Unified multimodal creation

WeGen2 WeGen2 wordmark with an original four-leaf pixel mark; all lettering is vector outlines.

One model. Many worlds.

One model. Many ways to create.

Bring an idea.
See where it goes.

Start with words. Add references. Keep creating in context. WeGen2 brings generation, editing, and visual composition into one native pixel-space model.

61B
Total parameters
7B
Activated per generation token
11
Generation & editing benchmarks
Native
pixels
Directly between RGB and the backbone

The film

A closer look
at what comes next.

From imagined worlds to reference-guided creation, then inside the architecture that brings them together.

Download 1080p ↗Visual transcript ↗

Model-generated images, presented through editorial motion design. Original instrumental score.

01 / Create

A wider
visual vocabulary.

Photographic light, intricate illustrations, tactile materials. Different visual languages, one model.

The Lantern Wood, a fantasy film poster generated with integrated English lettering.
Illustration & typographyThe Lantern Wood
A photographic portrait with an orange umbrella, rain reflections, flowers and a red tram.
Photographic detailA moment in the rain
A paper mountain city with a red train on a curved viaduct.
Material & craftA world made of paper
An intricate oil painting of a lantern-lit night market, stone bridges and glass pavilions.
Painterly scenesLight after dark

Words belong in the picture.

More to see.
More to read.

A WeGen2 creation-studio poster brings imagined landscapes, architectural spaces, and fashion scenes together with Chinese headlines and descriptive text.

Generated from one text prompt. The scenes, layout, and lettering are all part of the model output.

A Chinese illustrated reading guide generated from one text prompt.

And another voice:
a text-to-image reading guide.

A WeGen2 creation-studio poster with six miniature scenes and Chinese text about generation, editing, and composition.
Chinese text & illustrated information design

02 / Compose

Many references.
One new look.

Select a reference to trace its visible counterpart. Individual garments and accessories guide the resulting outfit.

Source person before the reference-guided outfit edit
The starting point
WeGen2 outfit composition guided by the person and eight visual references
WeGen2 output

Select a reference to explore the outfit.

03 / Edit

Keep the idea.
Change its world.

A reference image guides the next creation.

The paper world edited into a moonlit sceneThe original paper city in daylightSourceEdited

04 / Continue

There is always
a next idea.

Add a reply, then revise the conversation. Earlier images become context for successive edits.

Fictional chat interface before the edits
01 Starting context
Chat interface edited to include a reply and a penguin sticker
02 Add a reply
A subsequent edit revises the fictional messages
03 Revise the messages

Visual storytelling

A whole story.
One canvas.

A fox and a little robot make a cake. Eight richly illustrated panels, brought together as one complete comic page.

An eight-panel cake-making comic arranged as four columns and two rows
Eight-panel illustrated story · Panels generated sequentially
Explore another sequence
Robot café scene, image-edit state 1
Source
Robot café scene, image-edit state 2
Edit 1
Robot café scene, image-edit state 3
Edit 2
Robot café scene, image-edit state 4
Edit 3
Robot café scene, image-edit state 5
Edit 4

Five successive image states in a robot café scene. An image-editing demonstration.

Inside WeGen2

Native pixels.
Connected thinking.

Distinct roles for understanding and generation. A shared visual conversation, from context to the next image.

01

Directly in image space.

Learned patch modules connect RGB images and the backbone. No separate pretrained vision encoder or latent autoencoder is used at inference.

A foundation for further research.

61B total · 7B activated per generation token · No reinforcement learning or distillation

Measured across tasks

Top-tier results.
A broader picture.

The report evaluates WeGen2 across 11 generation, editing, and composition benchmarks.

Each benchmark uses its own metric. See the report for individual scores, comparison settings, and prompt-enhancement reporting.

Circular benchmark overview from the WeGen2 report
Generation · Editing · Composition

Start creating

A simple interface.
An open-ended idea.

Python / WeGen2
from wegen2 import WeGen2Pipeline

pipe = WeGen2Pipeline.from_pretrained(
    "./weights/WeGen2", devices="cuda:0,cuda:1"
)

session = pipe.session()
image = session.generate(
    "An intricate paper mountain city, a red train."
)
image.save("paper-world.png")

next_image = session.generate(
    "Change the scene to night. Light the windows."
)
next_image.save("paper-world-night.png")

Words. Images. Context.

A prompt starts a generation. Images provide references. A session retains context for the next edit.

Checkpoint
61B parameters · BF16
H20 planning estimate
2 × NVIDIA H20 · 96 GB each
Estimated weight memory
57.9 / 56.6 GiB across two H20 GPUs

Weights only; runtime memory is additional. Checkpoint-based estimates, not H20 measurements. H20 latency has not been measured.

Read the inference guide ↗

Code in private preview. Weights and final licensing forthcoming.

The team

WeGen2

Native Pixel-Space Modeling for Unified Multimodal Creation

Shaobin Zhuang1,*,§ · Zhipeng Huang2,* · Canmiao Fu2,† · Xiaoqiang Zhou2 · Hao Lu2 · Ying Zhang2 · Chen Li2,‡ · Jing LYU2 · Yali Wang3,‡

1 Shanghai Jiaotong University2 WeChat Vision, Tencent Inc.3 Shenzhen Institutes of Advanced Technology

* Equal contribution. † Project leader. ‡ Corresponding authors.
§ Work done during Shaobin Zhuang's internship at Tencent.

Open original ↗