DeepGen 1.0: A Lightweight Unified Multimodal Model for Advancing Image Generation and Editing
Paper • 2602.12205 • Published • 82
DeepGen SFT LLM Adapter (50k Steps)
Pretrained lightweight residual adapter (~10.2M parameters, ~80MB) evaluated on DeepGen 1.0 (Qwen2.5-VL-3B + SCB + UniPic2-SD3.5M-Kontext-2B DiT).
🔗 Links & Resources
- Code & Full Research Report: GitHub Repository
- Base Architecture: DeepGen 1.0 Base Repo
- Official Paper: DeepGen Technical Report
🚀 Quick Evaluation
import torch
from huggingface_hub import hf_hub_download
# Download adapter weights automatically
weights_path = hf_hub_download(
repo_id="OrliSpace/deepgen-sft-llm-adapter",
filename="deepgen_sft_llm_adapter_50k.pt"
)
state_dict = torch.load(weights_path, map_location="cpu")
print("Adapter weights loaded successfully!")
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support
Model tree for OrliSpace/deepgen-sft-llm-adapter
Base model
Qwen/Qwen2.5-VL-3B-Instruct