EasyCascadeLoader(cascadeLoader)

Description

The EasyCascadeLoader node loads a complete Stable Cascade pipeline from its three stage checkpoints (stage_c, stage_b, stage_a), a CLIP model, an optional LoRA, and a text prompt. It builds the EasyUse pipeline object, creates an empty latent at the requested resolution and compression, and also exposes the model, latent, and VAE for direct use.

Inputs

Required

• stage_c (COMBO): Selects the Stage C checkpoint. Stage C is the text-conditioned prior that creates the initial latent from your prompt.
• stage_b (COMBO): Selects the Stage B checkpoint. Stage B refines/upscales the Stage C latent before decoding.
• stage_a (COMBO): Selects the Stage A checkpoint. Stage A acts as the VAE/decoder that converts the final latent to pixel space. This model is also returned as the vae output.
• clip_name (COMBO): Selects the CLIP model used to tokenize and encode positive and negative into conditioning.
• lora_name (COMBO): Selects a LoRA to apply to the model and CLIP. Leave empty or select the "none" entry to skip the LoRA.
• lora_model_strength (FLOAT, default 1): Strength of the LoRA applied to the diffusion model. Range -10 to 10; negative values reduce or invert the LoRA's influence.
• lora_clip_strength (FLOAT, default 1): Strength of the LoRA applied to the CLIP text encoder. Range -10 to 10.
• resolution (COMBO, default "1024 x 1024"): Convenience preset for empty_latent_width and empty_latent_height.
• empty_latent_width (INT): Width in pixels of the image to generate. This is the target pixel width, not the latent width.
• empty_latent_height (INT): Height in pixels of the image to generate. This is the target pixel height, not the latent height.
• compression (INT, default 42, range 32 to 64): The latent compression factor used by the Cascade Stage A/VAE. The empty latent's spatial size is computed as empty_latent_width // compression by empty_latent_height // compression.
• positive (STRING, default ""): Positive prompt used to build the conditioning stored in the pipe_line output.
• negative (STRING, default ""): Negative prompt used to build the conditioning stored in the pipe_line output.
• batch_size (INT, default 1, range 1 to 64): Number of identical empty latent samples to generate.

Optional

• optional_lora_stack (LORA_STACK): Applies a stack of multiple LoRA configurations in addition to lora_name. Each entry in the stack supplies its own LoRA and model/CLIP strengths.

Outputs

• pipe_line (PIPE_LINE): EasyUse pipeline containing the loaded model, CLIP, VAE, empty latent, and the positive/negative conditioning created from the prompt fields. Use this with EasyUse pipeline-aware nodes.
• model (MODEL): The combined Stable Cascade model, assembled from Stage C/B/A and modified by any selected LoRAs.
• latent (LATENT): The empty latent tensor, with shape based on batch_size, empty_latent_width/height, and compression.
• vae (VAE): The Stage A VAE decoder, usable to decode sampled latents to images.

Usage Notes

• The actual output latent size is not empty_latent_width × empty_latent_height; divide both dimensions by compression. With the default compression of 42, a 1024 x 1024 image corresponds to roughly a 24 x 24 latent.
• The positive and negative prompt fields are encoded into conditioning that lives inside pipe_line. The direct model, latent, and vae outputs do not expose this conditioning or the CLIP model itself; if you need to apply the prompt outside the pipeline output, add a separate conditioning path.
• Keep compression at the value required by your Stage A checkpoint. Changing it to 32 or 64 does not change the model architecture, but it changes the expected latent size and will likely produce bad results unless the checkpoint was trained with that compression.
• The optional LoRA stack is applied on top of lora_name, so you can combine one quick LoRA selected in the main widget with a saved stack of additional LoRAs.
• Use the pipe_line output as the first input to EasyUse nodes such as Easy KSampler for a one-widget text-to-image setup.