Blender Input Load Diffusion Model

Newer architectures like Flux and SD3 don't ship as all-in-one checkpoints - the diffusion model lives in its own file, separate from the text encoder and VAE. BlenderInputLoadDiffusionModel is the ComfyUI-Blender version of the UNETLoader that handles those, so the model dropdown lands on the add-on's Blender panel. If your workflow is Flux/SD3-era, this is the model loader you'll actually use; the checkpoint-style loader in this pack is for the older SD 1.5/SDXL world.

The implementation is a thin subclass of ComfyUI's built-in UNETLoader: it inherits the parent's inputs (including the weight-dtype choice), adds the pack's order, default, and optional group, then in execute normalizes the file path and calls the parent's load_unet. There's no separate loader logic to learn - it is the native loader, with a Blender-facing wrapper.

Inputs and outputs

• unet_name - the diffusion model file from your ComfyUI diffusion_models folder. In Blender this is a dropdown of everything the server sees there.
• weight_dtype - how the model is loaded: default, fp8_e4m3fn, fp8_e4m3fn_fast, or fp8_e5m2. This is the "how much VRAM do I burn" switch. If the full model barely fits next to Blender, fp8 is the classic move - the community shorthand is that fp8_e4m3fn is "basically fine and much lighter," and fp8_e4m3fn_fast trades a little quality for speed. Leave default if you have room.
• default - which model the dropdown starts on.
• order - panel position, lower first.

Output is a single MODEL, which feeds your sampler the same way a native UNETLoader would.

Installing it

Whole-pack routine. ComfyUI side, ComfyUI Manager → search "ComfyUI-Blender", or:

cd ComfyUI/custom_nodes
git clone https://github.com/alexisrolland/ComfyUI-Blender

Restart ComfyUI. No extra Python dependencies, but you need a current ComfyUI - schema v3 nodes, and old builds won't load them (the README says latest required). The Blender add-on is the other half, a ZIP from the latest release installed via Edit > Preferences > Add-ons > Install from Disk.

Where people get burned

The dropdown reflects the server's diffusion_models folder, not your local disk - a remote server means you pick from its files. And the fp8 weights question comes up constantly: if you have the VRAM, default is the no-regrets choice; if Blender is hogging the card, fp8 is what keeps the shot rendering. People also grab this node when their workflow is actually SDXL-on-a-checkpoint and then wonder why nothing loads - checkpoint models don't live in diffusion_models. Pick the loader to match your architecture. Panel only appears after an API-format export and import, and the node title is the label - "Diffusion Model" beats "unet1".