Nodes/Diztraido Nodes/Load Flux.2 Models
ComfyUI Node

Load Flux.2 Models

The Flux.2 stack in one node — and yes, Flux.2 is worth the VRAM, if you have it

By jadervasque·Created 2 months ago·Updated 2 months ago· 0
Load Flux.2 Models
    • model
    • clip
    • vae
    unet_name
    weight_dtype
    clip_name
    type
    vae_name

    Flux.2 changed the loading story compared to Flux.1, and mostly not in your favor. The 32B transformer plus the Mistral-3 24B vision-language encoder that replaced Flux.1's dual text encoders means the "stack" is heavier, and the model is only practical on serious hardware - 18–24GB VRAM even quantized for the dev tier, with Klein 4B/9B (the size-distilled variants that landed in January 2026) fitting in roughly 13GB. Load Flux.2 Models from the Diztraido pack is the "make this not painful" node: it wraps ComfyUI's UNETLoader, CLIPLoader, and VAELoader into a single node so you configure one panel and get model/clip/vae out.

    Why this node instead of the native three? Same reason as the Flux.1 loader: it collapses three loaders into one and, more usefully, sets the CLIP loader's type default to flux2 - the exact value you'd otherwise have to know to pick. Flux.2 uses a single text encoder (the VLM), so this loader is actually simpler than the Flux.1 one: one clip_name, not two.

    How it works

    Same pattern as the pack's other loaders. It asks ComfyUI for the required inputs of UNETLoader, CLIPLoader, and VAELoader, merges them into one panel, defaults the CLIP type to flux2, and at execution instantiates the three native loaders and returns their outputs. Nothing magical, nothing risky - it's ComfyUI's own loading code wearing one widget panel.

    The inputs and outputs that matter

    • unet_name - the Flux.2 diffusion model from models/diffusion_models/. For dev-tier work people overwhelmingly run fp8; Klein has fp8 and NVFP4 quantizations that cut VRAM by another ~55%.
    • weight_dtype - default, fp8_e4m3fn, fp8_e4m3fn_fast, fp8_e5m2. Pick fp8_e4m3fn unless you have serious headroom.
    • clip_name - the single text encoder (the Mistral-3-based VLM), from models/text_encoders/.
    • type - the CLIPLoader type list (28 entries covering sd3, flux, wan, ltxv, hidream, and friends). Defaults to flux2; leave it.
    • vae_name - pixel_space.

    Outputs: model (MODEL), clip (CLIP), vae (VAE).

    How to install

    Ships in Diztraido Nodes. ComfyUI Manager → search "Diztraido Nodes" → install diztraido-nodes → restart. Or:

    cd ComfyUI/custom_nodes
    git clone https://github.com/jadervasque/ComyUI-Diztraido.git
    

    Restart and it's under the Diztraido flux category. No Python dependencies beyond ComfyUI itself, and the node downloads no models - you bring the Flux.2 files.

    Common issues

    The node isn't the hard part; the model is. An empty dropdown means the file isn't in the expected folder - diffusion_models/ for the UNet, text_encoders/ for the clip. And size a reality: if your card is under ~13GB, Klein is the entry point, not dev. One more thing to know about Flux.2 before you commit disk: the dev tier is non-commercial under BFL's license, Klein 9B is non-commercial too, and only Klein 4B is Apache 2.0 - worth checking which you're legally allowed to ship before you build a product on it. None of that is this node's business, but it's the question people actually hit next.

    CategoryDiztraido/flux

    Inputs (5)

    NameTypeDefaultDescription
    unet_nameCOMBO0 options:
    weight_dtypeCOMBO4 options: default, fp8_e4m3fn, fp8_e4m3fn_fast, fp8_e5m2
    clip_nameCOMBO0 options:
    typeCOMBO28 options: stable_diffusion, stable_cascade, sd3, stable_audio, mochi, ltxv, +22
    vae_nameCOMBO1 options: pixel_space

    Outputs (3)

    NameTypeDescription
    modelMODEL
    clipCLIP
    vaeVAE