工具｜PSD 模型预下载器

Here's the thing nobody tells you about the Tikpan Smart PSD Layering node: the first time you run it, it doesn't just generate a PSD - it installs dependencies and downloads models first. On a slow connection that's a fifteen-minute pause right in the middle of your workflow, with the progress bar sitting at zero and no way to tell what's happening. The PSD Dependency Downloader exists to make that stall happen on your schedule instead of its.

Drop this node on the canvas, run it once ahead of time, and the heavy lifting is done before you ever touch the layering node. It needs no API key, no Tikpan account, nothing cloud-based - this is the one genuinely local corner of an otherwise API-heavy pack. You're pulling open-source segmentation models and pip packages onto your own disk.

How it works

Two phases, both visible in the node's own logs. First it pip-installs the Python packages your chosen tier needs:

• 经济档 (300MB) - pytoshop, opencv-python, rembg
• 标准档 (500MB) - everything above plus git+https://github.com/facebookresearch/sam2.git and easyocr
• 极致档 (700MB) - the above plus simple-lama-inpainting (the LaMa infill model)

Then it downloads the actual AI model weights - the BiRefNet segmentation models and friends. The 包含补全模型 toggle adds the ~200MB inpainting model on top of whatever tier you picked. The outputs are a 下载日志 (STRING, the whole install+download transcript) and a 状态预览 (IMAGE that shows pass/fail visually). Both are diagnostics; you're not wiring them anywhere.

The model stack, briefly

This pack's PSD pipeline is built on well-known open-source parts. BiRefNet is the current default-recommendation for background removal - sharper edges than the old u2net, especially on hair and semi-transparent material (background-removal.md). SAM2 does the multi-element segmentation, GroundingDINO does text-prompted detection ("the price tag", "the logo") and LaMa fills in what's occluded. The README's honest bit: every model has a fallback chain - BiRefNet→rembg, PaddleOCR→EasyOCR, GroundingDINO→SAM2 dense segmentation - so a failed download degrades the output rather than killing the run.

Installing

Same pack install as everything else: Manager search "Tikpan", or git clone https://github.com/htrert/ComfyUI-Tikpan-Pro into custom_nodes, restart. Then just add this node and run it once. You don't need to install this and let the layering node auto-install - running the downloader once is enough; the layering node detects what's present.

Common issues

• SAM2 download stalls. The README and usage guide note that SAM2 weights come from Facebook's CDN, which can be slow or blocked in some regions. Use a proxy, or manually grab sam2.1_hiera_small.pt from the SAM2 releases and drop it in ComfyUI/models/sam2/.
• "缺少依赖" errors in the layering node later. That means the auto-install was disabled or failed silently. Running this downloader with the matching tier fixes it.
• Disk space. 300–700MB plus the optional 200MB inpaint model adds up, and SAM2 + BiRefNet + rembg all live in your ComfyUI install. Check you're not near-full before kicking off the 极致档.