Krea2 SVDQuant Checkpoint Loader

The short version

Normally, running Krea 2 in ComfyUI means juggling three separate files: the DiT, the Qwen3-VL text encoder (~8 GB), and the Qwen-Image VAE. People miss the second two all the time and then report the model as broken. This node exists to end that: it loads an all-in-one checkpoint that bundles the quantized diffusion model, a 4-bit text encoder, and the VAE into a single 12.77 GB file - and hands you model, clip, and vae from one loader.

It's the recommended starting point in the pack's README for exactly that reason: download one file, drag in one workflow, press Queue. If you'd rather keep the three-file modular setup, use Krea2 SVDQuant W4A4 Loader plus stock CLIPLoader / VAELoader instead - this node is the convenience option, and it's a genuinely good one.

How it works

Under the hood it's load_state_dict_guess_config on the combined file. If the checkpoint carries SVDQuant low-rank branches (the *.svdq_l1 / *.svdq_l2 tensors), they're pulled out, the model is built, and the 224 low-rank branches get attached on top of ComfyUI's native convrot_w4a4 kernel - same mechanism as the standalone W4A4 loader, just with the text encoder and VAE in the same container. If you point it at a branchless all-in-one file instead, it loads cleanly without branches.

The file to get is Krea2-Turbo-AllInOne-SVDQuant-W4A4-rank256-actaware-TEW4A4.safetensors from AlperKTS/Krea-2-SVDQuant-ComfyUI on Hugging Face, dropped into models/checkpoints/. The author picked that exact build (rank 256, activation-aware) because it's the one that holds up when you stack a LoRA on top.

The inputs that matter

• ckpt_name - the all-in-one checkpoint from your models/checkpoints/ folder. That's it for required inputs.
• vram_management - auto / classic, same as the W4A4 loader. Leave it on auto. classic pins the older patcher, and past the point where the model stops fitting VRAM that means per-module weight streaming at 30–100s an iteration. If you ever see that, it's this setting (or a KREA2_DISABLE_DYNAMIC=1 env var someone left in your launch script).

Outputs: model → KSampler, clip → CLIPTextEncode, vae → VAEDecode, plus a status string reporting rank, variant, size, and which kernel the quantized layers dispatch to. Read that status line once on your first run - it tells you whether the fast int4 kernel is actually in play, which is the difference between 2.4x speedup and a silent pure-Python fallback.

Install and gotchas

Same pack, same install:

cd ComfyUI/custom_nodes && git clone https://github.com/alperktt/Krea-2-SVDQuant-ComfyUI krea-2-svdquant

No extra Python dependencies - it uses ComfyUI's own comfy_kitchen backend. The one real gotcha is environmental and it's the pack's single most common support issue: quantized checkpoints need a cu130+ (CUDA 13) PyTorch build. Below that, ComfyUI disables the CUDA backend for these kernels and the whole thing runs slower than bf16. Verify with python -c "import torch; print(torch.version.cuda)" before you blame the file. Krea 2's weights also carry a trained-in safety filter (the open release went through an alignment pass the hosted model never got), which is a property of the model, not this loader - a LoRA-based uncensor is the community's usual answer if that's a problem for you.