MIT-licensed DSH plugin that puts a text-to-image workbench and `imagen_*` chat tools in front of free remote GPUs (Kaggle 2×T4 by default, Colab for the larger models), with seven models, version-tree rework, local-only output and idle auto-shutdown.
DSH integration
Native runtime
Author-claimed
Safety audit
Unaudited
Last verified
2026-11-21
License
MIT
01What can it help you accomplish?
Generate an image by describing it in chat, without hand-tuning prompts or ratios
The built-in assistant turns the request into an English prompt, infers the aspect ratio, checks the current model’s parameter limits and shows a confirmation card; the image is produced only after you confirm
DSH users who want pictures from a conversation but still want a checkpoint before compute is spent
Keep iterating on a picture in a dedicated workbench instead of starting over
Select a result and rework it in place — local inpainting, full repaint, Real-ESRGAN ×2 upscale or outpaint — with each product archived as a sub-version in the version tree
Creators iterating on one image (posters, character art, product shots) and needing version history
02How to install into DeepSeek Harness
Prerequisites
- A Kaggle account with phone verification completed — unverified accounts are downgraded to CPU-only, per the README
- Kaggle credentials written to `%USERPROFILE%\.kaggle\credentials.json` (browser OAuth, kaggle CLI 2.x, or a manual API token)
- `dsh` on PATH, or the DSH GUI plugin manager; replace `web` in the commands with your profile name (usually `desktop` on the DSH Desktop client)
- No local GPU needed — model loading and inference run on the remote kernel
Installation steps
- 01
Prepare Kaggle credentials: workbench → GPU status → Run settings → Advanced → Kaggle access → direct, then "在浏览器中登录 Kaggle"; or `pip install -U kaggle` + `kaggle auth login`; or paste an API token into `~/.kaggle/access_token` (chmod 600).
- 02
Install the plugin: `dsh plugin --profile web add dsh-our-imagen` (swap `web` for your profile), or use the GUI plugin manager with `dsh-our-imagen` / `github:nanxfu/dsh-our-imagen`, then restart DSH.
$ dsh plugin --profile web add dsh-our-imagen
- 03
To upgrade, remove the plugin first (`dsh plugin --profile web remove dsh-our-imagen`), reinstall and restart DSH.
$ dsh plugin --profile web remove dsh-our-imagen
- 04
Open the workbench, press power on / wake the cloud; the first Kaggle boot takes roughly 3–10 minutes to allocate resources and load weights.
- 05
Once the status reads `Online`, start generating — steady-state renders take seconds per image.
Verify the integration
Not specified by the author
03DSH integration and capability boundaries
Native DSH plugin (`dsh-our-imagen`) that adds a full-page image-generation workbench, `imagen_*` model tools and a local Host process that drives remote free GPU kernels (Kaggle by default, Colab optional)
Seven built-in models over two compute sources
Model choice, prompt, aspect ratio, resolution, steps, guidance and batch count→Renders from SDXL Lightning, SANA-Sprint 1.6B, FLUX.2 Klein 4B and Anima Turbo on Kaggle or Colab, plus Z-Image Turbo, Qwen-Image 2.1 and FLUX.2 Klein 9B on Colab only
consumes the free weekly GPU quota of the Kaggle account or the Colab plan in useLocal-only image storage with idle shutdown
A generated image→The file is written straight into the local workspace `imagen-output/` directory through a single-use random Bearer Token tunnel
nothing goes through a third-party public image host; the cloud kernel shuts down after 30 idle minutes (configurable) or on manual power-offChat tools with confirmation cards
A natural-language request in the chat window→`imagen_generate`, `imagen_batch`, `imagen_recipes`, `imagen_session` and `imagen_history` tools; the confirmation card can carry its parameters into the workbench panel
the recipe library only offers visible prompt templates — no hidden prompt assemblyReconnect without re-booting
A DSH restart or page refresh while a kernel is running→The local `.imagen/session.json` state is probed and the existing instance is taken over
avoids starting a second cloud kernel and burning quota twice
04Who is it for? When not to use it?
