Eric Krea2 Component Loader

A ComfyUI node in Eric/Krea2 with 15 inputs and 2 outputs.

By EricRollei·Created about a month ago·Updated 10 days ago· 5

krea2_pipeline
settings

◄base_pipeline_pathH:\Training\Krea-2-Turbo►

◄loader_presetcustom►

◄transformer_picknone►

◄transformer_path►

◄text_encoder_picknone►

◄text_encoder_path►

◄vae_picknone►

◄vae_path►

◄precisionbf16►

◄attention_backendauto►

◄devicecuda►

◄keep_in_vramtrue►

◄offload_vaefalse►

◄vae_tilingauto►

◄vae_tile_px0►

CategoryEric/Krea2

Inputs (15)

Name	Type	Default	Description
base_pipeline_path	STRING	H:\Training\Krea-2-Turbo	Full Krea 2 diffusers folder. Supplies model_index.json, scheduler, tokenizer, and any component you don't override below. NOTE: this folder's model_index.json is_distilled flag decides the whole sampling regime (Turbo: guidance forced 0, fixed distilled shift; Raw: real CFG + active negative prompt) - a Turbo checkpoint run against the Raw base renders blown/oversaturated. Default is Turbo as the common case (2026-07-28).
loader_presetopt	COMBO	custom	Load a saved loader recipe (paths / precision / attention / device) into the panel for this run. 'custom' = use the panel as-is. Use the ★ Save Preset button to add one; the list refreshes on graph reload.
transformer_pickopt	COMBO	none	Pick a transformer from models/diffusion_models/krea2/. When not 'none' this OVERRIDES the transformer_path field below. Single-file .safetensors/.gguf only - use the path field for a diffusers folder.
transformer_pathopt	STRING		Override transformer: diffusers folder, bare .safetensors/.bin/.pt, or .gguf (dequantized on load). Empty = use base. Ignored when transformer_pick above is set.
text_encoder_pickopt	COMBO	none	Pick a text encoder from models/text_encoders/. When not 'none' this OVERRIDES the text_encoder_path field below.
text_encoder_pathopt	STRING		Override Qwen3-VL text encoder: HF/transformers folder, bare .safetensors, or .gguf. Empty = use base. Ignored when text_encoder_pick above is set.
vae_pickopt	COMBO	none	Pick a VAE from models/vae/. When not 'none' this OVERRIDES the vae_path field below.
vae_pathopt	STRING		Override VAE (AutoencoderKLQwenImage): diffusers folder or bare .safetensors. Empty = use base. Ignored when vae_pick above is set.
precisionopt	COMBO	bf16	Compute dtype (bf16 recommended for Blackwell). GGUF is dequantized to this dtype.
attention_backendopt	COMBO	auto	Transformer attention kernel. auto = flash if available else SDPA.
deviceopt	COMBO	cuda	4 options: cuda, cuda:0, cuda:1, cpu
keep_in_vramopt	BOOLEAN	true	Cache the assembled pipeline between runs.
offload_vaeopt	BOOLEAN	false	Keep the VAE on CPU during transformer inference (saves ~1 GB).
vae_tilingopt	COMBO	auto	Tiled-decode/encode policy for the pipeline VAE, honored by standard_decode/standard_encode everywhere. auto: tile only above 128 latent (>1024 px) - previous behavior. force: always tile. off: never tile - tile-seam A/B diagnostic; large non-tiled Wan-family decodes can leave a bottom-edge band and spike VRAM.
vae_tile_pxopt	INT	00–2048	Custom tile size in PIXELS (0 = VAE default). Needs >=256 to take effect; stride is 75% of tile. If the pattern pitch moves when you move this, the pattern IS tile seams.

Outputs (2)

Name	Type	Description
krea2_pipeline	KREA2_PIPELINE	—
settings	STRING	—