DP Load UNET With Info

Models like Flux and SD3 usually ship as a standalone diffusion-model file rather than a full checkpoint bundle, so they need their own loader - this is that loader, with a weight-dtype picker and an info string tacked on. If you've been reaching for ComfyUI's core UNETLoader, this is a drop-in replacement that also tells you what got loaded.

How it works

unet_name picks the file from your models folder. weight_dtype is the setting that actually matters: default loads the model at its native precision, and the three fp8 options (fp8_e4m3fn, fp8_e4m3fn_fast, fp8_e5m2) cast it down to 8-bit floating point on load. This is the single most useful knob on the node, and it's worth understanding what it's actually trading off.

fp8 stores each weight in 8 bits - half the size of fp16, a quarter of fp32 - and for diffusion inference specifically, the quality cost is close to invisible. It's the default way large models like Flux run on consumer hardware: a 12B-parameter model that doesn't comfortably fit in VRAM at fp16 usually does at fp8, with a quality difference most people can't spot. fp8_e4m3fn is the standard pick - e4m3 trades numeric range for precision, which suits diffusion weights well. fp8_e4m3fn_fast is the same encoding with native fast compute enabled, which matters if you're on an RTX 40-series or newer GPU (Ada/Hopper/Blackwell architectures accelerate fp8 math natively) - on older cards, fp8 still saves the VRAM but doesn't get the speed boost, since the math falls back to higher precision anyway. fp8_e5m2 is the other standard 8-bit encoding, more range and less precision than e4m3 - it's the less common pick for diffusion models, where e4m3's precision trade generally serves image quality better.

The practical upshot: if a model is fitting fine at default, there's no strong reason to change it. If you're running out of VRAM, fp8 is the first thing to try before reaching for a heavier compression scheme like GGUF, and one you can drop in with basically no visible quality cost.

Inputs and outputs

Required:

• unet_name - dropdown from your diffusion-model files.
• weight_dtype - default, fp8_e4m3fn, fp8_e4m3fn_fast, or fp8_e5m2.

Outputs:

• model - the loaded MODEL, wire it into your sampler chain the normal way.
• model_info - a string with details about what was loaded, handy for a text-preview node or for logging which weight precision a given run actually used.

How to install it

Search ComfyUI-Desert-Pixel-Nodes in ComfyUI Manager, or:

cd ComfyUI/custom_nodes
git clone https://github.com/DesertPixelAi/ComfyUI-Desert-Pixel-Nodes

Restart, find it under Desert Pixel/loaders. No dependencies beyond a standard ComfyUI/PyTorch install - you'll need the actual model file in your models/unet (or diffusion_models, depending on your ComfyUI version) folder separately; this node loads what's already there.

Common issues & troubleshooting

Dropdown is empty. No files found in the folder ComfyUI scans for standalone diffusion models - check the file landed in the right subfolder for your ComfyUI version, then restart, since custom-node dropdowns populate at startup.

Still running out of VRAM even at fp8. fp8 roughly halves memory versus fp16, but a large enough model can still not fit - below roughly 10-12GB VRAM, fp8 often isn't enough on its own, and a more aggressive GGUF quantization (Q4-Q5) may be the better route even though it trades away some of fp8's speed advantage.

fp8_e4m3fn_fast isn't actually faster. The speed benefit depends on your GPU having native fp8 compute - RTX 40-series and newer. On older cards, the weights still load in fp8 (so you keep the memory saving) but the math runs at higher precision under the hood, so you won't see the speed bump the "fast" variant is named for.

Generation quality looks noticeably worse after switching to fp8. For most diffusion models this shouldn't be visible - if it is, double-check you didn't also pick a lower-precision text encoder or a heavier LoRA stack at the same time, since fp8's "near-zero quality loss" reputation is specifically about the diffusion weights themselves, not a guarantee against every other setting you might have changed alongside it.