Comfyui Flux

Generate images locally using ComfyUI + Flux.1 Dev model. Zero API cost, strong text rendering. Use when: generating images locally, creating slides/PPT grap...

Security Scan

The skill's code and instructions match its stated purpose (local image generation via a local ComfyUI instance); no unrelated credentials, external downloads, or hidden endpoints were found — but the documentation is inconsistent about ports and the scripts will upload any file you point them at to whatever COMFYUI URL is configured, so verify COMFYUI is truly local before using.

Purpose & Capability

Name/description, SKILL.md and the two included Python scripts consistently implement local image generation through ComfyUI (Flux.1 Dev and PuLID). No unrelated services or credentials are requested. Note: SKILL.md and docstrings disagree on default ports (8188 vs 8200) and the prerequisites include a user-specific path (/Users/chrisliu/ComfyUI), which is sloppy and may confuse users.

Instruction Scope

Instructions and scripts operate within the expected scope: they talk to ComfyUI endpoints (/system_stats, /prompt, /history, /view, /upload/image), save generated images to a local folder, and upload user-supplied reference images for PuLID. They do not read arbitrary system state or other credentials. Important caveat: the scripts use the COMFYUI_URL environment variable — if that is set to a remote host, reference images and prompts will be sent to that host, enabling data exfiltration. SKILL.md documents COMFYUI_URL override, so this behavior is explicit but should be treated carefully.

Install Mechanism

There is no install spec and no external downloads; the skill is instruction-plus-scripts only, which is the lowest-risk install model. The included Python scripts use only stdlib modules and make HTTP requests to the configured ComfyUI endpoint.

Credentials

The skill declares no required environment variables; in code it reads COMFYUI_URL as an optional override (documented). This is proportionate. However, because COMFYUI_URL controls the destination for uploads and prompt submission, setting it to a non-local host would route image data (including any reference image files you provide) off your machine. The scripts can read and POST any file path you pass to --ref-image, so avoid pointing to sensitive files.

Persistence & Privilege

The skill does not request persistent or elevated platform privileges (always:false). It does not modify other skills or system-wide agent settings. It simply runs scripts that interact with ComfyUI when invoked.

Assessment

This skill appears to do what it claims: talk to a ComfyUI instance on your machine and save generated images. Before installing or running it: - Verify which port your local ComfyUI uses (SKILL.md mentions 8200 in one place and 8188 in another) and set COMFYUI_URL explicitly if needed. The scripts default to 127.0.0.1:8200 in code but some messages/docstrings reference 8188, which will confuse runs if your ComfyUI is on a different port. - Ensure COMFYUI_URL is not pointed at a remote host unless you intentionally want to send images and reference files there — if pointed remotely, the skill will upload files (including any --ref-image path) to that host, which could leak sensitive data. - Do not pass paths to sensitive files as --ref-image; the script will read and upload whatever file you supply. - The SKILL.md contains a user-specific startup path (/Users/chrisliu/ComfyUI) — that is just an example and should be replaced with your actual ComfyUI start command. - Be aware the PuLID functionality is explicitly uncensored (can generate any content); follow your local policies and laws when generating images. If you want higher assurance, inspect and run the included scripts locally before giving the skill runtime access, and run ComfyUI in a restricted environment or container.

ComfyUI Flux.1 Dev — Local Image Generation

Generate images locally via ComfyUI API + Flux.1 Dev. No API cost, excellent text rendering.

Prerequisites

ComfyUI must be running (default port 8200):

# 启动
bash /Users/chrisliu/ComfyUI/start.sh
# 关闭
bash /Users/chrisliu/ComfyUI/stop.sh

Override URL via env: COMFYUI_URL=http://127.0.0.1:8200

Models (already installed):

• flux1-dev.safetensors (diffusion model)
• t5xxl_fp8_e4m3fn.safetensors + clip_l.safetensors (text encoders)
• ae.safetensors (VAE)

Usage

python3 {baseDir}/scripts/flux_generate.py --prompt "PROMPT" [OPTIONS]

Basic

python3 {baseDir}/scripts/flux_generate.py \
  --prompt "A professional presentation slide with title 'AI Revolution' in bold white text on dark gradient background"

Custom Size (e.g. 16:9 for PPT slides)

python3 {baseDir}/scripts/flux_generate.py \
  --prompt "..." \
  --width 1344 --height 768

Parameters

Parameter	Default	Description
--prompt	(required)	Image description
--negative	""	Negative prompt
--width	1024	Image width
--height	1024	Image height
--steps	20	Sampling steps (more = better but slower)
--cfg	1.0	CFG scale (Flux works best at 1.0)
--seed	random	Reproducibility seed
--output	auto	Output path (default: generated-images/flux-{ts}.png)
--timeout	600	Max wait seconds

Common Sizes

Use Case	Size	Flag
Square	1024×1024	(default)
PPT 16:9	1344×768	--width 1344 --height 768
Portrait	768×1344	--width 768 --height 1344
Landscape	1344×768	--width 1344 --height 768

PuLID Face-Consistent Generation

Generate images with a reference face using PuLID-Flux (zero API cost, zero censorship):

python3 {baseDir}/scripts/pulid_generate.py \
  --prompt "PROMPT" \
  --ref-image path/to/face.jpg [OPTIONS]

Basic (face-consistent portrait)

python3 {baseDir}/scripts/pulid_generate.py \
  --prompt "a girl sitting in a cafe, warm afternoon light, casual outfit" \
  --ref-image avatars/kimi.jpg

Custom size + weight

python3 {baseDir}/scripts/pulid_generate.py \
  --prompt "a girl on the beach at sunset, wearing a sundress" \
  --ref-image avatars/kimi.jpg \
  --width 1024 --height 1024 \
  --weight 1.2

PuLID Parameters

Parameter	Default	Description
--prompt	(required)	Image description
--ref-image	(required)	Reference face image path
--negative	""	Negative prompt
--width	768	Image width
--height	1024	Image height (portrait default)
--steps	10	Sampling steps (PuLID works well with fewer)
--cfg	3.5	Guidance scale
--seed	random	Reproducibility seed
--weight	1.0	Face similarity weight (0.0-5.0, higher = more similar)
--start-at	0.0	PuLID start timestep
--end-at	1.0	PuLID end timestep
--output	auto	Output path (default: generated-images/pulid-{ts}.png)
--timeout	600	Max wait seconds

PuLID Notes

• Default portrait orientation (768×1024) — good for face shots
• --weight 1.0 is balanced; increase to 1.2-1.5 for stronger likeness
• --steps 10 is enough for PuLID; increase to 20 for more detail
• First run is slower (loads PuLID + InsightFace + EVA-CLIP models)
• No content censorship — can generate anything locally

Strengths vs API (qwen-image)

• ✅ Free — no API cost
• ✅ Excellent text/typography rendering (great for slides!)
• ✅ High detail and consistency
• ⚠️ Slower on Mac (~1-3 min per image)
• ❌ No native face-consistency (use PuLID/IPAdapter for that)

Output

Prints MEDIA:<path> — auto-attached to chat reply. Default saves to generated-images/.

Troubleshooting

If ComfyUI is not running, the script exits with an error and prints startup instructions.