Diffusion Loader (Mango)

Modern models don't ship as a single checkpoint file. Flux, SD3, and Pony-era SDXL workflows split the "model" into a diffusion UNET plus separate text encoders (often two of them) plus a VAE - which is why the stock loader alone can't handle them. Diffusion Loader (Mango) is the pack's answer: one node that loads a diffusion model, a dual CLIP, a VAE, and up to five LoRAs, and hands you the same metadata-friendly unet_name / unet_hash outputs the rest of the pack expects.

If you've built a Flux or SDXL graph with the native "UNET Loader + DualCLIPLoader + VAELoader" trio, this node collapses those three into one. It's the modern-architecture sibling of Loader (Mango), and it's the loader you'd reach for on any split-file setup.

How it works

The node reads your models/diffusion_models folder for the UNET, models/text_encoders for the two CLIPs, and models/vae for the VAE (it also supports the tiny TAESD/TAEF1 preview VAEs if you have those installed). It loads the diffusion model with comfy.sd.load_diffusion_model, builds a dual CLIP via comfy.sd.load_clip with the CLIP type derived from your type selector, and loads the VAE directly from its state dict. LoRAs apply on top, same load_lora_for_models machinery as the rest of the pack. It computes a 10-character SHA-256 prefix of the UNET file, so unet_hash is honest metadata material.

Inputs and outputs that matter

• unet_name - dropdown of your diffusion_models. The core choice.
• type - sdxl, sd3, flux, or hunyuan_video. This sets the CLIP type, so get it right or your conditioning will be wrong in confusing ways.
• clip_name1 / clip_name2 - the two text encoders from models/text_encoders.
• vae_name - from models/vae, plus the taesd/taef1 options if present.
• weight_dtype - default, or fp8_e4m3fn, fp8_e4m3fn_fast, fp8_e5m2. The fp8 options quantize the diffusion model on load, which is how you fit big Flux files into tight VRAM. If you're on a 24 GB card you probably don't need them; if you're on 8 GB, they might be the difference between running and not.
• LoraName1…5 / LoraWeight1…5 - the usual five LoRA slots.
• device (optional) - default or cpu. "CPU" here forces the CLIP onto CPU, useful for giant text encoders when VRAM is short, at the cost of slower prompt encoding.

Outputs: model / clip / vae go to your sampler and decoder; lora_stack feeds the sampler for metadata; unet_name / unet_hash (STRINGS) feed the text fields on FluxSampler (Mango) so your saved images record which UNET you used and its hash.

Installing it

It's part of the Mango Node Pack:

cd ComfyUI/custom_nodes
git clone https://github.com/mang01010/MangoNodePack

Restart ComfyUI, or install "Mango Node Pack" via ComfyUI Manager. No extra model downloads beyond what you already have in models/diffusion_models, models/text_encoders, and models/vae - and note those folders may be empty in a fresh install, so point them at whatever model files you're already using.

Gotchas

The type selector is the trap. Pick flux when you're actually running a Pony SDXL model and your generations will be subtly broken in ways that are hard to debug. Also, fp8 weight dtypes trade a little quality for VRAM headroom - Q8-style quantization is "basically fp16 at half the size" territory, but the e5m2 variant is visibly rougher, so don't default to it unless you must. One more: because it loads the UNET from diffusion_models, your split-model files need to be in that folder - a checkpoint-style file in models/checkpoints won't show up in the dropdown. It's a small node that assumes you've already organized your models the modern way, and it's the right tool when you have.