Doom Flux1 Loader (GGUF / WAN)

Doom Flux1 Loader (GGUF / WAN) is the pack's universal diffusion-model loader, and the name is doing a lot of work. It loads Flux.1 - and quite a bit beyond it - by auto-detecting whether you handed it a full-precision .safetensors or a quantized .gguf, loading up to three text encoders (each FP or GGUF), and optionally loading a second model for the MoE dual-expert architectures like Wan 2.2. If you're running the pack's Flux samplers, this is the loader you wire into them; if you're doing GGUF-quantized Flux on a small card, this is the node that makes it a two-dropdown affair instead of a scavenger hunt.

One important framing: because it ships with its own GGUF backend, it doesn't depend on the ComfyUI-GGUF pack. The vendor's own gguf folder is bundled inside this pack.

How it works

The mechanism is extension sniffing plus fallbacks:

• model_name - a dropdown of both diffusion_models and .gguf files. If the name ends in .gguf, it's loaded through the bundled GGUF path (GGUFModelPatcher + GGML operations); otherwise a normal diffusion-model load. This is your Flux.1 dev/schnell, or any DiT.
• model_name_2 - the second model for MoE pairs like Wan 2.2's two experts. Both get a ModelSamplingFlux-style patch so the Flux-family samplers can drive them. If you're not doing MoE, leave it on None; the model_2 output is None too.
• clip_name1/clip_name2/clip_name3 + clip_type - up to three text encoders, each FP or GGUF (GGUF clips come from the clip folder), typed by clip_type (flux, flux2, wan, ltxv, minimax, krea2…). Flux.1's T5 + CLIP-L is a two-encoder job; Flux.2's 24B Mistral encoder is the third kind of beast entirely.
• vae_name - with a "Baked VAE" option for checkpoints that carry one.
• weight_dtype - default / fp8_e4m3fn / fp8_e4m3fn_fast / fp8_e5m2. The fp8 ladder is the modern fit lever for 40-series cards and up.
• device - default or cpu; the cpu option offloads loading to RAM, which is the low-VRAM emergency hatch.

Outputs: model, model_2, vae, clip - exactly what the samplers want.

GGUF, briefly

The GGUF ladder is your fit-to-VRAM tool: Q8 is basically fp16 at half the size, Q4/Q5 is where 12GB cards live, Q3 and below are desperation. The tax is dequantization overhead - worst when LoRAs force repeated dequantize/requantize cycles. The KB's GGUF panel has the full story. Rule of thumb: if it fits at fp8, use fp8; reach for GGUF when it doesn't.

Installing it

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

Restart ComfyUI (or ComfyUI Manager). No extra pip dependencies - the GGUF backend is bundled. Note the README still shows the renamed DoomAI_nodes.git; the live repo is Doom_Flux_NodePack.

Common issues

• GGUF options missing from dropdowns. The bundled backend fails gracefully (_GGUF_OK=False) if it can't import - GGUF files then vanish from the lists and you get a log warning. Usually means a stale ComfyUI; update first.
• "Which VAE?" For Flux.1 use the standard ae.safetensors; choose "Baked VAE" only when loading a checkpoint that embeds one. Loading a baked VAE with a diffusion-model Flux setup is a classic source of weird colors.
• The text encoder is its own VRAM budget. On Flux.2-class encoders, quantize the encoder hard and keep precision in the diffusion model - the encoder is frequently what decides whether the whole thing fits. (See the KB's troubleshooting doc on 2026's "second VRAM budget".)
• Wan 2.2 workflows expect the second expert wired from model_2 - leave it on None and you'll get half a model and odd results.

It's the most "actual product" loader in the pack: FP or GGUF, one or two models, three encoders, all in one node, and it feeds every Doom sampler without adapters.