VELVET VICE KREA — Select Krea 2 Model

"VELVET VICE KREA - Select Krea 2 Model" is the node that answers the first question everyone asks when they load the Velvet Vice Krea 2 workflow: where's the model? The answer is intentionally boring. This pack doesn't ship a creator-specific checkpoint - a deliberate choice the README calls out. The loader just lists every Krea 2-compatible diffusion model sitting in your diffusion_models folder and lets you pick one. Your Krea 2, your loader, no opinion about whose fine-tune is best.

Why "model-neutral" is the right call

The Krea 2 model family released in mid-2026 as a 12B base under a revenue-capped community license, and it's a workflow-driven ecosystem - people run the official release, community FP8 conversions, and various merges. A workflow that hard-codes one creator's filename would break for everyone else. So this loader lists the actual contents of your ComfyUI/models/diffusion_models folder at runtime instead. When you open the dropdown you'll see your real models; the placeholder "SELECT KREA 2 MODEL" just marks "you haven't chosen yet."

The README is explicit about the two important behaviors: the saved value stays the sentinel string (not a creator-specific filename), so the workflow survives being shared, and the node delegates the actual loading to ComfyUI's core UNET loader - meaning weight-dtype behavior is exactly what stock ComfyUI gives you.

Inputs and outputs

• model_name - enum, populated from your diffusion_models folder. Defaults to the "SELECT KREA 2 MODEL" placeholder.
• weight_dtype - default, fp8_e4m3fn, fp8_e4m3fn_fast, or fp8_e5m2. Marked advanced; you'll rarely touch it unless your checkpoint ships as a specific dtype.

Output is a single model (MODEL) that feeds your KSampler, plus the branch's CLIP and VAE come from the workflow's other loaders.

On weight_dtype: the community FP8 conversion of Krea 2 (roughly 12 GB versus ~24 GB BF16) is how most people run it on 16–24 GB cards, and you generally load it as-is with default. Only reach for an explicit dtype if your file genuinely needs it - you can't conjure precision that isn't in the weights.

The two errors you'll actually hit

Because the loader validates against your local folder, the failure modes are unusually helpful:

• You queue without picking a model → "Select your installed Krea 2 diffusion model in the KREA 2 model loader before starting this branch."
• You pick a model that isn't actually in diffusion_models anymore → "Selected diffusion model is not installed."

The fix for both is the same: put your Krea 2 checkpoint in ComfyUI/models/diffusion_models/ (or models/unet/, which ComfyUI's core loader treats the same way), refresh the node, and pick it. Also note the workflow runs three independent loaders - one per branch - and only the branch you're executing needs a valid selection. Bypassed branches can happily sit on the placeholder.

Installing

cd ComfyUI/custom_nodes
git clone https://github.com/Velvet-Vice/velvet-vice-krea

Restart ComfyUI, or use ComfyUI Manager and search "VELVET VICE - KREA". No pip dependencies. The loader itself downloads nothing - the Krea 2 model, its Qwen3-VL text encoder, and the VAE are separate downloads you bring to the workflow, and the encoder/VAE loaders elsewhere in the graph handle those.