Easy Apply PowerPaint

Worth saying upfront, since it changes how you should read this node: PowerPaint had its moment in April 2024 and the conversation around it has mostly gone quiet since - checked against the corpus, mentions cluster almost entirely in that one month and trail off to near-nothing by 2025. That doesn't mean it's broken or bad - it's a genuinely capable multi-task inpainting model - it means most people today reach for Fooocus-style inpaint, BrushNet, or a modern edit model like Qwen-Image-Edit instead, and this node exists mainly for the workflows and workflows-of-workflows still built around it.

What it actually does

PowerPaint's trick is that it's one checkpoint trained for five different inpainting jobs, selected by which "task" prompt you steer it toward rather than by loading five different models: text-guided edits (change the masked region based on a text prompt), shape-guided edits (the mask shape itself is the guidance), object removal, context-aware fill (blend with surroundings, minimal new content), and image outpainting. That versatility is genuinely rare - most inpainting models pick one job and do it well; PowerPaint tries to do several from one set of weights.

Where this node differs from easy applyInpaint's own powerpaint mode: this one is the dedicated, standalone version, where you explicitly pick your PowerPaint model and CLIP files rather than having them resolved automatically as one option among several inpaint engines.

The inputs that matter

pipe, image, and mask are the required trio, same as any inpaint node. powerpaint_model and powerpaint_clip are dropdowns for the specific PowerPaint weights and its paired CLIP - both need to actually be downloaded and placed in the right folders, since neither auto-fetches. dtype (float16/bfloat16/float32/float64) trades precision for VRAM and speed - float16 is the sane default on almost any modern GPU.

function is the task selector described above - this is the field that actually decides what PowerPaint does with your mask, so get this one right before touching anything else. fitting and scale tune how strongly the task embedding steers generation; start_at/end_at work like a ControlNet's step range, letting PowerPaint's influence taper off before sampling finishes rather than holding through the whole run. save_memory (none/auto/max) trades speed for VRAM headroom - worth bumping up if you're running PowerPaint alongside a large base model and hitting OOM.

Output is a single pipe.

Installing it

ComfyUI Manager: search ComfyUI Easy Use, install, restart. By hand:

cd ComfyUI/custom_nodes
git clone https://github.com/yolain/ComfyUI-Easy-Use

Windows: double-click install.bat. Linux/Mac: run pip install -r requirements.txt yourself. Restart ComfyUI. The PowerPaint model and CLIP weights are separate downloads you'll need to place manually - this node applies them, it doesn't fetch them.

Where people get stuck

The single biggest issue is picking the wrong function for the job - text-guided and shape-guided in particular produce very different results from the same mask, and if your output looks nothing like what you expected, that's the first field to check. Missing weights is the second most common problem: if powerpaint_model or powerpaint_clip show as empty dropdowns, the files simply aren't in your models folder yet. And given how thin current community discussion of PowerPaint actually is, don't expect to find much fresh troubleshooting for it if you hit an edge case - if you're chasing a specific artifact and it isn't obviously a function, fitting, or scale problem, it may genuinely be faster to switch to BrushNet or a modern edit-model-based inpaint approach than to dig for PowerPaint-specific answers that mostly aren't being written anymore.