ComfyUI Node
Eric Qwen-Image Component Loader
A ComfyUI node in Eric Qwen-Image with 11 inputs and 1 output.
Eric Qwen-Image Component Loader
- pipeline
◄base_pipeline_path►
◄transformer_path►
◄vae_path►
◄text_encoder_path►
◄precisionbf16►
◄devicecuda►
◄keep_in_vramtrue►
◄offload_vaefalse►
◄attention_slicingfalse►
◄sequential_offloadfalse►
◄vae_tile_blendcosine►
CategoryEric Qwen-Image
Inputs (11)
| Name | Type | Default | Description |
|---|---|---|---|
| base_pipeline_path | STRING | Path to complete Qwen-Image model (provides config, scheduler, tokenizer, and defaults for unset components) | |
| transformer_pathopt | STRING | Path to custom/fine-tuned transformer. Folder or single .safetensors file. Leave empty to use base. | |
| vae_pathopt | STRING | Path to custom VAE. Leave empty to use base. | |
| text_encoder_pathopt | STRING | Path to custom text encoder. Leave empty to use base. | |
| precisionopt | COMBO | bf16 | Model precision (bf16 recommended for RTX 40/50 series) |
| deviceopt | COMBO | cuda | Device to load model on |
| keep_in_vramopt | BOOLEAN | true | Keep model in VRAM between runs (faster but uses memory) |
| offload_vaeopt | BOOLEAN | false | Move VAE to CPU during transformer inference (saves ~1 GB) |
| attention_slicingopt | BOOLEAN | false | Enable attention slicing (reduces VRAM, slightly slower) |
| sequential_offloadopt | BOOLEAN | false | Aggressive CPU offloading — very slow but handles huge images |
| vae_tile_blendopt | COMBO | cosine | Tile-seam blending for tiled VAE decode/encode. • cosine — C¹-smooth blend (zero slope at tile edges), eliminates faint grid lines on smooth gradients. • linear — original diffusers behaviour. Slightly sharper at tile edges; may show faint seams at high resolutions. |
Outputs (1)
| Name | Type | Description |
|---|---|---|
| pipeline | QWEN_IMAGE_PIPELINE | — |