SamplerCustom (Mochi Wrapper)

If you've used ComfyUI's KSampler or SamplerCustom, you already know how to use this node - which is the point. SamplerCustom (Mochi Wrapper) is the same idea rebuilt for kijai's Mochi wrapper: you hand it a model, conditioning, sigmas, latents, and a SAMPLER, and it runs the sampling loop. In the logtd/ComfyUI-MochiEdit pack, it's the workhorse that both the Mochi Unsampler and Mochi Resampler plug into. On its own it does nothing interesting; with those two custom samplers it becomes the engine of the whole RF-Inversion video-edit trick.

The confusing part for beginners: this one node gets used in two completely different configurations depending on which direction you're going. That's not a bug, it's the design.

The inputs

Nine inputs, and the ones that matter depend on the pass:

• model - a MOCHIMODEL, i.e. the Mochi model as loaded by the wrapper's loader nodes. This type only exists because you installed ComfyUI-MochiWrapper.
• positive / negative - conditioning. In the unsampling pass (video → noise) both must be blank prompts. In the resampling pass (noise → edited video) positive is your target prompt and negative can be whatever works for normal Mochi.
• cfg - the classic guidance scale. Unsampling: always 1.0. Resampling: whatever Mochi likes, roughly 4.5.
• sigmas - must come through Mochi Prepare Sigmas, then flipped for unsampling, not flipped for resampling. This is the #1 thing people get backwards.
• latents - for unsampling, the encoded video; for resampling, the unsampled noise from the first pass. Do not feed the original video here on the second pass or the "edit" is just the original.
• add_noise - False for unsampling (the video is already your starting point). Leave it on for normal generation if you use this node outside the edit workflow.
• seed - effectively irrelevant in the edit workflow; the author says there's no reason to change it.
• sampler - the SAMPLER from Mochi Unsampler or Mochi Resampler. This is the knob that makes the pass go backward or forward-with-guidance.

Output is a single samples LATENT, ready for the wrapper's VAE decoder.

How it works under the hood

The source wires the model through a model_fn that runs the DiT and applies CFG exactly like ComfyUI's own samplers (out_uncond + cfg * (out_cond - out_uncond), or just conditional when cfg <= 1.0), then hands the whole thing to the SAMPLER function you connected. Notably, it also accepts conditioning straight from the standard CLIPTextEncode output, which saves you from hunting for wrapper-specific text-encode nodes. It's a thin, well-behaved adapter - which is exactly what you want from a node that has to be correct in two opposite modes.

Install

This pack has no Mochi support of its own, so install the wrapper first:

cd ComfyUI/custom_nodes
git clone https://github.com/logtd/ComfyUI-MochiEdit

Restart. No extra Python requirements - the pack ships no requirements.txt.

Gotchas

The classic failure is a mismatch between the wrapper and this pack: the repos were merged together by their authors (the git history literally contains a merge from kijai), so if the node errors or the MOCHIMODEL type won't connect, update the wrapper first - that was the fix people needed at launch. And set expectations about the ecosystem: this is a November 2024 tool for Mochi, a VRAM-hungry early video model that most people have since left for LTX or Wan. Niche, but if you're editing Mochi clips, this is the sampler you'll reach for every time.