Init SD3 LoRA Training

The SD3/SD3.5 sibling of the Init family. If you're training a LoRA for Stable Diffusion 3.5 and you'd rather do it in ComfyUI than spin up Kohya separately, InitSD3LoRATraining is the node that configures and launches the run. It's the least-trafficked of the three Init nodes for a reason - SD3.5 never grew the training community that Flux and SDXL did - but if that's your base model, this is your node.

Like the rest of the pack, it wraps kohya-ss/sd-scripts, so the knobs and their names are Kohya's. SD3.5 shares Flux's DNA in the ways that matter to a trainer: it's a flow-matching model with a T5 text encoder alongside the CLIP pair, which is why you'll see flow-shift and T5 options here that SDXL doesn't have.

How it works

You feed it three required inputs - sd3_models (type TRAIN_SD3_MODELS), a dataset JSON, and optimizer_settings from an OptimizerConfig node - and it builds a trainer. The output network_trainer (NETWORKTRAINER) goes into a FluxTrainLoop node to run steps, exactly like the Flux and SDXL paths; the loop machinery is shared. You also get epochs_count and the raw Kohya args.

The inputs that matter

• network_dim (default 16) and network_alpha (default 16) - a reasonable 1:1 starting point.
• learning_rate (default 1e-4) - a sensible default for this architecture family.
• max_train_steps (default 1500) - scale to your dataset.
• training_shift (default 3) - the flow-matching timestep shift, specific to SD3-style models. The default is a fine starting point; it's not a knob to fiddle with blind.
• train_text_encoder - this one exposes the full menu (clip_l, clip_l+T5, and fp8 variants). On modern flow-matching bases you generally leave the text encoders frozen and just cache their outputs, so disabled is the safe default. Caching (cache_text_encoder_outputs) is also your biggest speed win here.
• blocks_to_swap and fp8_base - the usual VRAM levers.

Installing the pack

Via ComfyUI Manager: search ComfyUI Flux Trainer, install, restart. Or manually:

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

Torch 2.4.0+ is recommended, and the examples pull in kijai's ComfyUI-KJNodes. Use the normal fp8/fp16 SD3.5 model files, and - as with Flux - make sure any VAE you supply is the standard non-diffusers safetensors, not the diffusers folder version.

Common issues

Fewer guides exist for SD3.5. This is the honest one. The training community largely skipped SD3.5, so you'll find far fewer copy-paste recipes than for SDXL or Flux. Lean on the pack's example workflows and the general flow-matching intuition (frozen encoders, cache the outputs, moderate learning rate) rather than hunting for an SD3.5-specific settings table that may not exist.

Dataset setup is the same trap as everywhere else. A bad or empty dataset JSON throws an opaque parse error. Build your dataset with the pack's dataset node, keep captions in .txt files next to the images, and make sure the folder path resolves before you start.

It's experimental. Kijai's README is blunt that the whole thing is work-in-progress and that his defaults aren't necessarily good. Validate on a short run before committing hours of GPU time.