Krea2 SVDQuant Quantize

The short version

The pack ships pre-quantized checkpoints, but if you want to quantize your own Krea 2 - a finetune, a merge, a BF16 file you didn't want to download someone else's build of - this is the node that does it from inside ComfyUI. It runs the exact same function as the quantize_krea2.py CLI, so the graph and the terminal can't drift apart. Drag workflows/krea2_quantize.json into the canvas, point source_model at your BF16 checkpoint, press Queue.

The output is a quantized DiT in models/diffusion_models/ (about 8 GB), loadable with the Krea2 SVDQuant W4A4 Loader if you pick svdq, or the stock UNETLoader for w4a4 / int8 / fp8.

The inputs that matter

Mostly you set four things and leave the rest:

• source_model - a BF16 Krea 2 checkpoint (~24 GB). An already-quantized file can't be re-quantized, with one exception: FP8 is unpacked back to BF16 first and works fine.
• format - svdq (default): 4-bit weights and activations plus a low-rank bf16 correction branch. w4a4: same without the branch - smaller and ~9% faster per step. int8: the most faithful, still ~2x FP8 on Ampere. fp8: storage only, no speed gain on non-FP8 hardware.
• rank - the size of the low-rank branch (svdq only). The tooltip's honest framing: with no LoRA, ranks 64/128/256 measure the same, so 64 is enough. With a LoRA loaded, 256 wins clearly. 64 is the sensible default.
• refine_iters - 0 is a single-shot SVD split (~54s); 100 refines the branch against the quantization error and early-stops (~5.7 min). Keep it on if rank > 16 - refinement is what makes rank behave. Without it, raising rank just costs file size.

Worth knowing before you queue: this node blocks the queue for the whole run, unloads whatever model is resident to take the GPU, and writes ~8 GB - refusing to overwrite an existing file unless you tick overwrite. That's not a bug, it's the author being careful with your last run.

The optional act_stats input is the cheapest quality on offer: point it at an activation-statistics file from the Capture nodes and the branch gets fitted against measured per-channel activation energy instead of assuming it's uniform. It's free at inference and the best-measured gain in the whole repo (LPIPS 0.3378 → 0.2825). The seed input (default 0) makes builds reproducible on the same GPU; -1 restores the old unseeded behavior. Same seed on a different GPU still differs slightly - CPU and CUDA don't draw the same numbers.

How to install

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

No extra Python dependencies. If you'd rather run the same thing from a terminal:

cd ComfyUI/custom_nodes/krea-2-svdquant
python quantize_krea2.py /path/to/krea2_bf16.safetensors --format svdq --rank 256 --act-stats krea2_act_stats_turbo.safetensors

Gotchas

• Only the 224 transformer-block linears are quantized; norms, modulation, text-fusion and the final layer stay at full precision. Expect a log line like quantized 224 layers - a run reporting 0 fails loudly, so you'll know.
• If it's slower than FP8 afterwards, you're on a pre-cu130 torch build. The fast kernels only exist on CUDA-13+ PyTorch; below that everything drops to a pure-Python fallback that's slower than bf16. Check python -c "import torch; print(torch.version.cuda)" first.
• The download-table calibration files aren't for your build. krea2_act_stats_turbo.safetensors describes the stock Turbo weights. Activation statistics describe the weights they came from - a finetune or merge needs its own capture pass (that's what the Capture nodes are for).