Krea2 SVDQuant Quantize All-in-One

Bakes DiT, 4-bit quantized Qwen3-VL 4B text encoder, and VAE into a single combined checkpoint in ComfyUI/models/checkpoints/ (~12 GB instead of 15.5 GB across three files).

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

summary

◄source_dit▾►

◄text_encoder▾►

◄vae▾►

◄formatsvdq►

◄te_formatw4a4►

◄rank64►

◄refine_iters100►

◄groupsize256►

◄variantturbo►

◄output_name►

◄overwritefalse►

◄act_stats►

◄seed0►

CategoryKrea2/SVDQuant

Inputs (13)

Name	Type	Default	Description
source_dit	COMBO		The Krea 2 DiT model: either a BF16 source (~24 GB, will be quantized first) or an already-quantized SVDQuant/W4A4/INT8 file.
text_encoder	COMBO		The Qwen3-VL 4B BF16 text encoder safetensors (~8.88 GB). Its 252 linear projections will be quantized to te_format, reducing it to ~3.2 GB without losing the vision tower.
vae	COMBO		The Krea 2 VAE (~0.51 GB, kept unquantized).
format	COMBO	svdq	Diffusion model format. svdq: 4-bit weights and activations plus a low-rank bf16 correction branch.
te_format	COMBO	w4a4	Text encoder format: w4a4 (convrot_w4a4, ~3.2 GB) or int8 (int8_tensorwise, ~5.2 GB).
rank	INT	648–1024	svdq only: size of the low-rank branch on the diffusion model.
refine_iters	INT	1000–200	svdq only (if source_dit is BF16). Refines the branch against quantization error. Ignored if source_dit is already quantized.
groupsize	INT	25632–1024	convrot rotation group size.
variant	COMBO	turbo	Which Krea 2 release this is.
output_name	STRING		Filename in models/checkpoints/. Leave empty to derive automatically.
overwrite	BOOLEAN	false	Allow overwriting existing output checkpoint.
act_statsopt	STRING		svdq only (if source_dit is BF16): activation statistics file.
seedopt	INT	0-1–4294967295	Random seed for reproducible SVD.

Outputs (1)

Name	Type	Description
summary	STRING	Size, tensor counts and loader instructions for the baked checkpoint.