VRGDG Krea 2 LoRA Train Chunk

VRGDG Krea 2 LoRA Train Chunk is the workhorse: it runs one Krea 2 LoRA training chunk through musubi-tuner and hands you the path to the trained .safetensors - in native Krea 2 form, no Comfy conversion attempted. "One chunk" is the design: you train a slice of steps, get a LoRA and a state save, then decide whether to continue. It's the node the LoRA Studio UI drives under the hood, and you can also wire it directly into a graph.

This is the modern LoRA training reality in one node. The old SDXL-era rules don't port to a 12B Qwen3-VL-based model - you don't train the text encoder, you cache embeddings once, and the rank/alpha tables people memorized for SDXL are wrong here. Krea 2's own docs (which the node's defaults follow) recommend rank 32, alpha 32, fp8 base weights, and a shift timestep sampling at 2.5. The node's defaults are the author's reading of those docs, which is the right place to start.

The inputs that matter

• dataset_images_dir - your training images (or a parent that gets organized into an images subfolder). It's everything: curation beats every other knob.
• workspace_dir - working folder for caches, logs, configs, checkpoints, and training state.
• steps_per_run (250) and total_target_steps (3000) - one chunk's length and when training stops. Lower steps_per_run = more frequent saves, easier iteration.
• resolution_width/height - 1024 default, matching Krea 2's native inference resolution.
• network_dim / network_alpha - rank and alpha, both 32 per Krea 2 docs.
• blocks_to_swap - how many Krea 2 main blocks to offload to CPU (default 4, max 26). This is your VRAM lever: on a tight GPU, raise it; it slows training but keeps you under the ceiling.
• fp8_base / fp8_scaled - both on. Krea 2 requires fp8_scaled whenever fp8_base is enabled, so keep them in lockstep.
• learning_rate_preset - quick picks (1e-4 default) or Custom to use the float below. 1e-4 is hot for a 12B; the KB's modern-model experience suggests going lower (5e-5, 3e-5) is often safer.
• cache_strategy - auto builds the cache only when needed, force always rebuilds, skip jumps straight to training. First run uses auto and pays the cache cost once.
• create_captions / caption_text / add_trigger_word / trigger_text - caption handling. Caption what should vary; leave fixed things undescribed. A trigger word helps but isn't magic.
• musubi_root, krea2_raw_dit, vae, text_encoder - the environment paths. These are exactly what VRGDG Krea 2 Musubi Installer outputs, so wire them through rather than typing them.

The outputs

• latest_lora_path - the trained native Krea 2 LoRA. This is the one you actually want.
• latest_state_path - training state, for continuing a chunked run.
• log_path, output_name, completed_steps, total_target_steps - run bookkeeping.

Installing it

Pack install (Manager → search vrgamedev, or git clone https://github.com/vrgamegirl19/comfyui-vrgamedevgirl.git), restart, hard-refresh. This node does not install the training stack - you need the Musubi installer to have run, with a Python 3.10–3.12 environment, or it will fail at first contact.

Common issues

• Errors before training starts. Almost always the environment: musubi not installed, wrong Python, or model paths that don't exist. Verify with the installer's report_path and the paths it output.
• OOM mid-run. Raise blocks_to_swap, or free VRAM first (the clear_memory_before_text_encoder toggle helps at the cache step). Drop resolution if it persists.
• Loss looks weird / output is garbage. Krea 2 training is young. Sample what you train (VRGDG LoRA From Path Model Only is made for this), compare chunks, and stop when the last saved chunk is worse, not the number of steps.

This is not a beginner's first training node - it assumes you have a stack stood up and a dataset worth the compute. If you do, it's the cleanest way to get a native Krea 2 LoRA out of ComfyUI.