ComfyUI-WepeNerd
Custom node pack featuring an interactive visual resolution picker and resolution suggestion tool with real-time dimension readout and divisor snapping. (Description by CC)
Nodes (24)
Your Qwen edit shifted the whole frame by three pixels. This node puts them back.
Keep only what Qwen actually changed — and keep the rest of the frame untouched
How far did Qwen actually move your image? This node gives you the number
The LoRA that only lives where you paint it
Stop typing resolutions — drag a box to set your output size
Your 111-Frame Video Came Back at 109 Frames. That's the Grid.
The two numbers that fix a clip that's almost right
Push-warp an image right inside the node
Wan wants 4n+1 frames — this loader hands them over without VHS
You shouldn't add this node — here's why it exists
The brush ComfyUI keeps making you leave the graph to find
Qwen Edit Has No Mask Input — So Draw One and Show It the Region
One dial, no cropping, no mental arithmetic
A resolution calculator that finally kills the 'not divisible by 64' errors
One file, sound inside, no FFmpeg scavenger hunt
Stop accepting whatever the scheduler dropdown gives you
The semantic LoRA slider
Sketch it inside the graph instead of opening Krita
Zoom Crop Doesn't Zoom, and That's Exactly Why It's Useful
ComfyUI-WepeNerd
Resolution tools, video loading and saving, painting and image warping, spatial LoRA masks and H3 Ripple video edit propagation for ComfyUI.
| Tool | What it does |
|---|---|
| Drag Resolution | Set dimensions visually with aspect-ratio and divisor controls |
| Resolution Suggest | Calculate dimensions from a target size or scale factor |
| Resize Image Megapixels | Resize an image batch to a target area without cropping |
| VAE DeGrid | Remove the 2px pixel grid that Qwen Image, Qwen Image 2.1 and Wan 2.1 VAEs leave in decoded images and video, without softening real detail |
| Zoom Crop | Remove width/height pixels and pick the kept area with a one-click 5×5 picker |
| Load Video (Upload) | Load a clip by frame cap and skip into an IMAGE batch, with audio, transparency mask, model frame presets, a readout of what will load, and a preview that loops the clip |
| Save Video | Save an IMAGE batch as MP4 (H.264, H.265, AV1), WebM, ProRes, lossless FFV1, GIF or WebP in one pass, with audio, transparency and the workflow embedded; no FFmpeg executable needed |
| Duration Selector | Pick a clip length that is legal for LTX, MiniMax H3, Wan, Hunyuan, Cosmos or Mochi, as frames or seconds kept in sync |
| Frame Slicer | Keep a run of frames (start to stop) from an IMAGE batch |
| Speedpaint | Basic painting on a blank canvas or imported image, updated directly downstream |
| Paint Mask | Paint a mask over an uploaded or connected image, updated directly downstream |
| Liquify Image | Push-warp files or connected image batches with editable strokes, updated directly downstream |
| Qwen Edit Mask | Paint where a Qwen-Image 2.1 edit should happen and get a guided reference image, mask and prompt |
| Image Edit Align Composite | Automatically align Qwen, Krea2 or other image edits and composite through a supplied or automatic mask |
| Image Edit Difference Mask | Align and isolate visual edits, with mask preview and cleanup controls |
| Image Edit Measure Drift | Optional measurement for inspecting or reusing an alignment fit |
| Sigma Curve | Draw a custom sigma schedule by dragging points on a graph, with presets and exact numeric entry |
| Slider | Remap slider loras to a standard strength range |
| Load LoRA Masked | Mask which areas of an image each lora influences |
| H3 Ripple | Edit a frame of a shot in any image editor; MiniMax H3 carries the edit through the whole shot |
H3 Ripple: Watch the examples and download the workflows.
Installation
From your ComfyUI custom_nodes directory:
git clone https://github.com/WepeNerd/ComfyUI-WepeNerd.git
Using the Python environment that runs ComfyUI:
python -m pip install -r ComfyUI-WepeNerd/requirements.txt
Restart ComfyUI and refresh the browser. This pack uses Pillow, numpy, SciPy, and PyAV, plus Torch supplied by ComfyUI. It does not install GPU wheels, models, or external runtimes.
Quick start
Add Speedpaint and connect its IMAGE output to Preview Image. Paint and click Queue to preview the result. Connect Resolution Suggest to its width and height inputs to control the canvas size, or VAE Encode to use the painting in a generation workflow.
