Eric Krea2 Component Loader

The plain Eric Krea2 Loader gets you running. The Component Loader is for when you want to run your way: it loads the same Krea2Pipeline, but every major component - transformer, text encoder, VAE - can be swapped for a different file, and the whole recipe can be saved as a named preset and recalled with one click.

Start with base_pipeline_path, which is the full Krea 2 diffusers folder. It supplies the model_index.json, scheduler, tokenizer, and any component you don't override. Two things to know about it. First, the default is H:\Training\Krea-2-Turbo - again Eric's machine, so change it to your checkpoint folder. Second, and this matters more than any other knob here: the is_distilled flag in that folder's model_index.json decides the entire sampling regime. A Turbo checkpoint run against a Raw base renders blown and oversaturated, so make sure your base matches the checkpoint you intend to use.

Then come the overrides. Each has a Pick dropdown (populated from ComfyUI's model folders) and a Path field, and the dropdown wins:

• transformer_pick / transformer_path - swap the DiT itself. Single-file .safetensors/.gguf for the pick; a full diffusers folder via path. The .gguf gets dequantized on load.
• text_encoder_pick / text_encoder_path - override the Qwen3-VL encoder, same story.
• vae_pick / vae_path - swap the AutoencoderKLQwenImage VAE. This is how people run the FP8 conversion or a different VAE build without editing the base folder.
• precision - bf16 default; GGUF is dequantized to whatever you set here.
• attention_backend - auto (FlashAttention if present, else SDPA), flash, sage, or sdpa. auto is ~3x faster than sdpa on Blackwell and lossless; sage is fastest with a slight quality trade.
• device - cuda, cuda:0, cuda:1, or cpu.
• keep_in_vram - cache the assembled pipeline between runs; leave on.
• offload_vae - keep the VAE on CPU during transformer inference to save ~1 GB of VRAM. Worth it on smaller cards.
• vae_tiling / vae_tile_px - tiled decode/encode policy (auto tiles above 1024 px). off is a diagnostic for tile-seam hunting; the tooltip's trick - if the pattern pitch moves when you change vae_tile_px, it's tiling - is genuinely how you tell seam artifacts from generation artifacts.

The loader_preset dropdown is where this node earns its keep over the plain Loader. Pick a saved recipe and the paths, precision, attention backend, and device all load in. The ★ Save button writes the current panel to loader_presets.json, and the dropdown refreshes on graph reload. The samples shipped in the preset list (DarkBeast-..., JibMixV10, Kreamania bf16) show this is a node built by someone who A/Bs fine tunes for a living - you're inheriting his habits, which is fine, but treat them as starting points.

Outputs are krea2_pipeline (into Generate or Multi-Stage Ultra) and settings, a string recipe that Merge Settings can fold into your saved PNG metadata so the whole loader state is reproducible later.

How to install

Same pack as everything else: ComfyUI Manager → search "Krea2_ComfyUI_Advanced", or

cd ComfyUI/custom_nodes
git clone https://github.com/EricRollei/Krea2_ComfyUI_Advanced

then restart. Requires a diffusers build that includes Krea2Pipeline (install from git with --force-reinstall, since the version string never bumped past 0.39.0.dev0) and transformers>=4.57.0 for Qwen3-VL. Weights are yours to supply - this pack drives the real Hugging Face pipeline, it doesn't bundle the model.

Common issues

• Blown/oversaturated Turbo renders. Base path points at a Raw folder (or the is_distilled flag is being read wrong). Load the matching checkpoint.
• Presets not showing up. The dropdown refreshes on graph reload - restart or reload the workflow.
• VAE seams in big decodes. Leave vae_tiling on auto; only flip to off to diagnose. Big non-tiled Wan-family decodes can leave a bottom-edge band and spike VRAM.