkward - it's just how the SD3 stack is packaged.

How it works

SD3 uses a diffusion transformer plus a trio of text encoders: two CLIP models and a big T5. SD3ModelSelect exposes a dropdown for each so you point at the exact files you've placed in your ComfyUI model folders, then emits a sd3_models object (type TRAIN_SD3_MODELS) that feeds the SD3 training-init node. Bundling all four selections into one output keeps the training graph tidy - one wire carries the whole model stack forward.

The inputs that matter

Four required file pickers, and they're the whole node:

• transformer - the SD3 diffusion model (the MMDiT transformer). The thing you're actually training against.
• clip_l and clip_g - SD3's two CLIP text encoders.
• t5 - the T5-XXL text encoder, the big one that handles natural-language prompts.

Each is a dropdown populated from the files you've dropped into ComfyUI's model directories. There's also an optional lora_path (a text field) if you want to start from an existing LoRA rather than from scratch. Output: sd3_models.

A practical note on those dropdowns: they list whatever model files ComfyUI can see. If a picker is empty, ComfyUI isn't finding files of that type in the expected folder - that's a placement problem, covered below, not a broken node.

Installing ComfyUI-FluxTrainer

ComfyUI Manager: search "ComfyUI Flux Trainer", install, restart. Or clone it:

cd ComfyUI/custom_nodes
git clone https://github.com/kijai/ComfyUI-FluxTrainer
cd ComfyUI-FluxTrainer
pip install -r requirements.txt

Restart ComfyUI. Torch 2.4.0 or newer recommended. Example workflows use ComfyUI-KJNodes.

You'll also need the SD3 model files themselves - the transformer and all three text encoders - placed where ComfyUI expects them so the dropdowns populate. As with Flux LoRA training, use full-precision (fp16) or the standard fp8 weights, not diffusers-repackaged variants.

Common issues

Empty dropdowns. The most common snag: you open the node and a picker has nothing in it. That means ComfyUI didn't find that file type in its model folders. Confirm the transformer and each text encoder are actually downloaded and sitting in the right directory, then restart ComfyUI so it re-scans. You need all four present - SD3 won't train missing a text encoder.

Right files, right roles. Because you're choosing four things by hand, it's genuinely possible to put a file in the wrong slot - a CLIP where T5 goes, say. Double-check each dropdown points at the file it's labelled for.

Manage expectations on SD3 itself. SD3's launch was rocky and the wider community never adopted it the way it did SDXL or Flux, so SD3 training here is a minority path within an already-niche pack. FluxTrainer's SD3 support exists and works, but you'll find far fewer worked examples and community configs than for Flux or SDXL. If you're training SD3 specifically, expect to do more of your own experimenting - and keep kijai's standing caveat in mind that this whole pack is experimental and the default settings are just what happened to work for him, not a tuned recipe.