dsh-comfyui

Local ComfyUI text-to-image for DeepSeek Harness.

• Generation panel (sidebar entry + main panel): prompt box, Krea-tuned prompt optimizer, size/steps/guidance/seed, live progress, last result, and an on-disk gallery.
• Auto-start: probes 127.0.0.1:8188 and starts ~/tools/ComfyUI/run.sh when nothing answers.
• Chat tool image_generate: the agent generates images into the conversation and copies them into the workspace generated/ folder.
• Model discovery: reads ComfyUI's /object_info, so GGUF (UnetLoaderGGUF + CLIPLoaderGGUF) and regular UNETLoader/CLIPLoader installs both work. The bundled preset is Krea 2 Turbo (GGUF Q4 + Qwen3-VL text encoder + Qwen-Image VAE).

Environment

The plugin drives a local ComfyUI; nothing is bundled. A full checklist — ComfyUI install, the three custom nodes, the Krea 2 model files, the LoRAs each feature needs, verification commands and known traps (FP8 on Apple Silicon, missing OpenCV, slow Hugging Face downloads) — lives in SETUP.md. Short version:

./scripts/download-models.sh   # 7.0 GB GGUF + 8.3 GB text encoder + VAE
./scripts/download-loras.sh    # optional style LoRAs

Install into a DSH profile

cd ~/.dsh/profiles/<profile>
# 1. install the package (verified form: pnpm add "<name>@file:<absolute path>")
pnpm add "dsh-comfyui@file:$HOME/project/dsh-comfyui"
# 2. append the row below to cordis.patch.yml
# 3. reload it (touch cordis.patch.yml, or restart the app), then refresh the page

The row in cordis.patch.yml:

- insert:
    - id: comfyui
      name: dsh-comfyui

First start may take a while: the plugin probes 127.0.0.1:8188, starts <comfyRoot>/run.sh when nothing answers, then loads ~15 GB of weights on the first job.

Panel quick tour

Sidebar → ComfyUI 生图. Write an idea, optionally press 优化提示词 (rules or model rewrite), pick a preset and size, then 生成. While it renders you get step progress, elapsed time and a 停止 button; the panel keeps its draft and in-flight job if you switch pages. Finished images land in the gallery below (hover a thumbnail → 删除 removes the file, second click confirms). Chat generations additionally copy into <workspace>/generated/.

Pixel game assets

Pick the 像素游戏资产 preset in the panel (or pass style: "pixel" / pixel_asset: true) to produce true pixel-art sprites:

1. a Krea 2 pixel LoRA (e-n-v-y/Krea-2-Pixel-Art, 32/64/128 variants) renders a 1024² sheet,
2. Krea2PixelArtRefiner collapses it onto the requested logical grid and reduces the palette (needs envy-ai/ComfyUI-Krea2-Pixel-Art-Refiner plus opencv-python-headless),
3. the gallery asset really is 32/64/128/256 px — not a downscaled illustration. The raw 1024² render is never saved anywhere,
4. a second refiner expands that same collapsed grid back to 1024² with hard blocks (scale_to_original: true). This block PNG is the preview: the gallery and the chat show it (larger than 64 px, still crisp pixel art), while the small sprite stays the downloadable asset. It lands next to the sprite as <name>.preview.png and gallery.json links it as preview.

The host also refuses to store a "pixel" job whose image is larger than the requested sprite size, so a graph change can never quietly put a smooth 1024² render in the pixel slot.

Panel options: sprite size (32/64/128/256) and palette size (2–256, default 24). Tool: image_generate({ prompt: "a wooden treasure chest", pixel_asset: true, asset_size: 64, palette_colors: 24 }).

Reference images (image + text)

krea2-turbo can take a reference image, wired through the Krea 2 edit path:

mode	LoRA	use
identity	krea2_identity_edit_v1_1_r64.safetensors	keep a character / persona consistent
style	krea2_style_reference.safetensors	copy the look of a picture

Install requirements once:

cd ~/tools/ComfyUI
git clone https://github.com/ostris/ComfyUI-Krea2-Ostris-Edit.git custom_nodes/ComfyUI-Krea2-Ostris-Edit
# LoRA files land in models/loras/

Panel: upload a reference in the 参考图 card, pick 人设 / 风格, then generate. Chat tool: image_generate({ prompt, reference: "/path/to/portrait.png", reference_mode: "identity" }).