Good for
- DSH users who want pictures from a conversation but still want a checkpoint before compute is spent
- Creators iterating on one image (posters, character art, product shots) and needing version history
Not for
- Kaggle only grants free GPUs to accounts that completed phone verification; an unverified account is downgraded to CPU-only, which makes the plugin unusable for generation.
- Z-Image Turbo, Qwen-Image 2.1 and FLUX.2 Klein 9B need Colab and will not appear in the model cards while the compute source is Kaggle.
- The project is in Preview: the Kaggle driver is built in and the Colab driver is wired up, while local GPU and self-hosted Worker backends are still planned.
05Compatibility, maintenance and safety notes
- Kaggle only grants free GPUs to accounts that completed phone verification; an unverified account is downgraded to CPU-only, which makes the plugin unusable for generation.
- Z-Image Turbo, Qwen-Image 2.1 and FLUX.2 Klein 9B need Colab and will not appear in the model cards while the compute source is Kaggle.
- The project is in Preview: the Kaggle driver is built in and the Colab driver is wired up, while local GPU and self-hosted Worker backends are still planned.
MIT · actively maintained (latest release v0.14.0, 2026-11-19); Preview status, Kaggle driver default, Colab driver available
06Frequently asked questions
Do I need a discrete GPU on my own machine?
No. Model loading and inference run on the remote kernel (Kaggle or Colab); locally the plugin only draws the UI and saves the images.
Why does Kaggle give me CPU instead of a GPU?
Kaggle only allocates free GPUs to accounts that finished phone verification in the account settings. Without it the account is downgraded to CPU-only, and the plugin cannot generate.
Are my images uploaded to a public image host?
No. The remote result comes back through a single-use random Bearer Token tunnel to the local Host and lands in the local workspace `imagen-output/` directory.
Why is the first boot so slow, and can I stop it?
A cold boot allocates the instance, checks the runtime and downloads model weights — typically 3–10 minutes. After that renders take seconds. The kernel shuts down automatically after 30 idle minutes (configurable) or immediately from the power button.
Why can’t I pick Z-Image Turbo, Qwen-Image 2.1 or FLUX.2 Klein 9B?
Those three need more VRAM and are Colab-only. Switch the compute source to Colab (Run settings → compute source → Colab Pro, then authorize in the browser) and they appear in the model cards.
07Related DSH workflows
dsh-image-gen
by shanliuling
Adds in-chat multimodal image generation via a `generate_image` tool with multi-provider support (Gemini/OpenAI/Seedream), interactive preview, copy/download, and session-attached results.
dsh-vision-router
by ysr666
Eyes for text-only DeepSeek Harness agents: built-in free vision chain (no key) + pixel-level vision tools (Q&A, grounding, crop, pixel diff, colors, OCR, SVG trace, cutout, screenshots). One-command install, no Python, image turns work like ordinary tool-calling turns.
dsh-vision-toolkit
by anionex
为纯文本 DSH Agent 提供 10 个结构化视觉工具:意图感知图片问答、长截图 OCR、原始像素 grounding、UI 还原、像素 diff 等
dsh-see-now
by saos-3624
Captures screen, OBS, or webcam frames, filters by visual change, packs them into image sheets, and delivers them to DSH sessions.
08Data and sources
**DeepSeek Harness 的开源文生图工作台与视觉 Agent 插件。**
宿主自动挂载 `imagen_generate`、`imagen_batch`、`imagen_recipes`、`imagen_session`、`imagen_history` 等工具
This page is generated from the project’s public documentation, repository metadata and a structured parse of DSH Plugins; last verified on 2026-11-21. Found an error? Submit a correction.
Best DeepSeek Harness Plugins
Twelve plugins worth installing first — picked from the whole catalog, across every category.
