Nodes/Krea-2-SVDQuant-ComfyUI/Krea2 SVDQuant LoRA Loader
ComfyUI Node

Krea2 SVDQuant LoRA Loader

Applies a LoRA to a Krea2 W4A4 quantized model as a parallel low-rank branch, leaving the 4-bit weight untouched. The stock loader has to dequantize, add and requantize, which puts the LoRA delta through 4-bit quantization along with the weight. LoKr/LoHa/OFT cannot fold into the branch and run per-forward instead; the 'adapters' input trades that exactness back for the stock loader's speed.

By alperktt·Created about a month ago·Updated 5 days ago· 11
Krea2 SVDQuant LoRA Loader
  • model
  • model
lora_name
strength1.00
adaptersbypass (exact, slower)
CategoryKrea2/SVDQuant

Inputs (4)

NameTypeDefaultDescription
modelMODELOutput of the Krea2 SVDQuant W4A4 Loader, or any Krea2 checkpoint with convrot_w4a4 quantized blocks.
lora_nameCOMBOA Krea2 LoRA. Targets 'blocks.N.{attn,mlp}.*' under either prefix ('diffusion_model.' or 'transformer.') for the quantized blocks; anything else it carries (txtfusion etc.) goes through ComfyUI's normal path.
strengthFLOAT1.00-10–100 passes the model through untouched. Negative values invert the LoRA. Chain more of these nodes to stack LoRAs - the whole stack is rebuilt each time, so strengths stay exact.
adaptersoptCOMBObypass (exact, slower)Only affects LoRAs that cannot fold directly into the low-rank branch (LoKr, LoHa, OFT); a plain LoRA is free in all modes. 'bypass' computes the adapter every forward without touching the 4-bit weight. 'svd delta' decomposes the delta into low-rank factors (zero per-step overhead). 'bake' rewrites the weight once: fast, but quantizes the delta to 4 bits.

Outputs (1)

NameTypeDescription
modelMODELThe model with the LoRA attached as a parallel branch.