EasyLoader (Kolors)(kolorsLoader)

The EasyLoader (Kolors) node (kolorsLoader) loads a complete Kolors inference pipeline in one step: the UNet, the VAE, the ChatGLM3 text encoder, and an optional LoRA. It is listed in the EasyUse/Loaders category. Besides returning the prepared MODEL and VAE, it also builds a PIPE_LINE that downstream EasyUse nodes can consume directly.

Description

Kolors is a text-to-image model family that uses ChatGLM3 as its text encoder instead of a standard CLIP-based encoder. This loader is tailored to that architecture: it loads the required model pieces and constructs an EasyUse pipeline, a model, and a VAE in a single node. Because ChatGLM3 is required for prompt encoding, the chatglm3_name widget is mandatory.

Inputs

Required

• unet_name (COMBO): Select the Kolors UNet file to load.
• vae_name (COMBO): Select the VAE file to use for decoding latents.
• chatglm3_name (COMBO): Select the ChatGLM3 text encoder file used to encode positive and negative.
• lora_name (COMBO): Select a LoRA to apply. The available options are the LoRA files known to ComfyUI.
• lora_model_strength (FLOAT, default=1, min=-10, max=10, step=0.01): Strength applied to the LoRA on the diffusion model weights. 0 disables the model-side LoRA effect; negative values invert it.
• lora_clip_strength (FLOAT, default=1, min=-10, max=10, step=0.01): Strength applied to the LoRA on the ChatGLM3 text encoder weights. 0 disables the text-encoder-side LoRA effect; negative values invert it.
• resolution (COMBO, default="1024 x 576"): A preset shortcut that sets empty_latent_width and empty_latent_height to standard values. Custom dimensions can still be entered manually after selecting a preset.
• empty_latent_width (INT): Width of the empty latent to create. This is typically set by the resolution combo, but can be overridden for custom sizes.
• empty_latent_height (INT): Height of the empty latent to create. This is typically set by the resolution combo, but can be overridden for custom sizes.
• positive (STRING, default=""): The positive prompt text. It is encoded using the selected ChatGLM3 text encoder.
• negative (STRING, default=""): The negative prompt text. It is encoded using the selected ChatGLM3 text encoder.
• batch_size (INT, default=1, min=1, max=64): Number of latents to create in the pipeline. Higher values use more VRAM.

Optional

• model_override (MODEL): When connected, the node uses this model instead of loading a UNet from unet_name.
• vae_override (VAE): When connected, the node uses this VAE instead of loading one from vae_name.
• optional_lora_stack (LORA_STACK): Connect a LoRA stack here to apply multiple LoRAs alongside the primary lora_name selection.
• auto_clean_gpu (BOOLEAN, default=false): Enable to attempt an automatic GPU cleanup after the load, which can reduce VRAM pressure when switching models frequently. Disable it when you want to keep the loaded model cached in memory for repeated samples.

Outputs

• PIPE_LINE: An EasyUse pipeline object containing the loaded model, VAE, prompts, latent width/height, and batch size. Connect this to pipeline-aware EasyUse nodes for sampling.
• MODEL: The effective model after loading and applying the selected LoRA / LoRA stack / model override. Use this with standard ComfyUI sampling nodes.
• VAE: The effective VAE after loading and applying any VAE override. Use this for decoding latents to images.

Usage Notes

• Kolors uses ChatGLM3 rather than the usual CLIP text encoders. To change the prompt encoder, use chatglm3_name.
• The resolution menu is a convenience for setting empty_latent_width and empty_latent_height. Selecting a preset overwrites those fields, so enter custom values afterward if you need a different aspect ratio. Keep the dimensions compatible with the latent format — multiples of 8 or 16 are commonly required.
• LoRA strengths are independent: lora_model_strength affects the diffusion model weights, while lora_clip_strength affects the ChatGLM3 text encoder weights. A value of 0 disables that side of the LoRA, and negative values are accepted by the widget.
• If you need multiple LoRAs, connect optional_lora_stack; the primary lora_name LoRA is still applied unless its strengths are set to 0 or a no-op entry is selected.
• Use model_override and vae_override when you already have a model or VAE loaded elsewhere in the graph. This avoids reloading the same files from disk and lets you reuse resources across multiple loader configurations.