Load Kiwi Edit Pipeline (Yogurt Kiwi Edit)

This node is the front door of the whole ComfyUI-YogurtKiwiEdit pack, and the name tells you exactly what it does: it loads the Kiwi-Edit pipeline and hands you a pipeline handle you'll feed to every other node in the graph. No magic, no API key, no cloud call - just a diffusers model folder being pulled into memory so the four stage nodes downstream have something to work with.

Kiwi-Edit itself is a video editor from showlab that shipped in March 2026. You feed it a source video plus an instruction ("add a red hat to the woman while keeping the facial identity unchanged" - that's literally the pack's example prompt) and it edits the clip. It's a research release built on Wan: the model folder declares a WanTransformer3DModel, a Wan 3D VAE, and a FlowMatch scheduler, with a Qwen-based multimodal encoder bolted on the front. Community reception at launch was lukewarm at best - the top comment on the announcement thread was "almost useless, zero good results so far" - and one of the few other people in that thread asked for a ComfyUI workflow. This pack is basically that wish, fulfilled by an author (yogurt7771) with no other footprint in the ecosystem. Expect research-model rough edges; the bridge is the useful part.

What it actually does

The node scans ComfyUI/models/KiwiEdit for any subfolder containing a model_index.json and turns the results into the model_name dropdown. Pick one, and it runs diffusers' DiffusionPipeline.from_pretrained with trust_remote_code=True. That flag matters: a research repo like Kiwi-Edit ships its own pipeline code (pipeline_kiwi_edit.py) in the model folder, and from_pretrained executes it. This is normal for this class of model, just worth knowing - the model folder is code, not only weights.

It's also why the download is big. The full instruct+reference model runs about 20 GB on disk, of which the 5B transformer is ~10 GB and the MLLM encoder another ~7.5 GB. Not a toy.

The inputs that matter

• model_name - dropdown, populated from models/KiwiEdit. If it shows [No models found in models/KiwiEdit], your model folder isn't where the node looks, or it's missing model_index.json.
• dtype - auto (the default) picks bfloat16 on cards that support it, otherwise float16, otherwise float32. Leave it on auto unless you have a reason not to.
• cpu_offload - on by default, and it's the whole point of this pack. With it on, enable_model_cpu_offload() keeps the enormous pipeline resident in RAM and shuffles only the needed module group onto the GPU per stage. Turn it off and pipe.to(device) tries to hold everything at once, which on anything short of a 24 GB card is how you get OOM.

The single output, pipeline, is a handle, not a tensor. Wire it into all four stage nodes - VAE Encode Inputs, MLLM Encode Context, Generate Latents, and VAE Decode - they check that the handle's model path matches theirs, so don't mix two pipelines in one graph.

Installing it

Two ways, both standard:

# ComfyUI Manager: search "ComfyUI-YogurtKiwiEdit", install, restart.
# Manual:
cd ComfyUI/custom_nodes
git clone https://github.com/yogurt7771/ComfyUI-YogurtKiwiEdit
cd ComfyUI-YogurtKiwiEdit
pip install -r requirements.txt   # diffusers>=0.36.0, accelerate
# restart ComfyUI

Then grab a model. All three variants live under the linyq org on Hugging Face: kiwi-edit-5b-instruct-only-diffusers, kiwi-edit-5b-reference-only-diffusers (the one the example workflow uses), and kiwi-edit-5b-instruct-reference-diffusers (both capabilities). Clone or download one and put the folder - not a .safetensors - at ComfyUI/models/KiwiEdit/<folder>/model_index.json. The folder name is what shows up in the dropdown (the example workflow lists it as linyq--kiwi-edit-5b-reference-only-diffusers).

Where people get stuck

Mostly the model. [No models found] almost always means the download landed as a nested folder (you dropped KiwiEdit/kiwi-edit-... inside itself) or you grabbed a single weights file instead of the diffusers directory. And remember the model is a March 2026 research release with a thin fanbase - if your first edits look mediocre, that's the model, not your wiring.