Experimental research release. Results vary by clip, and nothing described here is guaranteed.

This LoRA edits MiniMax-H3 video using the source clip as a frame-aligned guide, not as a loose reference. The goal is to return the same take with only the requested change applied, while better preserving timing, motion, framing, pose, and camera movement.

Two Versions (Super Important)

- minimax_h3_vfx_edit_v1.0_r128.safetensors (I recommend using this unless you want to try sending the first frame or if you want to test if the quality is higher) — the original version, trained with the 384 resolution setting. Its common 16:9 training bucket was 512×288, using 73-frame clips.

- minimax_h3_vfx_edit_v1.0_r128_ffp.safetensors (The worst part is that I also recommend this one haha, I really like the result of this one and also the possibility of passing the edited first frame. The problem is that with a strength of 1.0 it tends to change much more than it should, so the right thing to do would be to try with around 1.0/0.8/0.7) — the newer version. FFP means First Frame Propagate. It received later training at the higher 512 resolution setting, commonly 672×384 for 16:9 footage, using 124-frame clips.

The FFP version can also use an edited first frame sent through MiniMax-H3's native image-reference channel while the original video remains connected through Add Guide. This feature does not work perfectly on every clip. The prompt must explicitly refer to the supplied image, for example, <Picture 1> and ask the model to match or propagate its edit. The auto-prompter normally adds this wording for you, but always review the generated prompt.

Another important detail I recommend you test with both Taomate Turbo Lora, it was extracted from FL2VA but works reasonably well with the base REF2VA, there's another one that's basically one of the best for me, which is DMD Turbo Lora, it was extracted specifically for the base REF2VA model and often works without problems. If you're going to test the version with FFP in the name, use DMD Turbo Lora.

DMD 8 steps Turbo LoRA (recommended with ffp)

Taomate 3 steps Turbo LoRA (fl2va but works with more steps in the ref2va)

Quick Start

- Begin every prompt with `vfx_edit:`.
- Start with LoRA strength `1.0`.
- Connect the source video to **Add Guide**, with its VAE connected.
- Match the source aspect ratio and use the highest practical resolution.
- Keep clips short; `124 frames` is about five seconds at 24 fps.

Example:

vfx_edit: Turn the man's skin into melting synthetic tissue that gradually reveals robotic components underneath, while preserving his identity, pose, hands, clothing, camera motion, framing, lighting, background, and every unrelated element.


Write one concise sentence containing two things: what must change and what must remain unchanged. Name important details explicitly, but never ask the model to preserve something the edit must replace. Training prompts were roughly 16–68 words, with a median near 38; longer prompts can introduce contradictions and increase drift.

An automatic prompt generator can translate casual requests into this format, but it can misunderstand the scene or incorrectly assume that a reference image was supplied. Always review its output. If no reference was provided, say so clearly and remove any invented `<Picture 1>` or reference wording. Also verify that the preservation clause does not protect the area you want changed.

What It Usually Does Well

- Location and background changes
- Relighting, color, weather, and atmosphere
- Object replacement
- Effects attached to a subject or part of the frame
- Propagating the look of an edited frame across a short shot

Reference Images

You may provide one image showing either an element to insert or the finished look of an edited frame. Describe visible objects briefly, but when using an already-edited frame, simply ask the video to match it instead of re-describing it. Use an image at least as large as the output; small references may be ignored. Never mention a picture that was not actually supplied.

Limitations

This is not a face-swap or compositing tool. It does not produce layers, alpha channels, or roto. Exact identity, text, logos, tiny faces, fingers, hair, railings, and fast thin details may drift or deform. It cannot invent camera movement and works best on short, single-shot clips. Strong edits can still alter unrelated content.

Long clips consume substantially more time and VRAM because both source and output video are processed together. Cut sequences into individual shots, edit them separately, and reassemble afterward. Testing was performed on an RTX 5090; smaller GPUs may require lower resolution, which usually increases softness, flicker, and detail loss.

The aligned guide improves the odds of preserving the original take—it does not guarantee it. Treat every result as an experiment.