---
title: Zeta-Chroma Test Bench
emoji: 🎨
colorFrom: purple
colorTo: indigo
sdk: gradio
sdk_version: 5.33.0
app_file: app.py
pinned: false
hardware: zero-a10g
tags:
- ml-intern
---

# 🎨 Zeta-Chroma Test Bench

Generates images from [lodestones/Zeta-Chroma](https://huggingface.co/lodestones/Zeta-Chroma) — a **pixel-space** diffusion model based on the Z-Image architecture.

This Space runs a **custom inference pipeline** re-implemented from the [ComfyUI source code](https://github.com/Comfy-Org/ComfyUI/blob/main/comfy/ldm/lumina/model.py), since this model variant is not yet supported by the `diffusers` library.

## How it works

- **Model:** `NextDiTPixelSpace` — 6.5B parameter pixel-space diffusion transformer with `dec_net` decoder
- **Text encoder:** Qwen3 4B (layer -2, chat template) from [Tongyi-MAI/Z-Image](https://huggingface.co/Tongyi-MAI/Z-Image)
- **Scheduler:** Custom Euler sampler with AuraFlow shift=8.0 and beta sigmas (α=0.6, β=0.6)
- **No VAE** — model generates RGB pixels directly (patch_size=32)
- **Free** — runs on ZeroGPU, uses the visitor's own daily GPU quota

## Features

- **🧪 Test Suite** — 4 challenging furry art prompts (hyper-realism → toony) with fixed seeds for cross-version comparison
- **🎨 Custom Prompt** — full control over prompt, negative prompt, resolution, steps, CFG, shift, and seed
- **🔖 Checkpoint tracking** — displays the model's commit SHA so you know which training checkpoint is active
- **📡 Webhook** — auto-triggers on model updates (configure at [webhook settings](https://huggingface.co/settings/webhooks))

## Resolution presets

All resolutions are multiples of the model's patch_size=32:

| Ratio | Landscape | Portrait |
|-------|-----------|----------|
| 1:1 | 512², 768², 1024² | — |
| 4:3 | 1024×768 | 768×1024 |
| 16:9 | 1024×576 | 576×1024 |
| 5:3 | 1280×768 | 768×1280 |
| 5:4 | 1280×1024 | 1024×1280 |

## Privacy

All generations are private to your browser session. No images are stored server-side or shared between users. GPU time is billed to *your* account quota, not the Space owner's.

## ⚠️ WIP Model

This is a work-in-progress training checkpoint. Image quality and prompt adherence will improve as training progresses. The webhook will automatically pick up new checkpoint versions.