⚗ Alchemica Load Krea 2 (model + clip + vae)

Krea 2 needs three files, not one. The community's oldest Krea 2 mistake is grabbing the diffusion model, forgetting the ~8 GB Qwen3-VL text encoder and the Qwen-Image VAE, then posting that the model is broken. The second-oldest mistake is loading the text encoder with the wrong type, which fails with a conditioning-size error that explains nothing.

This node exists to make both of those impossible. Model, text encoder and VAE in one node, with the encoder always loaded as type krea2.

What it's doing

Under the hood it's exactly three core nodes stacked: UNETLoader for the diffusion model, CLIPLoader with type="krea2" for the text encoder, VAELoader for the VAE. Nothing clever, nothing quantised behind your back. The value is that the type is hardcoded - Krea 2 will not run on a text encoder loaded as qwen_image or stable_diffusion, it throws a conditioning-size error and stops, and the hardcode is the cure.

The four dropdowns read live from your ComfyUI model folders, so what you see is what you have:

• unet_name - from models/diffusion_models. The default points at krea2_turbo_nvfp4.safetensors if it's there.
• weight_dtype - the same list UNETLoader offers, default out of the box. Leave it unless you know your card wants something specific.
• clip_name - from models/text_encoders. You want qwen3vl_4b_fp8_scaled.safetensors. This one is always loaded as type krea2.
• vae_name - from models/vae. The Qwen-Image VAE, qwen_image_vae.safetensors. Around 250 MB.

Outputs are model, clip, vae, and that's the point: one node, three clean wires. model goes into your LoRA stack and then a sampler; clip goes to whatever writes or encodes your prompt (the pack's Prompt node, Text Encode, Prompt Enhancer, or the Hybrid/Face Blend nodes, which need a vision-language text encoder specifically); vae goes to the sampler's optional vae socket or a VAE Decode.

Which Krea 2 to load

Turbo is what most people should start with. It's guidance- and timestep-distilled: 8 steps, CFG 1, mu 1.15, renders up to 2K. The NVFP4 build is the small one and the pack's starter workflow is built around it; krea2_turbo_fp8_scaled.safetensors is the alternative, slightly bigger, and the pack's own comparison used FP8 as the reference for how much error NVFP4 introduces.

16 GB of VRAM is enough for 4 MP with NVFP4. If NVFP4 runs slower than FP8 on your 50-series card, your PyTorch isn't a cu130 build - that's a torch problem, and the pack's Calibration node will actually tell you so.

Installing

# ComfyUI Manager: search "AlchemicaSampler"
# or
comfy node install comfyui-alchemicasampler
# or
cd ComfyUI/custom_nodes && git clone https://github.com/Nessuno76/ComfyUI-AlchemicaSampler

ComfyUI 0.36.0 or newer - Krea 2 support and CLIP.generate (used by the prompt-writing nodes) both come from there. No pip dependencies; pyproject.toml lists an empty dependency list. Restart the ComfyUI server after installing, not just the browser.

Models are not included and not redistributed. Krea 2 is under the Krea 2 Community License - free commercially below $1M trailing revenue, with acceptable-use and content-filtering obligations attached. Qwen3-VL-4B is Apache-2.0; check the Qwen-Image VAE's own model page.

Where people get burned

An empty dropdown. The enum is generated from the folders ComfyUI knows about. Nothing in the unet_name list means the file isn't in models/diffusion_models (or you've got a nested folder - models/diffusion_models/krea2/... won't show flat).

Loading the encoder yourself and hitting the conditioning-size error. Use this node, or if you insist on core nodes, make sure your CLIPLoader says krea2. There is no other correct type.

Expecting the pack to install the weights. It doesn't. It also won't download anything for you, unlike some other loader packs.

Wondering why your LoRA does nothing to the prompt. The clip output here is the raw encoder - LoRA application happens in the LoRA Stack node, which has its own optional clip input and its own strength_clip. That's deliberate: the prompt-writing nodes want the plain encoder.