Output size defaults to 跟随参考图 (Kontext scale to ~1MP); pick a concrete size in the size dropdown to override it. Each generation costs more: the edit LoRA plus vision encoding make a step noticeably slower than plain text-to-image on Apple Silicon (~50-60 s per step here).

Presets

The panel offers exactly two presets, shown side by side in one 预设 card:

preset	id	what it does
画师风插画	illustration	finished digital illustration in a distinctive artist style — linework, brushwork, deliberate palette
像素游戏资产	pixel	true sprites: rendered at 1024² through the pixel LoRA, then collapsed to 32–256 px with a small palette

The list is a single registry (STYLES in lib/prompt.js): the panel tiles, the image_generate style enum and the rule-mode facets all read it, so adding a preset is one entry there (plus, if you want a custom mark, one branch in PresetGlyph in client.js). Tile captions come from the blurb field, localised through { zh, en }.

画师风格（illustration 的子选项）

Picking 画师风插画 reveals a chip row under the two tiles: 通用 plus 14 artist references (新海诚、吉卜力、京阿尼、米山舞、望月けい、村田莲尔、天野喜孝、安倍吉俊、CLAMP、Ilya Kuvshinov、WLOP、 loundraw、鸟山明、荒木飞吕彦). An artist is one record in ARTISTS (same file): its fragment replaces the generic illustration style sentence, short is the chip label, zh/en the tooltip. 通用 means no artist. The pixel preset ignores the artist row entirely.

These are style hommages, not attribution: some artists object to AI imitation, so treat them as references for personal work and never present the output as theirs.

Chat tool: image_generate({ prompt, style: "illustration", artist: "shinkai" }), or artist: "ghibli" and so on; omit artist for a generic illustration.

画风 LoRA（二次元向）

Under the look row there is a 画风 LoRA row listing the official Krea 2 style LoRAs that are actually installed in models/loras. Picking one patches the UNET (LoraLoaderModelOnly) and appends that LoRA's official trigger word to the prompt — the same recipe as the ComfyUI Krea 2 template.

chip	file	trigger word	look	strength
赛璐璐日间动漫	cel-shaded-daytime-anime-comfy	cel shaded daytime anime style	现代 TV 动画赛璐璐（社区，fal 系列）	1.0 (1.0-1.25)
复古动漫	krea2_retroanime	purple retro anime style	90s cel anime, purple palette	1.0
艺术水彩	krea2_softwatercolor	art deco watercolor style	art-deco watercolour	1.0
水墨笔触	krea2_darkbrush	monochrome ink wash style	sumi ink wash	1.0
点阵网点	krea2_dotmatrix	monochrome stippling style	halftone stippling	1.0
霓彩流挂	krea2_neondrip	textured abstract style	neon drips	1.0
雨窗	krea2_rainywindow	rainy window style	rain on glass	1.0
暮色残影	krea2_sunsetblur	ethereal motion blur style	long-exposure blur	1.0
复古塔罗	krea2_vintagetarot	vintage tarot style	tarot illustration	1.0
童趣速写	krea2_kidsdrawing	naive expressive sketch style	crayon sketch	1.0

Trigger words and the 1.0 strength come from the official Comfy-Org/Krea-2 card; 赛璐璐日间动漫 is a community LoRA (ilkerzgi/krea-2-cel-shaded-daytime-anime-lora, trained on Krea 2 Turbo, Krea 2 Community License — commercial use needs your own check). Only installed files show up as chips (/object_info decides), so the row stays honest.

./scripts/download-loras.sh            # 二次元四件套: celanime + retroanime + softwatercolor + darkbrush
./scripts/download-loras.sh all        # all ten (5.3 GB); the chips row then shows every one

Trigger words and the 1.0 strength below are checked against the official Comfy-Org/Krea-2 model card; only files that exist in models/loras become chips, so an un-downloaded row simply stays hidden.

