Diffusion Model Loader KJ

This is ComfyUI's stock "Load Diffusion Model" node with the performance switches you'd normally set via command-line flags or extra patch nodes moved onto the loader itself. Instead of loading your model and then bolting on a SageAttention node, an fp16-accumulation node, and a dtype override, you pick it all here in one place. For anyone who runs Flux, Wan, or the newer DiT models and cares about squeezing speed and VRAM, that's a genuine convenience.

What it actually does

Under the hood it's a UNET/diffusion-model loader - you point it at a weights file and it hands you a MODEL. What makes it the "KJ" version is the row of optimization options baked in:

• weight_dtype - how the weights get loaded: default, fp8_e4m3fn, fp8_e4m3fn_fast, fp8_e5m2, fp16, bf16, or fp32. This is the big one. fp8 halves VRAM versus fp16 at near-zero quality loss - it's been the default way to run Flux-class models since 2024, and if the fp8 file fits, you just use it. The _fast variant leans on native fp8 compute on RTX 40-series and newer cards for extra speed.
• sage_attention - turn on SageAttention right here, choosing the kernel (auto is the sane pick; the list runs through the int8/fp8 variants up to sageattn3). SageAttention is a memory-efficient attention implementation that gives a real speedup - the catch is it needs Triton, which is famously painful to install on Windows.
• enable_fp16_accumulation - the same thing ComfyUI's --fast flag does, as a per-model toggle. It's a meaningful speed boost on supported models, but some (Qwen, for instance) don't play nice with fp16 accumulation, so having it on the loader means you flip it per-model instead of restarting Comfy with a different flag every time you switch.

There's also compute_dtype (default/fp16/bf16/fp32) for the math precision, patch_cublaslinear (swaps in CublasLinear for the linear layers), and an optional extra_state_dict string. The model_name dropdown is your weights file - it reads from your diffusion_models folder, so if it's empty, your model isn't where ComfyUI expects it.

The single output is MODEL, which goes wherever a model goes - LoRAs, sampler, the works.

Why bother, versus the stock loader

Two reasons. First, tidiness: one node instead of a loader plus three patch nodes. Second, and more useful, is the per-model toggling. fp16 accumulation is the clearest example - it helps most models and hurts a few, and toggling it on the loader beats maintaining two ComfyUI launch shortcuts. That's exactly the pain this consolidates.

Installing it

Comes with kijai's KJNodes pack - kijai being the person who ships the fp8 conversions for basically every new model, so a loader that foregrounds fp8 and SageAttention is on-brand.

• ComfyUI Manager - search KJNodes for ComfyUI, install, restart.
• Manual - cd ComfyUI/custom_nodes && git clone https://github.com/kijai/ComfyUI-KJNodes, then pip install -r requirements.txt, and restart.

Common issues

SageAttention is the usual troublemaker. Selecting a sageattn option here does nothing if SageAttention (and Triton) aren't actually installed in your environment - the node exposes the switch, it doesn't install the dependency. On Windows that Triton install is the real work; budget for it. Leave sage_attention at disabled until you've confirmed the library imports.

Second, don't assume fp8 fits everything below 12GB. fp8 is a single point, not a ladder - under roughly 10–12GB the fp8 file stops fitting and you want GGUF's lower quant levels instead, which this loader doesn't cover. And enable_fp16_accumulation is a no-op (or worse) on models that don't support it, so if output degrades right after you flip it, that's your suspect - turn it back off for that model.