-
Speedpaint: choose New for a blank background. Load runs a connected
imageupstream and imports its first image; without a connection it opens the file picker. You can also drop an image onto the canvas. Use round or square brushes, colour, size, opacity, and pressure-to-size controls. Toggle the eraser (E) to restore the loaded image or blank background; B returns to painting. Ctrl/Cmd+Z undoes; Shift+Ctrl/Cmd+Z redoes; Alt-click samples a colour. -
Resolution tools: choose a divisor that matches your model's required dimensions. Megapixel resizing retains the whole image, with small aspect changes possible from rounding.
-
Load Video (Upload): upload/select a video under WepeNerd/Video and connect
imagesto an IMAGE input. Pick the clip withskip_first_framesandframe_load_cap(121 by default; 0 loads all frames); large skips jump straight there. The node shows what will load before you run it, and the muted preview loops over exactly that clip. A format preset trims the frame count to a model's rule and sets its frame rate. Connectaudiodownstream to keep the clip's soundtrack, or disableload_audioto skip audio decoding. -
Save Video: connect
images(andaudio/maskif you have them) to WepeNerd/Video › Save Video. MP4 · H.264 has broad player support; pick WebM, ProRes, FFV1, GIF or WebP when you need transparency. Drop the saved file onto the canvas to reload its workflow. -
VAE DeGrid: put it straight after VAE Decode (before any resize, sharpener or upscaler) and leave the defaults. It shows the grid strength it measured; a batch with no grid passes through untouched. See VAE DeGrid.
-
Liquify: load an image, or connect IMAGE and Queue once to see its preview. Drag to warp, use Undo/Redo, then Queue to render at the source resolution.
-
Masked LoRA: connect a native Krea2 MODEL, select a LoRA, and open Edit mask. Empty masks have no effect. This node is specific to Krea2 and remains in Beta.
-
Qwen Edit Mask: open or drop an image (or connect IMAGE and click Load input), paint the area to change, and type the edit in
instruction. Connectimagetoimage_1andprompttoprompton Text Encode Qwen Image 2.1, and set the encoder'sresolutionto 0. See the guide options. -
Image Edit alignment: under WepeNerd/Image Edit, add Image Edit Align Composite, connect the original to
sourceand the generated edit toedited, then savecomposite. Affine alignment and a cleaned automatic mask are enabled by default; Measure Drift is optional. Use Image Edit Difference Mask for a mask preview and finer cleanup controls. Pixels where the final mask is zero are copied exactly from the source. See the controls and limits. -
H3 Ripple: open the frame-upload workflow, or the local image-editor workflow with two complete Qwen Image 2.1 editing branches. Both use Ripple Clip → Ripple Apply → Ripple Output and require MiniMax H3 FL2VA, its text encoder and video VAE, and
H3_Ripple_v1. H3 uses Pika H3 Flex at strength 1, 6 steps, LCM, and the simple schedule. Pick a video, time window, duration and megapixels; preview the crop before editing. The starting size is 0.25 MP, with 107 frames / 4.46 seconds. Download and edit a marked frame, then upload, drop or paste it onto its card. Add more anchors when other views need the same change; each marker gets its ownanchor_Noutput and a matchingedited_anchor_Ninput on Ripple Apply. Changing only megapixels resizes the existing edits; changing the crop or frame needs a fresh review. The local-editor workflow starts with Ripple Output muted: run and review both Qwen images first, then enable Ripple Output and run the video. Source audio and a source | result comparison are saved. See H3 Ripple for model files, geometry and anchor routing.
See the node guide for controls, inputs, outputs, and limitations.
Saving paintings
Speedpaint supports dimensions from 64 to 4096 pixels and PNG, JPEG, or WebP imports. Use Load to import a file or the first image from the optional IMAGE input. Resizing painted work resizes the whole composition with Lanczos.
Paintings are stored in ComfyUI/input/wepenerd_speedpaint. Copy that folder with
your workflows when moving machines. Standard Save/Export and Queue wait for
pending writes. Undo and redo share a 30-operation / 128 MiB history budget;
history does not persist after reloading.
Compatibility
Use a current ComfyUI version with the native model nodes required by your workflow. Python 3.10 or newer is required.
Masked LoRA supports native floating-point and INT8 ConvRot Krea2. Spatial masks can influence pixels outside the painted region through attention and denoising. A full mask does not necessarily match a global LoRA loader. See the node guide for unsupported model and adapter formats.
More WepeNerd tools
- LocalAI: local prompt enhancement and captioning.
- Experimental: 3D and video utilities.
The packages can be installed independently. Use current versions when combining them; the older all-in-one core includes duplicate nodes.
Liquify is included in core from 0.2.0. If you also have Experimental installed, update it to 0.2.0 or newer before restarting ComfyUI. Existing Liquify nodes and saved paintings keep their IDs and outputs.
Report an issue with a minimal workflow and the relevant console output. Licensed under MIT.