ComfyUI-SuiteTea

Some good ComfyUI nodes for bad reasons.

Install

• ComfyUI Manager → Install from URL →
  https://github.com/teepunkt-esspunkt/ComfyUI-SuiteTea.git
• Or manually:
  git clone https://github.com/teepunkt-esspunkt/ComfyUI-SuiteTea.git
  into your ComfyUI/custom_nodes/ folder.

---

Nodes

Tea: Save & Reload Image (category: SuiteTea / IO)

A utility node to save VRAM on older GPUs. Many workflows pass images directly from one model to another → this can cause out-of-memory (OOM) errors on the first run. This node saves the image to disk and reloads it, forcing upstream tensors to unload.

• Works as both a detacher and a normal image loader.
• Has a file picker with preview (like the stock Load Image node).
• Cleaner defaults: Teafault.png, output/temp, output/saved.

Inputs

• image_in (optional IMAGE tensor) → triggers save→reload.
• image(picker) → choose/upload an image with preview.
• temp_folder (default output/temp)
• filename (default Teafault.png)
• also_save_perm (BOOLEAN, default false)
• perm_folder (default output/saved)

Output

• reloaded_image (BHWC float, shape 1×H×W×3)

Usage

• To break tensor lineage: Model Output → Tea: Save & Reload Image V2 → Next Node

• To just load a file: leave image_in unconnected and pick a file.

---

Tea: CheckpointLoader (category: SuiteTea / Loaders)

This Node is created to help run a python script run the same workflow with different models A string-based checkpoint loader to work with external Python batch scripts. Lets you loop the same workflow across multiple models without clicking through the dropdown.

Inputs

• ckpt_path (STRING, full path to .safetensors or .ckpt)

Usage (model loop workflow)

1. Create a suiteTea_local.json in suitetea/scripts/ with your private model folder path: { "MODELS_DIR": "C:/your/full/path/to/checkpoints" }
2. Run discover_models_flat.py → generates models_list.txt.
3. Build a workflow modelloop.json using Tea: CheckpointLoader instead of the dropdown loader.
4. Run run_all_models.py → will iterate through all models in models_list.txt using the same workflow.

---

Tea: Load Fram from Vid As Img (category: SuiteTea / IO)

Extract a single frame from any video and output it as an IMAGE tensor. Useful for extending clips from the last frame, grabbing a reference still, or snapshotting a timestamp—while keeping VRAM usage low. Optionally saves the frame as a PNG to reuse in later chains.

Inputs

• video(picker) → choose a video
• mode(first | last | index | time)
• video_path(STRING, optional override; if set, this path is used instead of the picker)
• frame_index(INT, used when mode=index)
• time_Sec(FLOAT, used when mode=time)
• max_side(INT, 0 = no resize; otherwise downscales keeping aspect, e.g. 1024)
• save_png(STRING, if empty, auto-names to output/tea_frames/<video>_<tag>.png)
• overwrite(BOOLEAN)

Outputs

• image(BHCW float, shape 1xHxWx3)
• saved_path(STRING; empty if save_png=false)
• picked_index(INT; returns the frame index in indexmode, otherwise -1)

Usage

• Extend a clip from its last frame: Tea: Load Frame From Vid As Img (mode=last) → (optional) VAE Encode → your i2v/video pipeline
• Grab a specific moment: mode=time, set time_sec=2.5 to pick the frame at 2.5s or mode=index, set frame_index=123
• Persist the still: Toggle save_png=true (and/or set save_path) to store a reusable PNG and break tensor lineage.

Notes

• Prefers ffmpeg (fast, robust). If ffmpeg isn’t on PATH, it falls back to OpenCV if installed.
• Works with common formats (mp4, mov, webm, mkv, …) as supported by ffmpeg.
• The node returns a CPU tensor; VRAM is only touched when you feed it into a VAE/model.
• If both video and video_path are provided, video_path takes precedence.

Scripts

Located in suitetea/scripts/ — helper utilities for batch workflows:

• discover_models_flat.py: Scans your private models folder (from suiteTea_local.json) and writes models_list.txt. Run this whenever you add/remove checkpoints.

• run_all_models.py Reads models_list.txt and your exported workflow (modelloop.json). Runs the workflow once for each model, saving results into a timestamped folder with the model name as filename prefix.

---

(More nodes will be added here as SuiteTea grows.)

License

MIT