H3 Ripple in ComfyUI: Edit One Frame, Change the Whole Video (MiniMax H3 Setup Guide)

Editing one frame of a video is easy. Making that edit stick for every frame after it, without flicker, is the hard part.

H3 Ripple is a new LoRA and workflow pair for MiniMax H3 that does exactly that: you edit one or two “anchor” frames in any image editor, and H3 carries the change through the rest of the shot.

It comes from WepeNerd, the developer behind the ComfyUI-WepeNerd node pack, and it went up on Hugging Face at the end of October. This guide covers the files, the exact folders, both workflows (Simple and Local Editor), the settings, and what to do when the edit drifts or the run runs out of memory.

New to MiniMax H3? Start with our MiniMax H3 ComfyUI guide, which lists every H3 workflow, file and fix in one place.

What H3 Ripple Is

H3 Ripple is a LoRA for the MiniMax H3 video model, plus two ready-made ComfyUI workflows. It adds one job to H3: edit propagation. You give it a source video and one or more frames that you have already changed, and it renders a new video where the change holds from start to finish.

Think of jobs like these:

• Changing the colour of a jacket in a short clip.
• Adding a hat, glasses or a logo to a product shot.
• Swapping the sky or the season in a landscape pan.
• Turning a plain wall into a mural.

Doing this frame by frame in an image editor would take hours and still flicker. Ripple asks you to edit only the anchors and lets the video model fill in the rest, using the source clip to keep the motion.

The official guide page describes it in three steps: choose your moment, make the change, and let it ripple.

How Anchor Frames Work

An anchor is a frame of the source video that you mark and then edit. Ripple treats each edited anchor as the truth for that point in time, and the source video as the guide for motion between them.

A few rules follow from that:

• One anchor is enough when the change is visible the whole time from roughly the same angle. The Simple workflow uses one.
• Two or more anchors help when the subject turns or the camera moves, so a new side of the object comes into view. The Local Editor workflow uses two.
• Anchors must agree. If the jacket is red in anchor one and orange in anchor two, H3 will try to blend the two, and you get a colour shift mid-clip.

You place the anchors in the Ripple Clip node, which is also where you pick the shot (the start and end of the clip) and the crop and output size.

Files and Folders

Ripple runs on top of a normal H3 FL2VA setup. If you already run H3 first-last-frame video in ComfyUI, you only need the Ripple LoRA and the node pack. The workflows also include a “Models & setup” card that lists the exact downloads they expect.

The guide page says the Simple workflow needs five local model files and the Local Editor needs eight. The three extra files are the Qwen Image 2.1 set, which both edit branches share.

If you do not have the Qwen Image 2.1 files yet, our Qwen Image model files guide lists each file and folder, and the Qwen-Image 2.1 GGUF guide covers smaller quants for 8 to 12 GB cards.

The H3 subfolder inside loras is optional. ComfyUI finds LoRAs in subfolders, and the dropdown shows them as H3/minimax_H3_Ripple-v1.safetensors. If you put the file straight in loras/, just reselect it in the LoRA loader after you open the workflow.

Install the Nodes

The Ripple Clip, Ripple Apply and Ripple Output nodes come from the WepeNerd pack. The pack does not install models, GPU wheels or external runtimes, so it is a light install.

1. Open a terminal in ComfyUI/custom_nodes.
2. Clone the repo: git clone https://github.com/WepeNerd/ComfyUI-WepeNerd.git
3. Install its requirements with the same Python that runs ComfyUI: python -m pip install -r ComfyUI-WepeNerd/requirements.txt
4. Restart ComfyUI and refresh the browser tab.

The pack depends on Pillow, numpy, SciPy and PyAV, and uses the Torch that ComfyUI already has. On the Windows portable build, run pip through python_embeded\python.exe -m pip install -r ComfyUI\custom_nodes\ComfyUI-WepeNerd\requirements.txt from the portable folder.

The README says it was tested with ComfyUI 0.34.0, frontend 1.51.10, Python 3.12 and Windows. If you are on an older ComfyUI, update first. The pack also ships other useful nodes (Speedpaint, Paint Mask, Liquify, Qwen Edit Align Composite), so it is worth having anyway.

Workflow 1: Ripple Simple

Use H3_Ripple_Simple.json when you want full control over the edit. You make the change in Photoshop, Krita, GIMP, an online editor or any image model, then hand the frame back to ComfyUI.

1. Drag H3_Ripple_Simple.json into ComfyUI. Install any missing nodes it flags.
2. In Ripple Clip, load your video, choose the shot, set the output size and place one anchor marker on the frame you want to edit.
3. Download that anchor frame from the node.
4. Edit it in your image editor. Keep the same size and framing, and change only what you want changed.
5. Upload the edited frame to the matching anchor input.
6. Write a short prompt that describes the edited scene, for example “a woman walking down a street in a bright red raincoat”.
7. Queue the run. Ripple Output is enabled in the download, so queuing also renders the H3 video and saves a side-by-side comparison.

The Simple workflow renders 90 frames at 24 fps, which is 3.75 seconds, from a single anchor.

