Realtime LoRA Trainer (SD 1.5 - sd-scripts)

Everyone moved on to Flux and Z-Image for the interesting stuff, but SD 1.5 is still the fastest thing you can train a LoRA on - and this node makes that speed usable inside a live ComfyUI graph instead of a separate CLI trainer you have to babysit.

What it is and why you'd reach for it

This is one of ten trainer nodes in ShootTheSound/comfyUI-Realtime-Lora, "Realtime LoRA Trainer" - a pack whose whole pitch is closing the loop between training and using a LoRA without leaving your workflow. The pack author calls SD 1.5 out specifically as underrated: on a modern GPU, 500 steps can finish in around two minutes, fast enough that "train a quick LoRA, then immediately test it in the same workflow" is genuinely practical rather than aspirational. That makes this node less about chasing SOTA quality and more about rapid iteration - testing whether a concept, a face, or a style is even learnable before you commit to a longer run on a bigger model.

Training here runs on Kohya's sd-scripts under the hood, the same engine most SD 1.5/SDXL LoRA tutorials on the internet are written against - so if you outgrow this node's exposed settings, the vocabulary (rank, alpha, learning rate) transfers directly to running sd-scripts by hand.

Inputs and outputs that matter

• images_path - point it at a folder of training images (with .txt captions for batch training), or use inputcount plus the image_1-image_4 / caption_1-caption_4 optional slots for a small hand-picked set.
• sd_scripts_path - where sd-scripts is installed (default ~/sd-scripts). The node calls into that install; you never open it directly.
• ckpt_name - a dropdown of SD 1.5 checkpoints ComfyUI already knows about.
• caption - the default caption used when you haven't supplied per-image ones; rare trigger words still apply here the same as anywhere else.
• training_steps (default 500), learning_rate (default 0.0005), lora_rank (default 16) - SD 1.5-era defaults, and unlike the newer flow-matching architectures in this pack, SD 1.5's settings tables are the settled, well-trodden ones - this is the one model here where a generic "rank 16-32, LR 1e-4 to 3e-4" internet guide still roughly applies.
• vram_mode - three presets, Min (256px) up to Max (768px); the shipped default is Low (512px).
• output_name, keep_lora - filename and whether to retain the output after the run.
• Output: lora_path - a string, feed it straight into lora_path_opt on the pack's other SD 1.5 nodes (the SD 1.5 Model Layer Editor, or a standard LoRA loader) or the Apply Trained LoRA node.

Installing it

The node comes with the pack. ComfyUI Manager: search "Realtime LoRA Trainer." Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/ShootTheSound/comfyUI-Realtime-Lora

Restart ComfyUI. That's the whole install for using the pack's loaders and analyzers. Actually training needs sd-scripts installed separately (kohya-ss/sd-scripts on GitHub - follow its own instructions), and the README's one Python-version warning applies here too: 3.10-3.12, with 3.10 the safest bet and 3.13 explicitly not recommended yet.

Troubleshooting

ckpt_name dropdown is empty. It only lists what's already in your checkpoints folder in a format ComfyUI recognizes as SD 1.5 - an empty dropdown means nothing matched, not that the node is broken.

Training either errors immediately or silently does nothing. Almost always sd_scripts_path pointing at a folder that isn't actually a working sd-scripts install, or a Python version mismatch - the README specifically calls out 3.13 as untested territory for both sd-scripts and AI-Toolkit training paths.

The trained LoRA doesn't look like anything changed. At the SD 1.5 scale of dataset (15-25 images for a character), 500 steps at the default settings is a reasonable starting point, but a genuinely undertrained result usually means too few steps or too low a learning rate for the dataset size - raise steps before touching anything else, since SD 1.5 tolerates more steps per image than the newer, larger architectures in this pack do.

Repeated runs on the same inputs seem to skip training entirely. That's a feature, not a bug - the pack caches automatically, so identical inputs reuse the existing LoRA instead of re-training from scratch.