Edit Anything v2: Multi-Instruction Video Editing for LTX-2.5

Edit Anything v2 (Hugging Face | Civitai) is a rank-128 instruction IC-LoRA for LTX-2.5 (22B dev) released by Alissonerdx. Unlike the earlier LTX-2.3 tracks, it performs several editing operations — Replace, Remove, Add, Restyle — in a single prompt, takes no reference image, and ships with a ready-to-use ComfyUI workflow.

What's new in v2

The first Edit Anything release (June, covered here) targeted LTX-2.3 with three separate tracks: motion transfer, a no-reference multitask LoRA, and a reference V2V build. Edit Anything v2 is a clean restart:

Built on LTX-2.5 — trained from scratch on LTX-2.5 22B dev, not a continuation of the LTX-2.3 LoRAs. 2750 aligned source/edited video pairs, rank 128 / alpha 128, at 704×384, 73 frames.
Multiple operations in one prompt — "Replace the purple Dodge Challenger with a green Lamborghini, and Add a cinematic apocalyptic scene with falling bombs, and Restyle the sky into a sunset" all in a single pass. v1 could only do one operation at a time.
No reference image — the source video goes in as the guide, the instruction goes in as text, the edited video comes out. No RoPE tricks, no sidecar modules, no loss masks.
Prompt enhancer — write in natural language and it is converted into the appropriate instruction prompt.

Examples

Each clip is a three-panel comparison: source guide | first pass | second pass (upscaled).

Replace the background, keep the dancer

edit_anything: Replace the green indoor background with a sandy beach under a bright sky, where the man dances in dark blue swim trunks and a white beach towel draped over his shoulder, holding a yellow inflatable ring around his waist.

Replace a car + keep the subject

edit_anything: Replace the purple Dodge Challenger in the background with a green Lamborghini Huracan, featuring black five-spoke wheels and a low front splitter. The woman remains fully visible throughout the video, wearing her brown suit as she dances in the foreground.

Two edits at once

edit_anything: Replace the woman's dark black hair with vibrant red hair. Add a metallic leather metalworker outfit, including a sleeveless top and pants, worn by her in the center of the frame.

ComfyUI usage

Load workflows/EditAnythingLTX2.5.json from the Hugging Face repo, or wire it by hand with the BFSnodes LTX Multiple Controls node. Settings that match training:

input	value
`guide_video`	your source video
`guide_source_id`	`0`
`guide_layout`	`overlap`
`guide_ref_resize_mode`	`match_target`
`guide_downscale_factor`	`1`

Leave identity_image, mask_video and identity_mask_image disconnected — the LoRA was trained without any of them.

Suggested start: LoRA strength 1.0, CFG 3–5, 30 steps, no LightX2V. The trained box is 704×384, 73 frames — other sizes and lengths are extrapolation and degrade in temporal consistency first.

Prompting rules

Every prompt starts with the trigger edit_anything: followed by one sentence per edit. The model understands only four verbs (most to least reliable): Replace, Add, Remove, Restyle. Rules that matter:

10 to 20 words per instruction; shorter loses description, longer drifts.
One object per Add; for a group or crowd, use Replace on an existing region instead.
Anchor each instruction to something visible ("to the left of the man", "on the desk on the right side of the frame").
Describe state, not manner — "standing in the green grass field", not "swimming gracefully".
Restyle targets a region, not the whole clip; for a whole-video look, write two or three Restyle instructions over the regions that carry it.
Make it <style> and Turn it into <style> do not work — use Restyle.

Limitations

The editing vocabulary is limited to the four verbs; edits that change very little of the frame (subtle removals) are the least reliable; there is no reference-image support ("add this object" with a picture remains unsolved); and clips longer than the trained box lose temporal consistency first. As with v1, these are research experiments — expect failures and iterate on the prompt.