Edit Anything
Alissonerdx/EditAnything
published May 2026 · updated May 2026
Edit Anything is an experimental video editing model that uses LoRAs on LTX-2.3 to enable motion transfer, prompt-driven multitask edits, and reference-based video-to-video transformation.
specs
| Task | Video Editing |
| Architecture | LTX-2.3 (22B) LoRA |
| Parameters | 22B base model + LoRA ranks 128/256 |
about this model
EditAnything is a set of experimental LoRAs for video editing built on LTX-2.3 (22B), hosted as a managed API. The collection comprises three independent training tracks, each offering distinct video editing capabilities.
Motion transfer LoRA (v0.1)
This track enables motion transfer: the user replaces the first frame of a guide video with an edited still (e.g., from Flux Klein), and the model propagates that appearance through the entire clip while copying the motion of the original guide. Two rank variants are available, extracted via truncated SVD:
| File | Rank | Size | Frobenius retention |
|---|---|---|---|
edit_anything_30k_v0.1_motion_transfer_r128.safetensors | 128 | 1.31 GB | ~99.4% |
edit_anything_30k_v0.1_motion_transfer_r256.safetensors | 256 | 2.62 GB | ~99.9% |
Limitations: hard scene cuts and very fast motion degrade output; poor blending in the first frame propagates. Works best when the inserted subject occupies a similar region to the original.
No-reference multitask LoRA (v1.1)
A prompt-only multitask editor trained on a balanced mix of Add, Remove, Replace, and Style edits. No reference image is required; the model is driven entirely by the text prompt. It offers greater temporal consistency than v0.1 for these tasks because it was trained directly on video pairs. Does not support motion transfer, compositional prompts, background changes, or global color grading.
Reference video-to-video (Ref V2V, experimental)
Two builds of an IC-LoRA plus sidecar modules designed for strong identity transfer from a reference image into a video, targeted at Add and Replace edits. Requires BFSnodes ComfyUI custom nodes to load sidecar modules; not yet available as a standalone API endpoint. Evaluation by the community is ongoing.
All variants are research experiments released for community testing; they are not production-ready and will fail on many inputs.
best for
- ·Motion transfer: replace first frame of a guide video with an edited still and copy motion
- ·Prompt-driven multitask editing: add, remove, replace objects or apply style using text only
- ·Reference video-to-video: transfer identity from a reference image to video edits
FAQ
Motion transfer, no-reference multitask editing (add/remove/replace/style), and reference-based video-to-video transformation. It is experimental and may fail on many inputs.
The LoRA files are 1.31 GB (rank 128) and 2.62 GB (rank 256). Inference speed depends on base model (22B parameters) and hardware; no specific benchmarks are provided.
Hard scene cuts and very fast motion cause artifacts. Multi-action prompts (e.g., "add X and remove Y") are not supported. No global color/lighting changes. Identity transfer requires a strong first frame.
Use the gigarouter OpenAI-compatible endpoint with your API key. Send video frames and a text prompt (e.g., "Replace the statue with a man in a raincoat") in the expected format.
The model card does not specify a license; it is released as a research experiment for community testing.
We're benchmarking and onboarding Edit Anything as a hosted, OpenAI-compatible API. Sign in for free credit and be ready when it lands, or tell us you want it and we'll prioritize it.