EasyLoader (Comfy)

A vanilla ComfyUI txt2img graph starts with a little pile of nodes: load checkpoint, maybe a separate VAE, a LoRA loader, two CLIP text encodes, an empty latent. easy comfyLoader collapses all of that into a single node that spits out a pipe - one fat wire carrying model, CLIP, VAE, both conditionings, and the latent - which the rest of the EasyUse chain reads from. Fewer wires, less spaghetti, faster to build. That's the entire philosophy of yolain's ComfyUI-Easy-Use pack, and this loader is the front door for standard SD1.5/SDXL work.

There are a few flavors of EasyUse loader; "comfy" is the one that parses prompts the ComfyUI-native way. Its sibling a1111Loader mimics Automatic1111's prompt weighting instead, for people chasing webui-identical results. If you don't have a specific reason to reproduce A1111, comfyLoader is the sane default.

How it works

It loads your ckpt_name checkpoint, optionally applies a LoRA (lora_name plus its model/clip strengths), encodes your positive and negative prompts, and builds an empty latent at the resolution you pick (or custom empty_latent_width/empty_latent_height and batch_size). All of that gets bundled into the pipe output. It also breaks out model and vae separately, so you can tap those directly if some node needs the raw model rather than the pipe.

The vae_name defaults to "Baked VAE," meaning it uses the VAE baked into your checkpoint - fine for most models. Override it only if you've got a checkpoint with a known-bad VAE and a better one to swap in.

One value worth understanding: clip_skip, which defaults to -2. Clip skip stops CLIP a layer or two early; -2 (skip the last layer) is the community-standard setting for SDXL and most anime/booru-lineage models, which is why the pack changed the default from -1 to -2 in v1.2.6. Leave it at -2 unless a model card tells you otherwise.

The inputs a beginner actually sets

• ckpt_name - your checkpoint.
• positive / negative - your prompts, right here in the loader.
• resolution - pick a preset (512×512 for SD1.5, a 1024-ish preset for SDXL) or choose "width x height (custom)" and set the latent dims.
• lora_name - optional; None if you're not using one.

Optional optional_lora_stack and optional_controlnet_stack inputs let you feed in stacks built by other EasyUse nodes, so you can apply several LoRAs or ControlNets without cramming them all into this one node.

Installing it

ComfyUI Manager → search "ComfyUI Easy Use" → install → restart. Or cd ComfyUI/custom_nodes && git clone https://github.com/yolain/ComfyUI-Easy-Use, run install.bat / pip install -r requirements.txt, restart. First launch may switch your theme to the pack's Obsidian UI - Settings → Color Palette to revert.

Common issues

The pipe is the thing people trip on. Downstream EasyUse nodes (easy preSampling, easy kSampler) expect that pipe, so wire the pipe output into them - not the raw model. If you mix EasyUse loaders with vanilla samplers, you'll be pulling model/vae out separately and rebuilding conditioning by hand, which defeats the point.

Empty ckpt_name dropdown means no checkpoints found - check your models/checkpoints folder. And if your images look flat or oddly toned, sanity-check clip_skip against what the model actually wants; -2 is right for most, but a model trained expecting -1 will drift if you force -2 on it.