Diffusers Pipeline Walkthrough + Speed/VRAM Notes

Install (use source builds)

The official docs recommend installing Transformers + Diffusers from source for GLM-Image. (GitHub)

pip install git+https://github.com/huggingface/transformers.git
pip install git+https://github.com/huggingface/diffusers.git

Minimal text-to-image code

import torch
from diffusers.pipelines.glm_image import GlmImagePipeline

pipe = GlmImagePipeline.from_pretrained(
    "zai-org/GLM-Image",
    torch_dtype=torch.bfloat16,
    device_map="cuda",
)

prompt = 'A modern poster with the headline "SPRING SALE" and the CTA "SHOP NOW".'
img = pipe(
    prompt=prompt,
    width=1024,
    height=1024,
    num_inference_steps=50,
    guidance_scale=1.5,
    generator=torch.Generator(device="cuda").manual_seed(42),
).images[0]

img.save("glm_image.png")

The pipeline class and call signature are documented in Diffusers. (Hugging Face)

Important defaults (don't "over-fix" randomness)

Diffusers notes the AR part uses sampling defaults (e.g., do_sample=True + temperature), and recommends not forcing deterministic decoding because it can cause degenerate outputs. (Hugging Face)

Resolution rules

Target width/height should be divisible by 32, otherwise you'll error. (GitHub)

Speed & VRAM reality check (H100 reference)

The GLM-Image repo includes measured end-to-end time + peak VRAM on a single H100 (Diffusers). (GitHub)

Examples:

• 1024×1024, batch 1 (T2I): ~64s, ~37.8GB VRAM (GitHub)
• 512×512, batch 1 (T2I): ~27s, ~34.3GB VRAM (GitHub)

They also note inference optimization is still limited and may require very large VRAM for practical use. (GitHub)

Practical settings for most people

• Start at 512 or 768
• Use guidance_scale ~ 1.5–4.0 (fal.ai suggests this range for balance) (Fal.ai)
• Keep poster text blocks short and structured