Nodes/ComfyUI Easy Use/EasyLoader (Full)
ComfyUI Node Runs on cloud

EasyLoader (Full)

One node that loads your checkpoint, LoRA, prompts and latent

By yolain·Created 3 years ago·Updated 8 days ago· 2,633
EasyLoader (Full)
  • model_override
  • clip_override
  • vae_override
  • optional_lora_stack
  • optional_controlnet_stack
  • pipe
  • model
  • vae
  • clip
  • positive
  • negative
  • latent
ckpt_name
config_nameDefault
vae_name
clip_skip-2
lora_name
lora_model_strength1.00
lora_clip_strength1.00
resolution512 x 512
empty_latent_width512
empty_latent_height512
positive
positive_token_normalization
positive_weight_interpretation
negative
negative_token_normalization
negative_weight_interpretation
batch_size1
a1111_prompt_stylefalse

This is the node that collapses your whole "start of workflow" mess into one box. Instead of a Load Checkpoint, a couple of CLIP Text Encodes, a LoRA Loader, and an Empty Latent Image all wired together, easy fullLoader does all of it and hands you a single pipe you can carry through the rest of the graph. It's the flagship loader of the Easy-Use pack, and the pack itself descends from TinyTerra's "pipe" idea - bundle a bunch of related outputs into one wire so your canvas stops looking like spaghetti.

If you've ever opened someone's ComfyUI workflow and immediately gotten lost, this is the antidote. One loader, everything configured in place.

How it works

You pick a checkpoint, optionally a LoRA, type your positive and negative prompts right on the node, set a resolution, and it runs the encode + empty-latent step for you. The clever part is the pipe output: it packs model, clip, vae, conditioning, and latent into one connection that Easy-Use's samplers (easy preSampling, easy kSampler) know how to unpack. But it also breaks everything back out as individual outputs, so you're never locked into the pipe if you want to wire a normal KSampler.

The inputs and outputs that matter

  • ckpt_name - your checkpoint. The one required pick.
  • positive / negative - multiline prompt boxes, encoded right here.
  • clip_skip (default -2) - note the default. -2 is the standard for SDXL and most anime/illustration checkpoints; the pack sets it there deliberately (older versions defaulted to -1 and it was changed on purpose).
  • lora_name + lora_model_strength / lora_clip_strength - bake one LoRA in without a separate node. Need more than one? Feed the optional optional_lora_stack input from an easy loraStack.
  • a1111_prompt_style (default false) - flip this to reproduce the exact image A1111/WebUI would give you from the same prompt. Real catch: it only works if you also have ComfyUI_smZNodes installed. Without it, leave it off.

Outputs: pipe (feed to Easy-Use samplers) plus model, vae, clip, positive, negative, latent broken out individually. You can also override the model/clip/vae with the model_override / clip_override / vae_override inputs if you loaded them elsewhere.

How to install it

ComfyUI Manager is easiest - open Manager, search ComfyUI-Easy-Use, install, restart. Or from the command line:

cd ComfyUI/custom_nodes
git clone https://github.com/yolain/ComfyUI-Easy-Use

then run install.bat (Windows) or pip install -r requirements.txt in the folder, and restart ComfyUI. The loader itself pulls no extra models - it uses the checkpoints and LoRAs already in your models/ folders.

Common issues & troubleshooting

a1111_prompt_style does nothing. It's a no-op without smZNodes. Install that pack first.

It's overkill for Flux/SD3. fullLoader shines on SD1.5 and SDXL where the checkpoint carries model+clip+vae together. For Flux you're loading UNet, CLIP, and VAE separately anyway, so reach for easy fullLoader's Flux-oriented siblings instead - this one's built around the classic checkpoint shape.

Prompts look weaker than a plain CLIP Text Encode. Check positive_weight_interpretation - it defaults to comfy, but if you're copying prompts from A1111 with (emphasis:1.3) syntax you'll want the A1111 setting so the weights parse the way you expect.

Node resizes itself when you switch LoRA. That's intentional - it redraws to show the LoRA fields. Not a bug.

CategoryEasyUse/Loaders

Inputs (23)

NameTypeDefaultDescription
ckpt_nameCOMBO1 options: None
config_nameCOMBODefault12 options: Default, anything_v3.yaml, v1-inference.yaml, v1-inference_clip_skip_2.yaml, v1-inference_clip_skip_2_fp16.yaml, v1-inference_fp16.yaml, +6
vae_nameCOMBO1 options: Baked VAE
clip_skipINT-2-24–0
lora_nameCOMBO1 options: None
lora_model_strengthFLOAT1.00-10–10
lora_clip_strengthFLOAT1.00-10–10
resolutionCOMBO512 x 51231 options: width x height (custom), 512 x 512, 512 x 768, 576 x 1024, 768 x 512, 768 x 768, +25
empty_latent_widthINT51264–16384
empty_latent_heightINT51264–16384
positiveSTRING
positive_token_normalizationCOMBO4 options: none, mean, length, length+mean
positive_weight_interpretationCOMBO5 options: comfy, A1111, comfy++, compel, fixed attention
negativeSTRING
negative_token_normalizationCOMBO4 options: none, mean, length, length+mean
negative_weight_interpretationCOMBO5 options: comfy, A1111, comfy++, compel, fixed attention
batch_sizeINT11–4096The number of latent images in the batch.
model_overrideoptMODEL
clip_overrideoptCLIP
vae_overrideoptVAE
optional_lora_stackoptLORA_STACK
optional_controlnet_stackoptCONTROL_NET_STACK
a1111_prompt_styleoptBOOLEANfalse

Outputs (7)

NameTypeDescription
pipePIPE_LINE
modelMODEL
vaeVAE
clipCLIP
positiveCONDITIONING
negativeCONDITIONING
latentLATENT