Kiwi Edit Generate Latents (Yogurt Kiwi Edit)

This is the heart of the pack: the actual denoising loop. Everything before it - VAE encoding, the MLLM context - is preparation. Everything after is just turning latents back into pixels. If you want to know what "the model" is doing, this node is where it's doing it.

Stage three of the split design, and the reason the split exists. Kiwi-Edit's transformer is a 5B-parameter Wan-family model, ~10 GB of the download in bf16. Running that alongside the MLLM encoder and the VAE at once is how you OOM a mid-range card. This node loads only the transformer plus the source/reference embedders onto the GPU, denoises, then shoves them back to RAM and frees cache. Peak VRAM is the transformer's, and only the transformer's.

The inputs that matter

• pipeline and conditioning and context - the three handles from the previous stages. All three must come from the same model; the node checks.
• num_inference_steps (default 50) - denoising iterations. 20–30 still works for flow-matching models; 50 is the safe default. Fewer steps = faster, slightly rougher.
• seed (default 0) - with a fixed seed the noise is fully reproducible (it's built on a CPU generator, so it's also CPU/GPU-independent).
• sigma_shift (default 5) - the flow-matching shift parameter. It bends how the scheduler spends its steps. The default is what the model was tuned with; leave it alone until you've got a feel for the model.
• denoising_strength (default 1.0) - this is the one to actually touch. At 1.0, the loop starts from pure noise: a fresh generation. If you connected an input_video in VAE Encode Inputs, it gets noise added at this strength - so 0.6-ish gives you a strong video-to-video edit, and lower values keep the input closer.
• guidance_scale (default 5) - here's the honest part. It's in the UI, it's in the schema, and the source code literally dels it on the first line of the function. The current Kiwi-Edit implementation has no CFG path, so this knob does nothing. Don't chase it.

The output is latents - the denoised latent video, still compressed in latent space. It goes to Kiwi Edit VAE Decode; nothing in ComfyUI's normal preview chain can show it as-is.

What happens under the hood

The node sets up the FlowMatch scheduler with your step count, strength, and shift, builds the noise tensor, then runs a loop: model forward pass, scheduler step, repeat - with a real ComfyUI progress bar so you can watch it crawl. If input_video latents are present it starts from the partially-noised input instead of pure noise, which is the whole video-to-video mechanism.

Installing it

Identical to the rest of the pack: ComfyUI Manager (search "ComfyUI-YogurtKiwiEdit") or git clone into custom_nodes, then pip install -r requirements.txt and restart. The loader node, the model folder under models/KiwiEdit, and both upstream stages all have to be in place first - this node is the middle of the chain, not the start.

Where people get stuck

Expect the long wait. Even on a decent card, 50 steps of a 5B Wan transformer over 81 frames takes a while, and CPU-offload swaps add to it - this is not a real-time tool. Don't be surprised by the guidance_scale knob doing nothing (that's upstream, not a bug in your graph), and remember there's no negative prompt in this version either. If your video-to-video edits look barely changed, raise denoising_strength; if they look like the source got destroyed, lower it. That's the knob with real power here.