ComfyUI Node
LC Model Quantize 🗜️
Saves a smaller copy of a diffusion model: fp8, mxfp8, nvfp4, int8 or int4 (native ComfyUI files) or GGUF Q8_0 to Q2_K. From the wired MODEL (LoRAs and merges baked in) or straight from a file, one layer at a time. Embedders, norms and the final layer stay full precision. The wired model is not changed.
LC Model Quantize 🗜️
- model
- report
◄source▾►
◄formatfp8 (scaled)►
◄protect_edgesfalse►
◄filename►
CategoryLC Modelbuilder
Inputs (5)
| Name | Type | Default | Description |
|---|---|---|---|
| source | COMBO | Where the weights come from. '(use the model input)' takes the wired MODEL, LoRAs and merges included (fine up to about 12-14B on 64 GB RAM). A file is read one layer at a time, so any size works. Pick the bf16 / fp16 file, not an fp8 one. | |
| format | COMBO | fp8 (scaled) | fp8: half size, near lossless, fast on 40 / 50 series. mxfp8, nvfp4: 50 series only. int8 (convrot): half size, runs on 30 series too. int4: a quarter. GGUF needs ComfyUI-GGUF to load: Q8_0 near lossless, Q6_K / Q5_K very close, Q4_K_M the usual pick for small cards, Q3 / Q2 for when nothing else fits. _M keeps the first and last two blocks at a higher type. |
| protect_edges | BOOLEAN | false | Keep the first and last transformer block at full precision too (a little bigger, a little closer to the original). |
| filename | STRING | Saved in models/diffusion_models. Empty = the source name plus the format. A subfolder works: 'Krea2/Krealism'. An existing file is never overwritten. | |
| modelopt | MODEL | Used when source is '(use the model input)'. |
Outputs (1)
| Name | Type | Description |
|---|---|---|
| report | STRING | — |