AIHub Expose Model

AIHub Expose Model is the node that lets the external app's user pick a model and attach LoRAs to it, and it outputs the full MODEL, CLIP, VAE trio your sampling needs. It's also the node the pack's own README tells you to think twice about: "Use of this selector is discouraged unless you want to specify an initial value, since it can be rather complicated to setup." That's the author being honest, so here's the honest version of the trade-off.

The complicated part comes from how model selection works in AIHub. To be selectable by a client, a model has to be exported - a JSON config file describing it (id, family, group, file, default cfg/steps/sampler/scheduler) lives in ComfyUI/aihub/models. You generate these with the AIHubMetaExportModel node or write the JSON by hand. On top of that, the node decides whether it's a regular checkpoint (loads from models/checkpoints) or a diffusion model (loads from models/diffusion_models via UNETLoader, with an fp8 weight-dtype choice). Add per-model LoRA lists, and you're juggling five moving parts just to show a dropdown.

What you actually set

• model - the default checkpoint to use, as a base value. The client can override it.
• loras / loras_strengths / loras_use_loader_model_only - comma-separated defaults for the LoRA stack: names, matching strengths, and a t/f per LoRA for whether it should skip the clip (loader-model-only mode). All three lists must line up.
• is_diffusion_model - and here's the trap: the default is true. Most people's checkpoints are not diffusion models. For anything in models/checkpoints, flip this off or the node goes hunting in the wrong folder.
• diffusion_model_weight_dtype - default, fp8_e4m3fn, fp8_e4m3fn_fast, fp8_e5m2. Only meaningful when is_diffusion_model is on; fp8 is how you run a diffusion model you can't fit in VRAM at full precision.
• limit_to_family / limit_to_group - narrow the picker to, say, family sdxl or group fancy_animal_model, so beginners can't select a model that breaks the workflow.
• disable_model_selection / disable_loras_selection - lock either half of the picker down.

Outputs are MODEL, CLIP, VAE - straight into a KSampler, exactly like a checkpoint loader's.

The one-liner verdict

The README's bluntest line is the real guidance: the standard node's actual use is selecting LoRAs for a specific model. If all you want is a model dropdown, use the Simple variant instead - it strips all of this to the essentials. Reach for this one when the workflow genuinely must let a client stack LoRAs onto a family of models. One friendly detail from the source: you can leave model, loras, and friends blank and the client-side UI just picks the first option it gets - handy for a quick default.

Install & gotchas

Pack install is the usual: ComfyUI Manager → search ComfyUI-aihub-workflow-exposer, or git clone https://github.com/otavanopisto/ComfyUI-aihub-workflow-exposer into custom_nodes, then restart. No requirements.txt, no extra pip deps. The model JSONs you export live in ComfyUI/aihub/models; keep them in sync if you rename files.

Biggest gotcha besides the is_diffusion_model default: lora lists that don't match counts. Three loras but two strengths and the client will send something your workflow chokes on. And remember exporting only matters for client-side selection - if the workflow is pinned to one model, the README notes you can skip all this and load directly with the AIHub Utils Load Model node on a string input.