Custom nodes
The parts under the hood.
Every Gmanski workflow is built from stock ComfyUI plus one node pack of our own. It ships inside the RunPod template, and the free workflows run with it out of the box.
LoRA & checkpoint router
Reads a caption of each picture and builds the right stack for it — checkpoint, content, body and clothing LoRAs at their authors' strengths — and reports what it chose.
- Krea 2 LoRA Router — caption in, checkpoint + LoRA stack + report out
- Apply LoRA Stack — applies the router's stack in one node
- Load Checkpoint By Name — loads a UNet chosen at run time, with fallback
- Apply LoRA By Name — a LoRA picked by a string, not a dropdown
Folders, not files
Run a whole folder through any workflow. Batch count is the number of files; each run takes the next one and names the output after it.
- Folder Image Loader — next image from a folder each run, with its name
- Load Image From Path — an image from any path on the pod
- Filename From Title — tidy output names from a title or source file
Long clips, in chunks
Upscale minutes of video on one GPU: the loader seeks one window per run, the writer stitches every window back with the original audio.
- Video Chunk Loader — one overlapping window per run
- Video Chunk Writer — stitches windows + audio into one file
- Video Scene Planner — turns an idea into shot-by-shot prompts
- Film Grain — matched grain so upscales don't look plastic
A library of people
Save a character once — description, reference photos, a LoRA per model — and load it by name in any workflow.
- Save Character / Load Character — one record, every model's LoRA slot
- Apply Character LoRA — picks the right LoRA for the loaded model
- Extract Character Traits — a reusable description from one photo
- H3 Reference Prompt — MiniMax H3 reference prompts done right
Prompts that write themselves
Scene-only captions for img2img — so the prompt describes the picture but never fights the face you're putting in it.
- Auto Caption — one node, sensible defaults
- Caption Batch — a folder of captions for LoRA datasets
- Prompt Enhancer — short idea in, full prompt out
The last ten percent
Faces, masks and resolution — the passes that make a render look photographed.
- Face Restore — GFPGAN / CodeFormer with blend control
- Person Mask — clean person masks for inpainting
- Tile Upscale — big upscales without seams
- Song Planner — structure and lyrics for MiniMax Music
Installing outside the template
A standalone repository for ComfyUI Manager is on the way soon. Until then the pack is pre-installed on the template, and the Pro membership includes direct support for custom setups.