Krea2 SVDQuant Quantize

Builds a quantized Krea 2 checkpoint from a BF16 one, without leaving ComfyUI. BLOCKS THE QUEUE while it runs (54s to ~6min), unloads any loaded model to free the GPU, and writes ~8 GB. Load the result with the Krea2 SVDQuant W4A4 Loader (svdq) or the stock UNETLoader (w4a4/int8/fp8).

By alperktt·Created about a month ago·Updated 5 days ago· 11

summary

◄source_model▾►

◄formatsvdq►

◄rank64►

◄rank_allocuniform►

◄refine_iters100►

◄groupsize256►

◄variantunknown►

◄output_name►

◄overwritefalse►

◄act_stats►

◄seed0►

CategoryKrea2/SVDQuant

Inputs (11)

Name	Type	Default	Description
source_model	COMBO		The BF16 Krea 2 checkpoint to quantize (~24 GB). An already-quantized file cannot be used as a source, except FP8, which is unpacked back to BF16 first.
format	COMBO	svdq	svdq: 4-bit weights and activations plus a low-rank bf16 correction branch. w4a4: the same without the branch - smaller and ~9%% faster per step. int8: the most faithful option and still ~2x fp8 on Ampere. fp8: storage only.
rank	INT	648–1024	svdq only: size of the low-rank branch. Only pays off with refine_iters > 0. Without a LoRA, 64 / 128 / 256 measure the same, so 64 is enough. With a LoRA loaded, 256 wins clearly and 64 loses most of its advantage - so 256 if you use LoRAs.
rank_alloc	COMBO	uniform	svdq only: how the rank budget is spread across the eight projection types. Same file size either way. uniform gives every layer the same rank. gqa moves the budget to attn.wk / attn.wv, which absorb ~2x the quantization error at a third of the branch cost because Krea 2 has only 12 kv heads. Measured and it does not pay: LPIPS 0.3523 vs 0.3403 for uniform, 5 of 10 prompts better, no mean effect. It does halve the spread across prompts and improve the worst one. Leave on uniform unless you are re-testing that.
refine_iters	INT	1000–200	svdq only. 0 is a single-shot SVD split (~54s); 100 refines the branch against the quantization error and early-stops (~5.7min). Keep this on if rank > 16: refinement is what makes rank behave. Without it, raising rank costs file size and buys nothing measurable.
groupsize	INT	25632–1024	convrot rotation group size. Unused for fp8.
variant	COMBO	unknown	Which Krea 2 release this is. Affects only the output filename and the recorded metadata - quantization is identical; what differs is the sampler settings afterwards.
output_name	STRING		Filename inside models/diffusion_models/. Leave empty to derive it from the variant and format.
overwrite	BOOLEAN	false	Off means an existing file of the same name is an error rather than 8 GB written over your last run.
act_statsopt	STRING		svdq only: an activation-statistics file from the Capture nodes (a bare filename is looked up in ComfyUI/output/). Fits the low-rank branch against measured per-channel activation energy instead of assuming it is uniform. Free at inference and the best-measured setting here - LPIPS to BF16 0.3378 to 0.2825 with no LoRA. Empty means the plain objective.
seedopt	INT	0-1–4294967295	svdq only: seed for the randomized low-rank SVD. Quantizing twice with the same seed on the same GPU now gives identical files; -1 restores the old unseeded behaviour, where it did not. The same seed on a different device still differs (~1e-4 per weight) - CPU and CUDA do not draw the same numbers.

Outputs (1)

Name	Type	Description
summary	STRING	Where the checkpoint was written, and what went into it.