Do not put krea2_turbo_lora_rank_64_bf16.safetensors on the Turbo checkpoint: it is the Raw→Turbo distillation delta, meant for krea2_raw (it rewrites the same capability that Turbo already has). It is deliberately absent from the panel's LoRA list.

Strength is editable next to the row (0-1.5). Official guidance is 1.0; drop to ~0.6-0.8 when a named artist prompt should dominate over the LoRA's own look.

Chat tool: image_generate({ prompt, style: "illustration", artist: "anime", lora: "retroanime" }).

Sampling defaults follow the installed model, not the plugin config: Krea 2 Turbo is 8 steps at CFG 1 (euler / simple, straight from the official ComfyUI Krea 2 template), and the panel locks the guidance field because CFG 1 means it cannot have any effect. LoRAs do not need extra steps — the official LoRA cards say their previews were rendered on Turbo at 8 steps, weight 1.0.

二次元尺寸与提示词

• Sizes are the official Krea 2 1K tiers: 1024², 4:5 928×1152, 2:3 832×1248, 9:16 768×1376, 4:3 1184×896, 3:2 1248×832, 16:9 1376×768, 2.35:1 1568×672, plus 512/768 drafts and a 2K tier (1440×1440, 2048×1152). Anime portraits work best from 832×1248 / 928×1152.
• The model is trained for 1K-2K and the official guidance is to render natively at the target size — there is no hires-fix stage here. On this Mac a 1K render is minutes and a 2K render is roughly ten times that, so 2K is a "final render" switch, not a default.
• Write the medium explicitly, because Krea 2 leans photographic otherwise: anime illustration, crisp cel-shading, flat shading, 1990s cel animation, detailed digital painting.
• No negative prompt: at CFG 1 the negative branch is skipped. (word:1.2) weighting also does not apply (the Qwen3-VL encoder reads whole sentences); move the important phrase earlier instead.
• LoRA trigger words are appended at the end of the prompt, which is exactly where the official and community cards say to put them.

Prompt modes (rules / llm / auto / off) are unchanged; rules assembles subject + medium + lighting + perspective + texture + orientation, llm rewrites the idea as one English paragraph aimed at the selected preset (no photographic-realism cues for pixel).

mode	what happens
rules	deterministic facet assembly, no model call
llm	one completion rewrites the idea as one English paragraph for the preset
auto	llm when the configured model is available, otherwise rules
off	the prompt is sent as written

Configuration

- insert:
    - id: comfyui
      name: dsh-comfyui
      config:
        comfyRoot: ~/tools/ComfyUI
        comfyUrl: http://127.0.0.1:8188
        autoStart: true
        provider: deepseek-account  # LLM used ONLY for prompt rewriting
        model: deepseek-flash       # that LLM's model id (not the image model)
        optimize: auto              # auto | llm | rules | off
        defaultModel: krea2-turbo   # image preset; auto-detected, rarely needs setting
        steps: 8                    # fallback only: the model preset wins
        guidance: 0                 # fallback only: locked at CFG 1 for Krea 2
        width: 1024                 # fallback when the request has no valid size
        height: 1024
        saveToWorkspace: true
        workspaceSubdir: generated
        timeoutMs: 900000           # per-image deadline

provider / model configure the prompt-rewrite LLM; defaultModel picks the image preset. Model files are auto-detected from /object_info; override with unetName, clipName, clipLName, vaeName if needed.

Notes

• Krea 2 Turbo renders 8 steps at CFG 1 (euler / simple). Measured here, 24 GB M5: ~20 s per step for plain text-to-image (~3 min per 1024² image), ~50-60 s per step with a reference image, ~3 min for a pixel sprite.
• Krea 2 weights are under the Krea 2 Community License (commercial conditions and a revenue threshold); third-party LoRAs carry their own licenses.
• The gallery lives in <comfyRoot>/dsh-gallery/ (plus gallery.json and jobs.json); chat generations are also copied to <workspace>/generated/.

Verification

node scripts/smoke-panel.mjs renders the panel headlessly against a fake host (tiny React stand-in) and checks the preset card: two tiles in one card, no leftover look dropdown, sprite options only for the pixel preset, reference card hidden for sprites. LEGACY_HOST=1 runs it against a pre-0.5 status payload to exercise the fallback path.

License

MIT