REQUIREMENTS:

• ComfyUI-BitDance Custom Node Repository

• BitDance Main Model FP8 (Place in models/diffusion_models)

• BitDance Text Encoder FP8 (Place in models/text_encoders)

• BitDance VAE FP16 (Place in models/vae)

DESCRIPTION:

• Low VRAM: Utilizing the built-in "model to offload" setting in the Text Encode node, this workflow automatically shifts the massive main model off your GPU right before heavy text encoding begins, saving your hardware from crashing.

• Speed: By quantizing to fp8 e4m3fn scaled and routing the main model to the graphics card, generation times are drastically reduced on standard consumer hardware.

• Quality: You retain near-full visual fidelity. Using the required euler_maruyama sampler, it flawlessly decodes the massive Binary Tokenizer (2^256 states) to generate perfect micro-details, accurate text (like reading a newspaper), and complex material blending from simple everyday English prompts.

LINKS: Video Tutorial: https://www.youtube.com/watch?v=4O9ATPbeQyg

| Full Documentation: https://aistudynow.com/how-to-fix-the-generic-face-bug-in-bitdance-14b-optimize-speed/

Models: https://huggingface.co/comfyuiblog/BitDance-14B-64x-fp8-comfyui/tree/main

Github: https://github.com/aistudynow/Comfyui-bitdance