FunPack Diffusion Model Loader

ComfyUI's built-in UNETLoader gets you a model in and out, but that's about it. The FunPack Diffusion Model Loader is what happens when someone got tired of stacking a GGUF pack, an fp8 loader and an attention override into the same graph - it's a drop-in loader that puts weight dtype, compute dtype, an attention backend and even MiniMax H3's sparse-attention speedup into one node, and then prints a status string telling you exactly what it decided. It ships inside ComfyUI-FunPack, DigitalGarbage's hobbyist UI/utility pack, and for regular ComfyUI use it's the pack's most useful node that doesn't require the Cutting Room editor at all.

What it actually does

The required inputs read like a list of things you usually patch in with three separate packs:

• model_name - picked from ComfyUI/models/diffusion_models. The neat part: .gguf files are listed right alongside .safetensors ones, and the loader reads GGUF itself, so you don't need ComfyUI-GGUF installed for it to work.
• weight_dtype - how weights sit in VRAM: default, fp8_e4m3fn, fp8_e4m3fn_fast, fp8_e5m2, fp16, bf16, fp32. The _fast variant also runs fp8 matmuls, which is quicker on Ada and newer at a hair of precision cost; plain fp8_e4m3fn roughly halves memory with barely visible loss. This is the same fp8 story as everywhere else in ComfyUI - on a 12–16GB card it's usually the first knob you touch.
• compute_dtype - what the model computes in, separately from how weights are stored. Set it and the node stops casting every layer, so an fp8 model can compute in bf16.
• attention - the attention backend for this model only: default, pytorch, sage, split, sub_quad, whatever's installed on your machine (only actually-available backends appear in the list). default just uses what ComfyUI was launched with. This is the same mechanism the --use-sage-attention / --use-flash-attention flags control, but per-node instead of per-launch.

Optional inputs worth knowing: fp16_accumulation (faster fp16 matmuls on recent NVIDIA cards; silently ignored if your torch build doesn't support it) and the whole sla_* family, which is the interesting part.

The SLA knob is why this node exists

sla is block-sparse attention built for MiniMax H3's long packed sequence - the inference path lightx2v's SLA turbo LoRA was distilled against, which is why that LoRA gives no speedup on its own. Turn it on and you get roughly 3.7x attention throughput at 768p/15s, with the chosen attention backend running alongside it as the dense fall-through for the text refiner and masked calls. It's skipped, with the reason in the status string, on anything that isn't H3 or on a machine without Triton.

The sub-knobs are all validated at their defaults, and the tooltips are unusually honest about the traps. sla_sparsity defaults to 0.90; break-even is around 0.60, so lowering it below that makes the kernel slower than dense - a loss, not a safe fallback. sla_block_size of 128 forces 1.6 seconds of speech down one attention pattern, so leave it at 64 unless the audio doesn't matter. And the classic "soundtrack is garbage on H3" complaint comes from 4-step distillation, not sparsity - use 6 steps instead of fiddling with this node. Keep sla_protect_audio on; it costs ~7% but stops top-k from dropping the audio prefix entirely.

Outputs

You get a MODEL that wires straight into a KSampler, plus a status string - newline-joined notes about the GGUF backend, dtype choices, and whether SLA installed. It's a genuinely useful diagnostic that most loaders don't bother with.

Install

FunPack is on the Comfy Registry and the regular routes all work:

# ComfyUI Manager: search "ComfyUI-FunPack" and Install
cd ComfyUI/custom_nodes
git clone https://github.com/digital-garbage/ComfyUI-FunPack

Then pip install -r requirements.txt - that pulls gguf, insightface, opencv-python-headless and the rest, with a baseline of transformers >= 5.0.0. The pack ships no models; whatever you load has to be in your diffusion_models folder already.

Gotchas

One thing the loader is quietly proud of: it decides GGUF by file content, not extension, so a GGUF renamed to .safetensors still loads - and it'll tell you in the status string. On the flip side, loading MiniMax H3 at fp16 or fp8_e4m3fn_fast is measured to corrupt the audio first, so keep H3 at bf16 weights. And since FunPack v4, the author updates the Cutting Room editor first and only tests nodes against it - this loader works fine in stock ComfyUI, but treat the whole pack as a solo hobby project, bug reports welcome on GitHub issues rather than expected to be fixed instantly.