tinyLoader

tinyLoader is the node that replaces the boring three-node intro every ComfyUI workflow starts with. Where you'd normally chain Load Checkpoint → CLIP Text Encode → Empty Latent, this one checkpoint and the empty latent it comes with in one box - and it hands you the width and height as actual outputs, which is the part you didn't know you wanted until you've hand-typed a resolution into three different nodes and prayed they matched.

It's part of the tinyterraNodes pack, which is otherwise best known for its xyPlot tools and pipe nodes. tinyLoader is the "keep it simple" cousin: no pipe, no prompt text, just a model, a latent, and the model-math settings that people actually fiddle with.

How it works

Under the hood it calls ComfyUI's standard checkpoint loader, then builds an empty latent from your chosen aspect ratio. Two things make it more than a merged Load Checkpoint + Empty Latent:

• sampling - options like eps, v_prediction, lcm, and x0 patch the model's sampling mode via ModelSamplingDiscrete. This is how you run an LCM checkpoint or a v-prediction model without hunting for a separate patch node. Leave it on Default unless you know your model wants something else.
• zsnr and cfg_rescale_mult - zero-SNR noise scaling and CFG rescaling, applied as a model patch. For most people these stay off/0. If you're testing a model that was trained with zero-SNR, flipping zsnr on is the difference between muddy and clean output.

The inputs you'll actually touch

Most of these you can leave alone, but a few earn their keep:

• ckpt_name - your checkpoint. config_name defaults to Default, which is right for the vast majority of modern checkpoints; the SD1.5-era yaml entries are there for old custom configs.
• clip_skip - defaults to -1, the standard "skip the last CLIP layer" setting people use with many SD1.5 and anime models. Set to 0 to skip nothing.
• empty_latent_aspect - a preset list of common resolutions (512×512, 768×768, 910×910, portrait/landscape ratios, plus the SDXL-style tall ones). Pick width x height [custom] and the two width/height boxes take over.
• vae_name - Baked VAE by default. If your checkpoint ships without a baked VAE, pick the .safetensors VAE from this list and it's loaded instead.

Outputs are model, latent, vae, clip - the standard wiring into any sampler - plus width and height as INTs, which you can feed straight into other nodes that need to know what you're generating at. That alone kills a whole class of "my upscale dimensions don't match" bugs.

Installing it

tinyLoader ships with the whole tinyterraNodes pack. Easiest route is ComfyUI Manager: search for tinyterraNodes (repo name ComfyUI_tinyterraNodes) and install. Or do it by hand:

cd ComfyUI/custom_nodes
git clone https://github.com/TinyTerra/ComfyUI_tinyterraNodes.git

Restart ComfyUI (or hit the reload button in Manager). The pack has no requirements.txt and downloads no models - everything it needs is already in ComfyUI, so install is painless.

Gotchas

Because it's a base-loader node, the usual loader rules apply: if you change empty_latent_aspect mid-workflow, results won't change until you re-run the whole graph. And when a checkpoint gets re-downloaded or swapped, reload the node (right-click → Reload Node, a tinyterra interface feature) so the enum options refresh. One real trap: clip_skip at -1 with a model that doesn't want it - most Flux-family and some newer checkpoints ignore CLIP skip or behave badly with it, so if output looks off, try 0 before blaming the prompt.