Weialbert2/jade-hadi-dual-bootstrap

Jade Hadi dual-LoRA RunPod job

This job trains two supplemental character LoRAs from the accepted dataset_krea2_v1 set:

1. Z-Image Turbo for the Matrix/Hadi image workflows.
2. Wan 2.1 14B for the Matrix/Hadi Wan video workflows.

Both runs go to 3,000 steps and save every 250 steps. All intermediate checkpoints are retained so the winner can be selected by a fixed-seed test; the final checkpoint is not assumed to be best.

The existing Krea 2 LoRA remains Jade's main image LoRA. The two outputs from this job do not replace it.

Locked input

• Local source: ../dataset_krea2_v1
• Expected: 38 images and 38 paired captions
• Trigger: j4d3ln
• Rejected ../jadeified_v1 images are never uploaded or used.

RunPod storage

The training Pod uses a 150 GB local volume mounted at /workspace, matching the Hadi/Matrix AI Toolkit recipe. It deliberately does not attach the 500 GB network volume. The network volume remains available for persistent ComfyUI models and workflows; AI Toolkit training uses the separate local Pod volume.

Repositories

• Dataset: Weialbert2/jade-hadi-lora-input-38
• Z-Image output: Weialbert2/jade-zimage-lora-hadi
• Wan output: Weialbert2/jade-wan21-lora-hadi
• Public bootstrap: Weialbert2/jade-hadi-dual-bootstrap

Launch

Prepare the private input/output repositories and publish the bootstrap:

python3 prepare_hf.py

Launch one RTX 5090/RTX Pro 6000 Pod. It trains Z-Image first, then Wan, uploads checkpoints continuously, and terminates itself:

python3 launch.py

The launcher writes the Pod id to ~/.runpod-hadi-dual-podid.