JBARU/lora-training-scripts

RunPod setup for kyrael + sorelith Krea2 LoRA training

What's already done

• Both datasets are uploaded as private HF dataset repos:
  • JBARU/kyrael-dataset
  • JBARU/sorelith-dataset
• No captions yet -- setup_pod.sh captions them on the pod (Qwen2.5-VL-7B, fast on a real GPU).

Files in this folder

• setup_pod.sh -- full bootstrap: installs musubi-tuner, downloads models, downloads datasets, captions, caches, trains both LoRAs
• caption_dataset.py -- auto-captioning script (called by setup_pod.sh)
• dataset_kyrael.toml / dataset_sorelith.toml -- musubi-tuner dataset configs

Steps you need to do yourself (I can't do these for you)

1. Accept gated model access

Before anything else, visit https://huggingface.co/krea/Krea-2-Raw while logged in and accept/request access (it auto-approves). Skip this and the download in step 4 of setup_pod.sh will fail.

2. Create a RunPod account + pod

1. Sign up at https://runpod.io (this needs your own payment method -- I can't do this part)
2. Go to Pods -> Deploy
3. Pick a GPU: RTX 4090 (24GB) is the recommended sweet spot for this job (~$0.34-0.69/hr)
4. Pick a template with PyTorch + CUDA pre-installed (e.g. "RunPod PyTorch 2.x")
5. Deploy the pod, wait for it to start
6. Open its web terminal (or connect via SSH if you set up a key)

3. Get the scripts onto the pod

Once you have the pod's terminal open:

mkdir -p /workspace
cd /workspace

Then either:

• Easiest: use the RunPod web UI's file upload to drop caption_dataset.py, dataset_kyrael.toml, and dataset_sorelith.toml into /workspace/
• Or paste their contents directly using cat > filename.py << 'EOF' ... EOF in the terminal

4. Run the bootstrap

Upload setup_pod.sh the same way, then:

chmod +x setup_pod.sh
./setup_pod.sh

It'll pause at hf auth login for you to paste your token interactively -- same rule as before, paste it only when prompted, never on the command line.

5. When it's done

Trained LoRAs land in /workspace/output/kyrael/kyrael_lora.safetensors and /workspace/output/sorelith/sorelith_lora.safetensors. Push them back to HF (so you can grab them locally) with:

hf upload <your-username>/kyrael-lora /workspace/output/kyrael --repo-type model
hf upload <your-username>/sorelith-lora /workspace/output/sorelith --repo-type model

Then locally: hf download <your-username>/kyrael-lora --local-dir D:\ComfyModels\loras\kyrael

6. Don't forget to stop the pod

RunPod bills by the hour while running -- stop/terminate it once training's done so you're not paying for idle GPU time.

Vaelith run: what changed after kyrael

• Network volume: use at least 100GB, not 50GB. Kyrael's pod hit disk-full twice at 50GB -- base models alone (33GB) plus the Qwen2.5-VL-7B captioning model's cache (16GB, easy to forget about) leave almost no margin at 50GB.
• num_repeats dropped from 10 to 3 in dataset_vaelith.toml. Kyrael's LoRA came out overtrained/rigid (locked pose, completely resistant to style LoRA blending even at 5.0 weight) -- traced back to 34 images x 10 repeats x 16 epochs = 5,440 total training exposures on a small dataset. Lower repeats should fix this without hurting identity retention.
• setup_pod_vaelith.sh cleans up the captioning model cache immediately after captioning finishes (rm -rf /workspace/.cache), instead of leaving it sitting there until disk fills up mid-training like last time.

If something fails partway through

Each numbered section in setup_pod.sh is independent enough to re-run on its own -- if training crashes on kyrael, you don't need to redo the downloads or sorelith's caching, just re-run that one training command.