Eric Krea2 Loader

Every Krea 2 workflow in this pack starts here. The Eric Krea2 Loader is what turns the diffusers folder you downloaded into a living KREA2_PIPELINE - the thing every other Eric node hands around. It also sets two decisions that quietly determine everything downstream: which checkpoint you're on (Raw or Turbo) and how fast the transformer runs.

The important mental model is that this pack doesn't reimplement Krea 2. It drives the genuine Hugging Face Krea2Pipeline - the real 12B single-stream MMDiT with Qwen3-VL conditioning - loaded straight from weights. So the loader's only real job is to point at those weights and pick sensible runtime knobs. The model_path input is the whole ballgame: it defaults to H:\Training\Krea-2-Raw, a path that lives on Eric's Windows box, not yours. Point it at wherever your diffusers-layout Krea 2 folder actually is, whether that's Krea-2-Raw (the undistilled base, guidance-hungry, the one you train LoRAs on) or Krea-2-Turbo (the 8-step distilled one). The node auto-detects the is_distilled flag from the folder's model_index.json, and that flag decides the whole sampling regime - Turbo gets guidance forced low and a fixed shift, Raw gets real CFG. Mismatch them and Turbo renders blown and oversaturated.

The rest of the required inputs are runtime choices:

• precision - bf16 is the default and the right answer on Blackwell; fp16 and fp32 exist but you'll rarely want them.
• attention_backend - auto picks FlashAttention if it's installed, otherwise SDPA. On a Blackwell card this is roughly a 3x speedup and it's lossless, so auto (or explicit flash) is the obvious choice. sage is fastest but trades a hair of quality. sdpa is the safe fallback.
• device - cuda, or cuda:1 if you want the pipeline off your primary GPU (say you're also running an LLM for Magic Prompt).
• keep_in_vram - defaults on, caching the assembled pipeline between runs so you're not reloading 24 GB of weights every queue. Leave it on.

The single output, krea2_pipeline, wires into Eric Krea2 Generate for a quick text-to-image, or into Multi-Stage Ultra when you want the multi-stage high-res path. If you have presets, there's no dropdown on this node - the Component Loader is the version with the preset system, in case you keep several checkpoints around.

How to install

This is part of the EricRollei/Krea2_ComfyUI_Advanced pack, so it installs with the whole pack. Easiest route is ComfyUI Manager - search "Krea2_ComfyUI_Advanced" and install. Manually:

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

Then restart ComfyUI. Two dependencies won't be handled for you, and both are mandatory:

python_embeded\python.exe -m pip install --upgrade --force-reinstall --no-deps git+https://github.com/huggingface/diffusers.git
pip install "transformers>=4.57.0"

The diffusers install is the sneaky one: Krea2Pipeline merged after the 0.39.0.dev0 tag and the version string never got bumped, so a plain pip install diffusers looks up to date and isn't. That's why --force-reinstall is required. And the pack does not download Krea 2 weights for you - grab krea/Krea-2-Raw and krea/Krea-2-Turbo yourself.

Common issues

• The default path doesn't exist. Obviously - it's Eric's. Change model_path to your folder or nothing loads.
• Krea2Pipeline not found / import errors. You installed stock diffusers instead of the git build. Run the --force-reinstall command above.
• Turbo looks blown out. The is_distilled flag is being read from a Raw base folder (or vice versa). Load the right checkpoint and the whole sampling regime snaps into place.
• Slow generations. attention_backend is falling back to SDPA because FlashAttention/SageAttention isn't installed. On Blackwell these are optional but recommended, and they're the difference between "works" and "fast".