Models & Sampling

Despite the name, the Models & Sampling node never loads a model. That's its whole personality: it's the intelligence-and-planning node of the pack. The mode dropdown's 14 operations split into three jobs - model intel (read .safetensors metadata without loading it, fingerprint a file, list installed models in a folder, extract LoRA trigger words), resolution planning (aspect ratio picker, SDXL bucket picker, megapixel calculator, snap-to-multiple), and seed/sampler tooling (seed cycle, multi-seed batch, seed history, sampler/scheduler pickers, a sampler-params bundle). Outputs are model, clip, vae, name, sampler, seed, dim, cfg, and params (DICT).

The two modes I'd genuinely reach for: safetensors_metadata and lora_extract_triggers. The first reads the JSON header of a .safetensors file - no tensors, no VRAM - and hands you the __metadata__ section plus tensor count. That's how you audit a downloaded model before trusting it, check its base model, or read its training info; the pack even ships a 52_safetensors_audit.json workflow for exactly that. The second pulls trigger words out of a LoRA's metadata (ss_tag_frequency, modelspec.trigger_words, and friends), which is the classic "I downloaded this LoRA and can't remember what to type" fix. Both read from a plain path - point them at your models/loras or models/checkpoints folder.

How it works

Same dispatch skeleton as the rest of the pack. The honest caveat is the model/clip/vae outputs: no mode produces them, so they're always empty. This node plans your sampling; it doesn't run it. The outputs that actually carry data:

• safetensors_metadata, list_installed_models, multi_seed_batch, seed_history, sampler_params_bundle → the params DICT.
• lora_extract_triggers, model_fingerprint, scheduler_pick → the name STRING.
• aspect_ratio_pick, sdxl_bucket_pick, megapixel_calculator → width and height land in the seed and dim INT slots (generic socket naming - don't let it throw you), with details in params.
• seed_cycle → the seed slot; sampler_pick → sampler; sampler_params_bundle → cfg plus the full params dict.

A few inputs worth knowing: aspect_ratio_pick takes a preset (eight ratios) and long_side; megapixel_calculator takes target_megapixels + ratio_w/ratio_h; seed_cycle takes seed and a cycle_mode - fixed, increment, decrement, random, or from_string, the last of which derives a seed from text via SHA-256 so the same phrase always gives the same seed.

Installing it

Ships with the pack:

cd ComfyUI/custom_nodes
git clone https://github.com/IxMxAMAR/ComfyUI-Utility-MegaPack

Restart ComfyUI, or ComfyUI Manager → "Utility-MegaPack". No model files and no heavy deps - this node deliberately avoids loading models, so it's cheap.

Where people get tripped up

The big one is expecting it to be a model loader: it isn't, and the empty model/clip/vae sockets are a feature, not a bug. Wire name/seed/params into a normal checkpoint loader and KSampler instead. Second, aspect_ratio_pick snaps both dimensions to multiples of 8 - SD UNet/VAE demand it, and older versions produced 1000×562 from long_side=1000, which crashed the model. If you need odd sizes, use snap_to_multiple and pick your own multiple. And if you're coming from the seed-cycling scene, note the sampler/scheduler dropdowns are fixed lists (euler, euler_ancestral, dpmpp_2m, karras, sgm_uniform, and so on) - they're pickers that emit strings, so anything you type into a KSampler by hand still works, but for consistency it's nicer to source it from this node's dropdown.