EasyLoader (PixArt)(pixArtLoader)

The EasyLoader (PixArt) node (pixArtLoader) is a convenience loader for PixArt models in the ComfyUI-Easy-Use package. It loads all components needed for PixArt — the diffusion model (from a checkpoint or standalone model file), VAE, CLIP text encoder, and SD3-style T5 text encoder — then builds prompt conditioning from the positive and negative text fields, creates empty latents at the requested dimensions and batch size, applies LoRA modifications, and packages everything into an Easy-Use pipeline.

Description

This node is designed as a one-stop loader for PixArt workflows. Instead of chaining separate checkpoint, VAE, CLIP, T5, LoRA, and empty-latent nodes, you can use EasyLoader (PixArt) to prepare a complete PIPE_LINE for sampling. It also exposes the raw MODEL and VAE outputs if you need to connect them elsewhere.

The node expects raw text prompts, not pre-computed conditioning. The positive and negative strings are encoded internally using the selected CLIP and T5 encoders.

Inputs

Required inputs

• ckpt_name (COMBO): Selects a checkpoint file containing PixArt weights. Choose the checkpoint that matches your installed model files.
• model_name (COMBO): Selects a standalone PixArt diffusion model file. Use either this or ckpt_name, depending on how your PixArt weights are stored.
• vae_name (COMBO): Selects the VAE used to decode latents into images.
• t5_type (COMBO): Text encoder type used for the T5 portion of PixArt conditioning. Currently the only available option is sd3; PixArt relies on the SD3-style T5 text encoder.
• clip_name (COMBO): Selects the CLIP text encoder used alongside T5 for conditioning.
• padding (INT, default 1, range 1–300): Sets the padding length used when encoding text with the T5/SD3 text encoder. Higher values add padding tokens around the prompt.
• t5_name (COMBO): Selects the T5 model file used for text encoding.
• device (COMBO, default cpu): Determines where the loaded components run. Options:
  • auto: let ComfyUI choose the device automatically.
  • cpu: force CPU execution.
  • gpu: force GPU execution.
• dtype (COMBO): Controls the precision used for the loaded model. Options:
  • default: use the model’s default precision.
  • auto (comfy): let ComfyUI decide the appropriate precision.
  • FP32: full float32 precision.
  • FP16: half precision, useful for reducing VRAM usage.
  • BF16: bfloat16 precision, useful on supported hardware.
• lora_name (COMBO): Selects a single LoRA file to apply to the PixArt model.
• lora_model_strength (FLOAT, default 1, range -10 to 10, step 0.01): Controls how strongly the selected LoRA affects the model. Negative values are allowed.
• ratio (COMBO, default "1.00"): Selects an aspect ratio preset by name. The default "1.00" corresponds to a square 1:1 aspect ratio.
• empty_latent_width (INT): Width of the generated empty latent in pixels. No default is set, so specify the desired width.
• empty_latent_height (INT): Height of the generated empty latent in pixels. No default is set, so specify the desired height.
• positive (STRING, default ""): Positive text prompt used for conditioning.
• negative (STRING, default ""): Negative text prompt used for conditioning.
• batch_size (INT, default 1, range 1–64): Number of latent samples to create in the empty latent batch.

Optional inputs

• optional_lora_stack (LORA_STACK): A stack of additional LoRAs to apply. This is useful when you need to apply more than one LoRA to the model, or when using a LoRA stack from another Easy-Use node.

Outputs

• PIPE_LINE: An Easy-Use pipeline object containing the loaded model, VAE, conditioning, latents, and other settings needed by downstream Easy-Use sampler nodes.
• MODEL: The loaded PixART diffusion model with LoRA changes applied.
• VAE: The loaded VAE, available for direct connection to decoder or sampler nodes.

Usage Notes

• The node does not expose a separate CONDITIONING output. Conditioning is built internally from positive and negative and included in the PIPE_LINE.
• If you load a model using model_name instead of ckpt_name, make sure your VAE, CLIP, and T5 selections match the model’s requirements.
• PixArt uses an SD3-style T5 text encoder, so t5_type is currently fixed to sd3.
• The padding value is applied during T5 text encoding. If you are working with long prompts, you may need to increase it to avoid truncating or misaligning prompt tokens.
• device defaults to cpu; switch to gpu or auto if your hardware supports faster T5/CLIP encoding.
• ratio is a convenience aspect-ratio preset. Actual output latent resolution is controlled by empty_latent_width and empty_latent_height.
• Use optional_lora_stack when applying multiple LoRAs; LoRA strength values can be negative to invert or reduce the effect.