⚗ Alchemica · Load Krea 2 [1.0, use Alchemica Load Krea 2 (model + clip + vae)]

Krea 2 is not a drop-a-checkpoint-in case. It wants a diffusion model, a text encoder, and a VAE, and stock ComfyUI makes you load all three from three separate nodes - UNETLoader, CLIPLoader, VAELoader. Worse, CLIPLoader has a type dropdown with a couple dozen entries in it, and exactly one of them works. This node exists to make that failure impossible.

What it actually does

It is a thin wrapper. Under the hood it calls ComfyUI's own UNETLoader, CLIPLoader and VAELoader and hands you back the three objects - so it isn't a reimplementation and it inherits whatever behaviour core ComfyUI has. The one thing it hardcodes is the part people get wrong: the text encoder is loaded as type krea2, always, with no option to change it. The author's tooltip says it plainly, and it's worth repeating because the error it prevents is cryptic - load the Qwen3-VL encoder as qwen_image or flux and Krea 2 dies with a conditioning-size error that reads like a corrupt file rather than a wrong dropdown.

You get three outputs: model, clip, vae.

• model goes to your sampler - AlchemicaKrea, AlchemicaKreaPro, the advanced AlchemicaSampler, or a plain KSampler.
• clip goes to anything that encodes text. For this pack that means ⚗ Prompt enhancer, ⚗ Text, or ⚗ Two-phase prompt - and note that the enhancer both writes and encodes with this same encoder.
• vae goes to VAEDecode (and VAEEncode if you're doing img2img).

The fields worth caring about

Four widgets, no more.

modello - the diffusion model, from models/diffusion_models. The default in a fresh install is krea2_turbo_nvfp4.safetensors; if you downloaded the FP8 conversion instead you'll pick something like krea2_turbo_fp8_scaled.safetensors. Krea 2 Raw works too, and so do the community finetunes and Krea2 checkpoints that flooded Civitai within days.

pesi - the weight dtype, and default is the right answer for almost everyone: it leaves quantised files (NVFP4, FP8) alone and only casts plain FP16/BF16 weights. The fp8_e4m3fn options exist for the case where you deliberately want to cast down a non-quantised file.

text_encoder - the Qwen3-VL-4B encoder, from models/text_encoders. If this dropdown is empty, the file isn't where ComfyUI looks.

vae - the Qwen-Image VAE, from models/vae.

That's it. Three files, one node, three wires out.

Installing

ComfyUI Manager: search AlchemicaSampler and install. Or by hand:

cd ComfyUI/custom_nodes
git clone https://github.com/Nessuno76/ComfyUI-AlchemicaSampler

Then restart the server. Reloading the browser tab does not register new nodes - this trips people up constantly with this pack, because the nodes genuinely do not appear until the process restarts.

There is nothing to pip install. The pack's pyproject.toml declares no dependencies, and it only imports ComfyUI itself plus torch. The real requirements are ComfyUI 0.36.0 or newer (this loader is fine on older builds, but the prompt enhancer isn't - see that node) and the three model files in models/diffusion_models, models/text_encoders, models/vae.

Where people get burned

Empty dropdowns. The node lists whatever is in the standard model folders. If you dropped qwen3vl_4b_fp8_scaled.safetensors into models/clip instead of models/text_encoders, the text_encoder field will offer you a bunch of unrelated CLIP files and the run will fail somewhere confusing.

NVFP4 that runs slower than FP8. This is a PyTorch build problem, not a node problem - 4-bit on RTX 50-series needs a CUDA 13.0 (cu130) PyTorch. You can confirm which build you're on from the ⚗ Calibration report, which prints your torch and CUDA versions.

The conditioning-size error. Nine times out of ten it's somebody using a stock CLIPLoader set to the wrong type in an old workflow. Swap in this node, or set the type manually to krea2.

Missing VAE symptoms. A wrong or absent VAE doesn't usually crash - it gives you washed-out or over-smoothed images, which is easy to misread as "Krea 2 is just soft". Krea chose the Qwen-Image VAE deliberately for non-photoreal work, and said on the record that Flux's VAE is the photorealism pick; but a different model's VAE loaded here is a different thing entirely and will wreck your colour.