Diffusion Text-to-Image
π Reading Time: ~13 min read
π
Updated 2026-03
Stable Diffusion 3.5 Large: MMDiT Architecture, T5 Text Encoder & ComfyUI Workflow Guide
Stability AI SD 3.5 Large (8B) multimodal diffusion transformer. Multi-modal DiT attention, FP8 text encoder acceleration, and photorealistic generation pipelines.
π¨ SD 3.5 Large and MMDiT architecture analysis
Stable Diffusion 3.5 Large uses a multimodal diffusion transformer (MMDiT, Multimodal Diffusion Transformer). Text embedding and image latent variables are processed using independent Transformer weights and intersected in both directions at the attention layer, which greatly improves the accuracy of complex compositions and English font rendering.
β‘ Image pulling and ComfyUI directory archiving
export HF_ENDPOINT="https://hf-mirror.net"
COMFY_DIR="/data/ComfyUI"
# Download SD 3.5 Large weight (approx. 16GB)
huggingface-cli download stabilityai/stable-diffusion-3.5-large sd3.5_large.safetensors \
--local-dir "$COMFY_DIR/models/checkpoints" \
--local-dir-use-symlinks False
# Download dedicated text encoder (CLIP G / CLIP L / T5XXL)
huggingface-cli download stabilityai/stable-diffusion-3.5-large text_encoders/t5xxl_fp8_e4m3fn.safetensors \
--local-dir "$COMFY_DIR/models/clip" \
--local-dir-use-symlinks False