Doom Super Loader

Modern ComfyUI workflows have a loader problem: one model isn't one node anymore. Flux wants a diffusion model, a T5, a CLIP-L, and a VAE. Flux 2 and Krea 2 want even more encoders bolted on. The stock answer is a growing pile of Load Diffusion Model, CLIPLoader, VAELoader, and CLIP Vision Loader boxes that take up half your canvas before you've done anything. Doom Super Loader is the pack's attempt to fold that pile into one node: checkpoint or diffusion model, VAE, up to three CLIPs, CLIP Vision, and a Style Model, all in a single blue box.

It's the node you reach for when you're setting up a fresh Flux / Flux 2 / Krea 2 graph and you want the loading section to take one row of canvas instead of three. It also has GGUF support for quantized diffusion models, which matters if your VRAM is the thing deciding whether a 12B Krea 2 or a 56B Flux 2 is even runnable today.

How it works

The trick is that every component overrides independently. Set a checkpoint_name and it wins for the baked VAE/CLIP; set a diffusion model_name instead and it's loaded directly. Mix and match: a checkpoint for the base, then point clip_name1 and clip_name2 at a better encoder pair and vae_name at a preferred VAE. The outputs are the standard MODEL, CLIP, VAE, plus CLIP_VISION and STYLE_MODEL - so it slots into existing graphs without you re-wiring anything.

The inputs that matter

• checkpoint_name / model_name - one of these is your actual model. Both default to "None"; if both are "None", you get nothing. If both are set, the checkpoint overrides.
• vae_name - including the special "Baked VAE" option to use whatever's inside the checkpoint.
• clip_name1 / clip_name2 / clip_name3 - up to three text encoders. Modern DiT models often need two; a few need three.
• clip_type - the one people forget. This tells the node what family the CLIP belongs to: flux, flux2, ltxv, wan, minimax, krea2, and so on. Set it wrong and your "aha" prompt comes out as garbage because the encoder is being interpreted as the wrong architecture.
• clip_device - set cpu to park the CLIP in system RAM. Slower encoding, but it can be the difference between running and OOM on an 8 GB card.

Install is the pack install:

cd ComfyUI/custom_nodes
git clone https://github.com/PeterMikhai/Doom_Flux_NodePack

Restart ComfyUI, or find it in Manager by searching "DoomAI Nodes". No pip dependencies beyond what ComfyUI already ships (torch, PIL). Everything it loads comes from the standard diffusion_models, text_encoders, vae, clip_vision, and style_models folders.

Where people get burned

• The clip_type trap. This is the number one source of "my Flux prompt renders like a dream" problems. The node will happily load a wrong-typed encoder and produce text that's technically encoded but semantically scrambled. Check the dropdown against your model family.
• Both loaders set. Checkpoint silently wins, which is confusing if you thought you were loading a diffusion model.
• "None" everywhere. The node doesn't error loudly; it just gives you an empty MODEL. If nothing comes out, check that you actually picked a model.
• Unfamiliar CLIP types. This pack tracks the newest encoders (Minimax, Krea 2) faster than most stock nodes do - which is the upside, but it means the dropdown assumes you know what a given model family needs.

One honest caveat: it's a single-maintainer pack, Russian-authored, with basically no community footprint - you're adopting it early. But for the job of "get three encoders into one box", it does exactly what it says.