🚀

Image Generation - DeepGen-1.0

🛠

Unified Multimodal Image Generation & Editing

A lightweight 5B-parameter unified multimodal model supporting general image generation, general image editing, reasoning image generation, reasoning image editing, and text rendering, all in a single model

What it does

DeepGen 1.0 is a unified multimodal image generation and editing model built on a hybrid VLM + DiT (Diffusion Transformer) architecture. With only 5B parameters (3B VLM + 2B DiT), it integrates five core capabilities in a single model while remaining competitive with or surpassing models 3× to 16× larger. The model is designed to demonstrate that massive scaling is not the only path to high-performance multimodal generation.

Problem it solves

  • Lightweight efficiency – Achieves state-of-the-art performance at just 5B parameters, far smaller than comparable unified models (e.g., BAGEL at 14B, Hunyuan-Image-3.0 at 80B)
  • Unified capabilities – Combines generation, editing, reasoning generation, reasoning editing, and text rendering in one model, eliminating the need for separate specialized models
  • Semantic understanding – Overcomes limitations of lightweight models via Stacked Channel Bridging (SCB), enabling fine-grained semantic control
  • Reasoning-aware generation – Supports complex, knowledge-intensive prompts that require multi-step reasoning before generating or editing an image
  • Creative workflows – Suitable for design, content creation, and multimodal applications requiring both generation and editing

Input/Output

  • Input:
    • Text-to-image: Natural language text prompt describing the desired image
    • Image-to-image: One or more reference images + a text instruction for editing or transformation
      • Supports: general editing, style transfer, reasoning-based editing, identity preservation
  • Options:
    • Diffusion steps (10 - 50): Diffusion steps control how many refinement iterations the model uses to generate an image (default: 50)
    • Random seed
    • Output image height & width (default: 512×512)
  • Output: Generated or edited image
    • Quality: High-fidelity, semantically accurate, fine-grained detail
      • image

Accuracy & Speed

  • Quality Metrics:

General Image Generation:

Benchmark
Score
Rank
Geneval ↑
0.87
🥈 among open-source models
DPGBench ↑
87.90
🥈 among open-source models
UniGenBench ↑
75.74
🥈 among open-source models

General Image Editing:

Benchmark
Score
Rank
GEdit-EN ↑
7.17
🥉 among open-source models
ImgEdit ↑
4.14
🥉 among open-source models

Reasoning Image Generation:

Benchmark
Score
Rank
WISE ↑
0.73
🥇 among open-source models
T2I-CoREBench ↑
46.5
🥈 among open-source models

Reasoning Image Editing:

Benchmark
Score
Rank
RISE ↑
10.8
🥉 among open-source models
UniREditBench ↑
75.7
🥈 among open-source models

Technical Details

Architecture
Hybrid VLM (3B) + Diffusion Transformer / DiT (2B)
VLM Backbone
Qwen2.5-VL-3B-Instruct (fine-tuned)
Key Innovation
Stacked Channel Bridging (SCB): extracts hierarchical features from multiple VLM layers, fused with learnable "think tokens" to guide the generative backbone
Training Strategy
3-stage: (1) Alignment Pre-training on image-text pairs & editing triplets; (2) Joint Supervised Fine-tuning on generation, editing & reasoning tasks; (3) Reinforcement Learning with MR-GRPO for quality and human-preference alignment
Total Parameters
5B (3B VLM + 2B DiT)
Checkpoints Available
Pre-training, SFT, and RL
Supported Tasks
General generation, general editing, reasoning generation, reasoning editing, text rendering

Compliance & Provenance

Provider
Open-source
Provider type
Specialized
License
Apache 2.0
EU AI Act risk class
Limited Risk
Art. 50 transparency
Required — outputs are marked. See AI Policy §2.
Region availability
Available globally
Training data summary
Pending — provider has not yet published per Art. 53(d)

For more on how we classify models and mark outputs, see our AI Policy.

Model Source