SD3 Train Validation Settings

This is the SD3 version of FluxTrainer's validation-settings node. It defines what the sample images look like when your SD3 training run pauses to preview itself - resolution, sampling steps, guidance, and the seed. You wire its output into the training/validation loop, which then renders check-in images using these settings at whatever interval you set.

It's the same idea as the Flux validation-settings node, trimmed for the SD3 path. Nothing about it changes how the model trains; it only controls the previews you look at along the way.

Why it's worth wiring in

The point of validation is to watch the LoRA learn instead of training blind and finding out at the end. And the single thing that makes previews actually useful is a fixed seed - with the seed pinned, every checkpoint generates from the same noise, so flipping between step 500 and step 1,500 shows you the LoRA changing rather than the seed changing. Leave the loop on its built-in defaults and the previews float; connect this node and pin the seed and they line up. That's most of the value right there.

The inputs that matter

• seed - fix it (default 42) so checkpoints are comparable. The most important setting here.
• width / height - preview resolution, defaulting to 1024×1024 (SD3's native resolution). Fine to keep, though smaller previews render faster and cost less training time if you're validating often.
• guidance_scale - 4 by default. This is your CFG for the sample images; the default is a reasonable middle. Nudge it if previews look over- or under-baked, but remember that's a preview setting, not a fix for the LoRA.
• steps - sampling steps per preview, default 20. Enough to judge progress without wasting time.

Output is validation_settings (type VALSETTINGS), which plugs into the validation_settings input on the train/validate loop.

Installing ComfyUI-FluxTrainer

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

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. For SD3 training you'll also need the SD3 model files (transformer plus the three text encoders) in place - see the SD3ModelSelect node.

Common issues

The usual mistake is skipping this node and letting the validation loop use defaults, then finding the previews don't tell you anything because the seed keeps moving. If you want to read your training by eye, connect this and pin the seed. Simple, and it's the whole reason the node exists.

The second is confusing preview settings with training settings. Guidance 4, 1024×1024, 20 steps - none of that touches the weights. If your samples look bad, first decide whether the problem is the LoRA (dataset, captions, learning rate, step count - the real levers) or just an aggressive preview setting. Don't tweak validation guidance hoping to fix a training problem.

And the honest context: SD3 never caught on the way SDXL and Flux did after its bumpy release, so the SD3 path in this pack sees little traffic. It works, but there's far less community wisdom to lean on than for Flux. Keep validation cheap, pin the seed, save intermediate checkpoints, and trust the samples over any default - the pack is experimental, and its defaults are just what worked once for the author.