---
title: UniSpace
emoji: 🛰️
colorFrom: blue
colorTo: gray
sdk: gradio
sdk_version: 5.50.0
app_file: app.py
short_description: Unified text-to-image generation and instruction editing
python_version: "3.12"
startup_duration_timeout: 1h
pinned: false
license: mit
---

# UniSpace

Interactive demo of **UniSpace: Unified Visual Representation and Scalable
Multimodal Modeling** — an 8B Qwen3 Mixture-of-Transformers built on a single
patch-reparameterized ("PR") visual tokenizer that handles both image
generation and instruction-based image editing.

* Paper: <https://huggingface.co/papers/2608.08676>
* Weights: <https://huggingface.co/yjb6/UniSpace>
* Code: <https://github.com/yjb6/UniSpace>

## What runs here

The Space loads the released `unispace-sft-0012000` checkpoint together with
the `pr-qwen-vit` tokenizer/decoder and exposes the authors' two reference
inference paths:

| Tab | Reference script | Settings |
|---|---|---|
| Text to image | `eval/gen/gen_images_qwen3_unified_mot.py` | 1024², 50 steps, CFG 10, `timestep_shift=0.112`, `cfg_renorm_min=0.9` |
| Instruction editing | `eval/gen/gen_images_imgedit_qwen3_unified_mot.py` | `--match-ref-size --ref-use-mar` (mar_1024 buckets), 50 steps, CFG 10, `timestep_shift=0.112`, `cfg_renorm_min=0.0` |

Both share one model instance built with the release configuration
(`max_latent_size=96`, `share_unified2llm=False`, `use_spatial_merge_und=True`,
`use_spatial_merge_gen=False`, `latent_patch_size=1`).

The sampling code in `unispace_infer.py` is copied verbatim from those two
scripts; only the argparse/torchrun driver was removed. Model construction
differs from the scripts in two ways that are required by the hosting
environment, not by the model:

1. The base Qwen3-8B weights are never downloaded — every `language_model.*`
   tensor is present in the UniSpace checkpoint, so only its config and
   tokenizer are fetched.
2. The 17-shard fp32 checkpoint is streamed one shard at a time, cast to
   bf16 and released, and the PatchReparam encoder loads only the Qwen3-VL
   vision tower (instead of the full 17.5 GB VLM with `device_map="cuda:N"`,
   which is incompatible with ZeroGPU).

## Credits & licensing

UniSpace code and weights are MIT-licensed (© 2026 UniSpace Authors); the
vendored `unispace/` and `patch-reparameterization/` sources carry the
upstream Apache-2.0 / MIT notices (`LICENSE-UNISPACE`, `unispace/NOTICE`,
`THIRD_PARTY.md`).

The four editing examples in `examples/` are the input images from the
UniSpace project page gallery, which come from **GEdit-Bench** (MIT, see
`THIRD_PARTY.md`). Text-to-image example prompts are taken from the UniSpace
repository's own test prompt set.