VRGDG Krea 2 LoRA Studio

VRGDG Krea 2 LoRA Studio is the front-end for the pack's Krea 2 training workflow. It opens the Studio UI, which bundles the whole train-and-judge loop into one window: preset-based chunk training, sampling with the result, and comparison grids so you can actually see how the LoRA is evolving instead of guessing from loss curves. It's a launcher node with two small inputs and two outputs that mostly echo where things live.

Why "chunk training"? The pack's Krea 2 approach trains in chunks - fixed step slices - saving a LoRA and state at each chunk's end. That's the iteration-friendly model: sample after a chunk, look at the grid, decide whether to keep going. Training a fresh architecture like Krea 2 blind is a recipe for burning a weekend and ending up with an overfit mess; the Studio's loop is built to keep you looking at output, not logs. The KB's LoRA material backs this up hard: intermediate saves are worth more than the final one, because the last epoch is usually not the best.

The inputs

• project_root - the parent folder where Studio projects get created (default A:/MUSUBI/Training/Krea2Studio).
• project_name - the name of the session folder the UI creates (default Krea2Studio).

The outputs

• project_root and project_name - echoes of the two inputs, handy if you want the rest of the workflow to know where the project lives.

Installing it

Pack install as always - Manager, search vrgamedev, or:

cd ComfyUI/custom_nodes
git clone https://github.com/vrgamegirl19/comfyui-vrgamedevgirl.git

Restart and hard-refresh the browser. Launcher/UI nodes live entirely in the frontend, so a stale tab means a dead button. Before the Studio is useful you'll also want the backend pieces it drives: a Krea 2-ready Musubi install (via VRGDG Krea 2 Musubi Installer) and the training models on disk.

Common issues

• Clicking the node does nothing. Stale frontend - hard-refresh. If it still won't open, check that the pack's JS loaded (same fix, harder press).
• The Studio opens but training errors out. The Studio is a UI over the training stack; a missing Musubi install or wrong Python (needs 3.10–3.12) will surface as errors inside the Studio. Fix the environment, not the UI.
• Comparison grids look useless. Grids show whatever you sampled - if you sampled one preset at one step count, there's nothing to compare. The preset system is the point: sample the same prompt across presets or chunk points.

If you're going to train Krea 2 LoRAs through this pack, this is the window you'll live in. Treat the preset presets as starting points - Krea 2 is new enough that "recommended" is still being discovered, and the Studio exists so you can find what works for your dataset.