dsh-codex-tools: Adding Codex Search, Image Generation, and Image Recognition Tools to DeepSeek Harness
`dsh-codex-tools` is a plugin for DeepSeek Harness (DSH), maintained by SPYQWER1, providing three tools: `codex_web_search`, `image_gen`, and `image_vision`, which respectively enable public web search summarization, bitmap generation, and local image understanding. This plugin reuses existing ChatGPT/Codex login sessions and does not offer standalone login or LLM services. Search supports customizing the number of sources and freshness; image generation supports multiple sizes and formats, with output paths restricted to the DSH session workspace, prohibiting absolute paths, parent directories, and symbolic links, and not overwriting existing files; image vision recognition only accepts local files within the workspace that meet specific format and size constraints. Installation requires Node>=22 environment, supports adding via Git or npm, and after installation
Read MoreDirect Image Generation in DeepSeek Harness: Getting Started with the dsh-image-gen Plugin
dsh-image-gen is an open-source DeepSeek Harness image generation plugin by shanliuling, which supports direct image generation via the generate_image tool within web dialogs. It is compatible with multiple providers including Gemini, OpenAI, Volcano Ark Seedream, and Tongyi Wanxiang, and comes with built-in gallery management and workspace persistence. This article covers installation commands, API Key configuration, typical usage, as well as plugin permissions and security considerations.
Read MoreGenerate bitmap assets in the coding workflow using Codex's imagegen Skill
# Introduction to OpenAI Codex System-level Skill `imagegen` This skill enables generation and editing of bitmap assets including photos, illustrations, mockups and sprites within coding conversations via either the built-in `image_gen` command or an explicit CLI. Details including default auto-installation location, two operating modes, prompt templates, save path conventions, applicable boundaries and official CLI examples are provided to facilitate directly integrating visual assets into project repositories.
Read MoreExperimental Measurement of Z-Image: An Efficient Image Generation Model with 6B Parameters
Z-Image is an efficient image generation model with 6B parameters, achieving or even surpassing the performance of mainstream competitive models with 8 inference steps (8 NFEs). It can run smoothly on consumer-grade devices with 16G VRAM. The model has three variants: Turbo (lightweight and real-time, suitable for AIGC applications and mini-programs), Base (undistilled for secondary fine-tuning), and Edit (specialized for image editing), with Turbo being the most valuable for practical deployment. In practical tests, the generation time for 1024×1024 resolution is 0.8 seconds (with Flash Attention + model compilation), and the peak memory usage is 14G. Technically, its S3-DiT architecture enhances parameter efficiency, the Decoupled-DMD distillation algorithm enables 8-step inference, and DMDR fuses RL and DMD to optimize quality. Its strengths lie in bilingual text rendering, photorealistic generation, low-VRAM deployment, and image editing. Limitations include only Turbo being open, and the need for optimization in extreme stylized generation and model compilation time. Z-Image balances performance, efficiency, and practicality, making it suitable for small and medium-sized teams and developers to reduce deployment barriers.
Read More