ComfyUI-SDXS1B

Custom ComfyUI nodes for AiArtLab/sdxs-1b - a fast text-to-image model using a Qwen 3.5 text encoder, modified SD 1.5 UNet (1.5B params), and asymmetric FLUX 2-based VAE.

Nodes

Node	Description
Load SDXS-1B Qwen Text Encoder	Loads the Qwen 3.5-2B text encoder and tokenizer
Load SDXS-1B UNet	Loads the 1.5B param UNet and flow-matching scheduler
Load SDXS-1B VAE	Loads the asymmetric VAE (8x encode / 16x decode)
SDXS-1B Qwen Text Encode	Encodes text with Qwen 3.5 (chat template, layer -2 pooling)
SDXS-1B Sampler	Denoising sampler with progress bar
SDXS-1B VAE Decode	Decodes latents to images

All three loader nodes share a pipeline cache, so the model is only loaded into VRAM once.

Workflow

Qwen Loader --CLIP--> Text Encode (+) -->
                       Text Encode (-) --> Sampler --> VAE Decode --> Preview
UNet Loader --MODEL------------------->
VAE Loader  --VAE--------------------------------------------->

Default settings (from model card)

• Resolution: 1024x1024
• Steps: 40
• CFG: 4.0
• Sampler: Flow Match Euler Discrete

Requirements

• ComfyUI (any recent version)
• Python packages (install in your ComfyUI venv):

pip install diffusers transformers>=4.52 accelerate

transformers>=4.52 is required for Qwen 3.5 model class support.

Installation

1. Install the custom nodes

Copy this folder into your ComfyUI custom_nodes directory:

ComfyUI/
  custom_nodes/
    ComfyUI-SDXS1B/
      __init__.py
      sdxs-1b-workflow.json
      README.md

2. Download the model

The model must be cloned as a full diffusers directory (a single .safetensors file will not work):

cd <your ComfyUI models/diffusers directory>
git lfs install
git clone https://huggingface.co/AiArtLab/sdxs-1b AiArtLab--sdxs-1b

The model directory should contain unet/, text_encoder/, vae/, tokenizer/, scheduler/, and pipeline_sdxs.py.

StabilityMatrix users: The diffusers directory is typically at Data/Models/Diffusers/. If the model doesn't appear in the node dropdown, try placing it in Data/Packages/<your ComfyUI package>/models/diffusers/ instead.

3. Load the workflow

Drag sdxs-1b-workflow.json into ComfyUI. Select AiArtLab--sdxs-1b in each loader node's model_name dropdown.

Troubleshooting

Problem	Solution
ImportError: cannot import name 'Qwen3_5ForConditionalGeneration'	Upgrade transformers: pip install transformers>=4.52
ModuleNotFoundError: No module named 'diffusers'	Install: pip install diffusers accelerate
Model not appearing in dropdown	Ensure the model is a cloned directory (not a single file) in your diffusers models path
VAE decode very slow / hangs	Make sure you're using the latest __init__.py (latent size should use // 16, not // 8)

License

Apache 2.0 (same as the model)