☁️BizyAir Nunchaku FLUX DiT Loader

What it is

Nunchaku is the community's name for SVDQuant - a 4-bit quantization scheme that shrinks Flux down to a fraction of its normal VRAM footprint while staying close to full-precision quality. It's genuinely popular: people run Chroma and Flux merges through it and report renders in the 10-second range on cards that would otherwise choke. This node is BizyAir's own cloud implementation of that idea, loading an SVDQuant-quantized Flux checkpoint on BizyAir's hardware rather than yours.

Worth knowing up front, because it trips people up: this is not the same thing as mit-han-lab/ComfyUI-nunchaku, the original local Nunchaku node pack. They share a name and a quantization format, but one runs on your GPU and one runs on BizyAir's. There's a real history of ComfyUI Manager's older missing-node auto-installer mixing the two up - users following a Nunchaku tutorial would get BizyAir silently installed instead when Manager guessed wrong about which pack a workflow needed. Comfy Org fixed that inference bug, but if you've ever wondered why BizyAir showed up in your custom_nodes folder uninvited, that's very likely why.

How it works

Point it at one of the pre-quantized SVDQuant Flux checkpoints and it loads that model on BizyAir's servers, ready to feed into the rest of a BizyAir-typed sampling chain. The quantization itself works by compressing weights to int4 while keeping a small set of outlier values at higher precision - SVDQuant's actual trick - which is why it holds up so much better than naive 4-bit rounding.

Inputs and outputs that matter

model_path picks the quantized checkpoint - the sample options include svdq-int4-flux.1-dev, svdq-int4-flux.1-fill-dev, and int4 builds of Kontext and Krea variants. cache_threshold is a speed/quality trade: it works like WaveSpeed's residual_diff_threshold, letting the model skip recomputing similar steps. The tooltip is direct about it - "a typical setting is 0.12," and 0 disables the effect entirely, which is the safe starting point if you want full quality first and speed second. attention defaults to nunchaku-fp16, offering roughly a 1.2× speedup over standard flash-attention2; the tooltip notes 20-series GPUs are locked into this option regardless. cpu_offload (auto by default) shifts the transformer to CPU when GPU memory is under 14GB - though since this loader runs on BizyAir's own hardware, that setting is really about BizyAir's infrastructure, not yours. device_id picks the GPU index, and data_type defaults to bfloat16 (20-series cards need float16 instead, per the tooltip). The optional i2f_mode only matters on 20-series GPUs and controls their GEMM implementation.

The single output is BIZYAIR_MODEL, which only connects to other BizyAir-typed nodes - it won't plug into a stock ComfyUI KSampler or CLIPTextEncode.

Installing it

Search "BizyAir" in ComfyUI Manager, or git clone https://github.com/siliconflow/BizyAir.git into custom_nodes and restart. You need a BizyAir API key configured before this or any other node in the pack will run - the in-app login prompt handles registration.

Common issues

If you were actually trying to install the local Nunchaku pack and got this instead, uninstall it via Manager's Custom Nodes Manager and go get mit-han-lab/ComfyUI-nunchaku directly - they solve the same VRAM problem in opposite ways (yours vs. BizyAir's GPU), and mixing them up wastes an install cycle. Beyond that, since BIZYAIR_MODEL doesn't interoperate with stock types, a "no matching node" or red-wire error usually just means you've got one BizyAir node feeding a non-BizyAir one somewhere upstream or downstream - trace the chain and make sure it's all-cloud or all-local, not a mix.