---
title: Agate Preview 002 4-step Live
emoji: ⌨️
colorFrom: red
colorTo: gray
sdk: gradio
sdk_version: 6.22.0
python_version: "3.12"
app_file: app.py
pinned: false
license: mit
short_description: A new image on every keystroke, 4-step Agate on GPU
models:
- ML-Intern-lab/agate-preview-002-4step
- Logolabs/agate-preview-002
- Falconsai/nsfw_image_detection
---

# Agate 4-step, live

Type a prompt and the image redraws as you type. The model is
[ML-Intern-lab/agate-preview-002-4step](https://huggingface.co/ML-Intern-lab/agate-preview-002-4step), an **unofficial** 4-step, guidance-free
distillation of LogoLabs' [Agate Preview 002](https://huggingface.co/Logolabs/agate-preview-002): 4 network passes per
256 px image instead of 100.

## Runs on any hardware

The app picks its serving mode and precision from the hardware at startup; switch hardware in the Space settings, no
code changes needed.

| hardware | precision | how keystrokes are served |
|---|---|---|
| ZeroGPU | bf16 | the first keystroke opens one GPU session that stays open while you type; it streams a new image for each prompt change and closes after 6 s without typing (each ZeroGPU call costs ~1.2 s of CUDA warm-up, so one call per keystroke would be slow) |
| T4 | fp16 (Turing has no bf16) | one request per keystroke, model warm, each step replayed from a CUDA graph |
| L4, A10G, L40S, A100, H100, ... | bf16 | same as T4 |
| CPU | fp32 | same, slowly |

Measured: ZeroGPU about 170-200 ms per image inside a live session (plus ~2 s to open one); T4 about 190 ms per image,
about 0.5 s per keystroke end to end from a test client with a ~285 ms network round trip to the Space.

Other details: `trigger_mode="always_last"` drops keystrokes typed during a render except the newest; images are sent
inline as WebP so the browser needs no second request; every image carries Agate's invisible watermark (`AGATE002`);
an NSFW classifier (`Falconsai/nsfw_image_detection`) blurs flagged outputs.

In-browser version: [ML-Intern-lab/agate-preview-002-4step-webgpu](https://huggingface.co/spaces/ML-Intern-lab/agate-preview-002-4step-webgpu).
Not affiliated with LogoLabs. Licence: MIT.