RUM FLUX.2 Load Native Model

If you've ever clicked a RUM workflow and it immediately errored, this is the node that started it. RUMFlux2LoadNativeModel is the entry point to ComfyUI-RUM, the ComfyUI adapter for RimoChan's RUM project - a cross-architecture distillation that squeezes SDXL-class anime knowledge (the Illustrious-family waiNSFWIllustrious teacher) into the much newer, much cheaper FLUX.2-Klein 4B architecture. You load a RUM checkpoint here, and everything downstream hangs off that one decision.

It's a diffusion_models loader, not a regular checkpoint loader. Point it at the RUM weights in ComfyUI/models/diffusion_models/ - model-checkpoint-1158000.safetensors for the text-to-image checkpoint, model-checkpoint-1202000.safetensors for the edit one - and you get back a ready-to-sample native BF16 MODEL. The repo ships T2I and edit workflows in examples/, and both start here.

What the inputs actually mean

Most of this node is one field you must not get wrong:

• base_text_tokens - how many of the Qwen text encoder's tokens the model treats as its "base" conditioning. The tooltip is blunt: it must match the text-encode and match-patch nodes downstream. The diffusers-match workflows use 200; the plain native 512-token path uses 512. Mismatch here means corrupted images, not an error - the pack added consistency checks in later versions, but you should still treat this as a commitment.
• rum_checkpoint_name - dropdown of everything in models/diffusion_models. Dropdown names vary between ComfyUI builds; on Aki you might see Unknown\no tags\rum-flux2-klein-4b-preview.safetensors.
• rum_checkpoint_path (optional) - an absolute path to a .safetensors file that overrides the dropdown. Handy for testing, fiddly for sharing.

Outputs are model (into the sampler chain) and a status string that tells you how many weights loaded and the resolved checkpoint path. Read it once when something looks off - it's the fastest way to confirm you loaded the right file.

Install and what it needs

Install the pack, then fetch the checkpoints:

cd ComfyUI/custom_nodes
git clone https://github.com/Rinne414/ComfyUI-RUM
cd ComfyUI-RUM
pip install -r requirements.txt
# then download the model weights
python scripts/download_models.py --comfy-root /path/to/ComfyUI --include-teacher-clip

Restart ComfyUI after installing. Note the requirements.txt is tiny (safetensors, huggingface_hub, numpy, pillow) - there's no diffusers dependency, that's the whole point of the native adapter. But you do need a recent ComfyUI: the nodes rely on newer internals like CLIP.load_model, SamplerCustomAdvanced, and nested latents, and the README pins verification at v0.26.2. An old ComfyUI will fail at load or sampling time.

Where people get burned

• The checkpoint goes in diffusion_models, not checkpoints. The dropdown literally reads from models/diffusion_models/; drop the file in the wrong folder and the node has nothing to offer.
• base_text_tokens mismatch - 200 vs 512 is the classic. If your image comes out structurally wrecked rather than subtly wrong, check this before anything else.
• The model dropdown shows a placeholder named put_rum_checkpoint_in_models_diffusion_models.safetensors when the folder is empty. That's not a bug, that's the author yelling at you.

The honest take: if you only ever run the shipped workflows, you'll set the dropdown once and touch base_text_tokens never again. The node is the boring, load-bearing front door of the pack - and when the door is broken, nothing else in RUM works.