Workflow 2: Ripple Local Editor

Use H3_Ripple_Local_Editor.json when you want the whole job to stay in ComfyUI with no outside editor and no API. It runs two Qwen Image 2.1 branches to edit two anchors from text instructions, then passes both images into Ripple.

1. Load H3_Ripple_Local_Editor.json and make sure the three Qwen Image 2.1 files are in place.
2. In Ripple Clip, load the video and place two anchors, ideally where the subject looks different (for example, front-facing and side-on).
3. Replace both example edit instructions with your change. Use the same words in both, such as “change the jacket to bright red leather”.
4. Check the two edited anchors in the preview. If they disagree, fix the instruction or the seed before you render the video.
5. Queue the run.

This workflow renders 124 frames at 24 fps, which is about 5.17 seconds, from two anchors. The Qwen edits run at 25 steps, euler / simple, CFG 1, according to the model card.

For more on writing clean edit instructions, see our Qwen-Image 2.1 edit guide.

Settings

The model card and the guide page give slightly different speed setups, so here are both. Start with whatever the workflow you downloaded already uses.

If you already use Pika H3 Flex, our Pika H3 Flex setup guide has its files and LoRA loader order. Either speed LoRA works the same way here: load the base model, then the speed LoRA, then the Ripple LoRA, each with a Load LoRA (Model Only) node.

You can lower the output size in Ripple Clip. The guide’s own advice is to start small, check that the edit holds, and only then render at the full size.

Tips for Clean Edits

• Start with a short, steady shot. A slow pan or a person walking toward the camera is far easier than a spinning dancer.
• Change one thing at a time. A colour or a single object is reliable. A full outfit and a new background in one go is not.
• Keep the anchor the same size as the source. Cropping or resizing the edited frame shifts everything and the model has to guess.
• Match the light. If you paste in an object, match its shadows and colour to the scene, or it will look stuck on.
• Add an anchor where new detail appears. If the back of the jacket only shows at the end, put the second anchor there.
• Describe the edited scene in the prompt, not the change. “A man in a green cap” works better than “make the cap green”.

Ripple vs Other H3 Editing Tools

Ripple is the only one of these that starts from your own footage and keeps its motion while changing what is in it.

Troubleshooting

Ripple Clip, Ripple Apply or Ripple Output show in red

The WepeNerd pack is missing or failed to load. Check the ComfyUI console for an import error, usually PyAV or SciPy. Re-run the pip install with ComfyUI’s own Python, then restart.

The Ripple LoRA is not in the dropdown

Make sure the file is in ComfyUI/models/loras/ or its H3 subfolder, then press R in ComfyUI to refresh the model lists. Reselect the file in the loader, because a saved workflow stores the path it was built with.

Out of memory

H3 is a large model. Lower the output size in Ripple Clip, use the int8 or W4A8 FL2VA build, and close other apps using the GPU. On the Local Editor workflow, the Qwen branches add their own load, so try the Simple workflow first.

Our H3 out of memory fix for 16 GB lists every setting that helps.

The edit fades or flickers part way through

This is the most common problem, and the guide warns about it. Fast motion, something passing in front of the subject, or a big change of angle make the edit hard to carry. Add an anchor near the point where it fades, shorten the shot, or pick a steadier clip.

Things I did not edit changed too

The model card says generated frames can alter details outside the intended edit. Keep the prompt close to the source scene, and keep edits small. For the Local Editor, WepeNerd’s Qwen Edit Difference Mask + Composite node can help keep the anchor’s untouched pixels identical to the source before Ripple sees them.

Colours jump between the two anchors

The two anchors disagree. Use the same instruction and seed for both Qwen branches, or edit both frames by hand from the same reference.

FAQ

Is H3 Ripple free?

Yes, the LoRA and workflows are free to download. The node pack code is MIT licensed. The LoRA card does not state its own licence, and it notes that the MiniMax H3 base model is under the MiniMax H3 Community License Agreement, so check that before commercial use.

Do I need an API or cloud service?

No. The Local Editor workflow runs the anchor edits with Qwen Image 2.1 on your own GPU. The Simple workflow lets you use any editor you like.

How long can the video be?

The two downloads render 90 frames (3.75 s) and 124 frames (about 5.17 s) at 24 fps. Longer clips are harder to keep consistent, so build up from these.

How much VRAM does it need?

The guide does not give a figure. It needs hardware that can already run MiniMax H3 FL2VA, and the Local Editor also runs Qwen Image 2.1. The Ripple LoRA itself is small.

Can I use more than two anchors?

The guide says to add more anchors if other views need the same change. Keep every anchor consistent with the others.

Does it keep the original audio?

The demo video has audio and the example comparison clips are silent. H3 FL2VA can generate audio, so check the workflow’s audio branch if you need sound.

Which speed LoRA should I use?

Use whichever the workflow you downloaded is built for. The model card lists the H3 Turbo 4-step LoRA, and the guide page says its downloads are set for Pika H3 Flex at 6 steps with LCM.

Sources

• WepeNerd/H3_Ripple on Hugging Face (model card)
• H3 Ripple guide page
• WepeNerd/ComfyUI-WepeNerd on GitHub
• Comfy-Org/MiniMax-H3 and lightx2v/Minimax-h3-Turbo