dsh-comfyui-image
从 DeepSeek Harness 驱动本地 ComfyUI 图像工作流（Z-Image-Turbo、Qwen-Image-2.1）
插件会安装到这里；不确定时保持 web。
npx -y @deepseek-ai/dsh plugin --profile web add github:QXingYShu/dsh-comfyui-image#29f385bce273ceedbbc2b92ec12b967c0122a07dComfyui Image 以 dsh-comfyui-image 发布，当前版本为 0.1.0。Plugin Hub 会校验它的 manifest，并保存精确安装来源，便于复现安装结果。
Generate image assets from DeepSeek Harness using a local ComfyUI install — no API keys, no cloud, no per-image cost. The agent picks a model, writes the prompt, queues the workflow on your own GPU, and gets back a real image file.
Two workflows ship wired up:
| Workflow | Good at | Speed | Negative prompt |
|---|---|---|---|
| Z-Image-Turbo | drafts, batches, fast iteration | seconds | no (CFG 1, prompt only) |
| Qwen-Image-2.1 | final assets, legible in-image text, precise composition | minutes | yes |
comfyui_generate — generate one or more images and save them to disk.comfyui_status — report which workflows are ready on this machine and
what defaults they use. Useful when generation fails.comfyui-image — model selection and prompt craft, so the agent
chooses deliberately instead of always reaching for the same model.You do not need ComfyUI to be running. The plugin starts it headless on first use and reuses it afterwards.
dsh plugin --profile desktop add https://github.com/<you>/dsh-comfyui-image.git
Or from a clone:
git clone https://github.com/<you>/dsh-comfyui-image.git
dsh plugin --profile desktop add ./dsh-comfyui-image
Restart the Harness and ask for an image.
A restart is required, not optional: the profile applies its plugins once when
it boots, so a change to lib/ — or to this plugin after a git pull — is not
picked up by the running process. Only the Web GUI's client plugins hot
reload, and only while pnpm run dev:web is rebuilding them; this plugin's
tools and skill are server-side.
The plugin prefers an already-running server, and otherwise launches one itself:
8188–8192, for a live /system_stats.extra_model_paths.yaml (which is what makes Comfy Desktop's shared model
directory visible).All optional. Set environment variables, or just rely on the defaults.
| Variable | Effect |
|---|---|
DSH_COMFY_PORT | Force one port for every workflow (default: probe 8188–8192) |
DSH_COMFY_PORT_Z_IMAGE_TURBO | Force a port for Z-Image-Turbo only |
DSH_COMFY_PORT_QWEN_IMAGE_2_1 | Force a port for Qwen-Image-2.1 only |
DSH_COMFYUI_PATH | Point at a manual ComfyUI checkout instead of Comfy Desktop |
DSH_ZIMAGE_MODELS | Override Z-Image-Turbo model file names (JSON) |
DSH_QWEN_IMAGE_MODELS | Override Qwen-Image-2.1 model file names (JSON) |
Override the model names when your weights differ from the defaults, e.g. if you quantised Qwen-Image-2.1 yourself:
export DSH_QWEN_IMAGE_MODELS='{"unet_name":"qwen_image_2.1_fp8.safetensors","clip_name":"qwen3vl_8b_fp8.safetensors","vae_name":"qwen_image_2.1_vae_bf16.safetensors"}'
draw a small green plant in a terracotta pot, soft window light, photograph
The agent loads the comfyui-image skill, picks Z-Image-Turbo for a draft, and
saves the result under ./generated-images/.
Ask for more control when you want it:
use qwen-image-2.1 to make a 16:9 poster that says "DEEP DIVE" in bold letters above a stylised submarine, screen-print texture, negative prompt: blurry text
Both graphs live in workflows/ as ComfyUI API-format JSON with
{{placeholder}} strings. They were derived from the official Comfy template
workflows and flattened out of their subgraphs, so they queue over plain HTTP
without the frontend.
To add or tune a workflow:
{{snake_case}} names.workflows/ and add a spec in lib/workflows.js.A parameter that is missing or misspelled fails loudly before anything reaches the GPU — you get an error naming it, not a silently wrong image.
npm test # hermetic: schema shape, parameter resolution, apply() on a fake host
npm run test:host # the installed Harness's own skill validator, extracted from app.asar
npm run test:smoke # discovery + a real generation of both workflows through the tool
node scripts/boot-check.mjs <node_modules> <plugin-dir> # apply() under the real Cordis loader
npm test is hermetic and needs neither ComfyUI nor the Harness. The other
three need one of the two, and boot-check additionally needs a checkout that
supplies @deepseek-ai/cordis.
define-tool.mjs asserts the shape of the compiled tool schema, not just the
behaviour: JSON Schema defines required as an array of property names owned by
the containing object, so a required: true left on a property node makes the
schema invalid, and a gateway that validates the tool list it is handed rejects
the whole list — every tool, not just the broken one.
test:host is the interesting one. ctx.skills.register() only validates
name, description and invocation, so a skill missing its source field
registers cleanly, shows up in the catalog, and then fails every time the model
tries to load it. That second validation lives in @deepseek-ai/dsh-skill; this
script lifts the real function out of the installed app.asar and runs the
plugin's skill through it, with a negative control proving the check can still
fail. npm test carries a transcription of the same rules so a regression is
caught without an installed Harness.
test:smoke generates with both workflows using the minimal argument shape a
model actually produces — prompt only, every optional argument omitted — because
that is where a default-erasing bug shows up.
node_modules. Using only
ctx.tools and ctx.skills keeps the plugin installable from anywhere.
lib/define-tool.js is a small local defineTool that compiles the parameter
DSL to JSON Schema and validates arguments before dispatch.MIT — see LICENSE.