1
What you will build
This workflow creates a prompt-based edit of an existing video clip. In the source video, Prompt Mastery demonstrates edits such as adding fire to hands, turning an object into a hammer, changing hair color, and making a person appear behind the speaker. The key trick is not to regenerate the whole final video. Instead, you cut out only the segment that needs the effect, run that segment through WAN2GP with the Edit Anything LoRA, then place the generated segment back into the original edit.
1. Film or select source clip
2. Cut at edit point
3. Export short control video
4. Generate in WAN2GP
5. Upscale, color match, stitch
Recommended path: Work on a short 5-second section at
480p, use one phase generation, keep the control video strength at 1, and upscale the result afterward. That is the path shown in the video and it is much faster than trying to generate the final resolution directly.2
Source limits and assumptions
This guide is based on the YouTube video “LTX 2.3: Edit Anything LoRA Works? (WAN2GP Version)” by Prompt Mastery, its auto-generated English transcript, the video description, and current model/workflow pages for the EditAnything LoRA. The transcript clearly captures the WAN2GP settings and creative process, but it does not include the creator’s exact full prompt text or Patreon resource files.
- The video description says the creator’s raw output videos, metadata, prompt structures, LoRA download link, and installation notes are available in a free Patreon resource pack.
- Because that pack was not accessed here, the prompt examples below use the public EditAnything prompt patterns from Civitai/Hugging Face plus the demonstrated WAN2GP settings from the video.
- Version-sensitive model behavior may change because the EditAnything LoRA is experimental and still being updated.
3
Prerequisites
Software and workflow
- WAN2GP installed and working.
- LTX 2.3 model support available inside WAN2GP.
- The EditAnything / Edit Anything LoRA for LTX 2.3, preferably the
9000stepscheckpoint used in the video. - A timeline editor such as CapCut, Premiere, DaVinci Resolve, or Final Cut.
- An upscaling path, such as ComfyUI RTX Super Resolution, Topaz, or another video upscaler.
Hardware expectation
- The creator tested on an RTX 3090 with 24 GB VRAM, AMD Ryzen 9 5950X, and 64 GB RAM.
- The LoRA file is roughly 1.2–1.3 GB, so leave VRAM headroom.
- For a 5-second clip at
480p, the creator reported about 89 seconds generation time on an RTX 3090.
Do not start at 720p or 1080p. The video specifically recommends
480p because the workflow is compute-intensive and can be upscaled later.4
Prepare the control clip
- Open the original video in your editor. The video demonstration uses CapCut, but the same logic applies anywhere you can cut and export a section.
- Find the exact frame where the edit should begin. For example, if you want a person to appear behind you, place the playhead at the first frame where they should appear.
- Add a marker at that frame. In CapCut, place a marker at the transition point so you can cut cleanly.
- Split the clip at the marker. Keep the unedited part before the marker unchanged.
- Export only the segment that needs the AI edit. This exported segment becomes the control video for WAN2GP.
- Keep the segment short. The video example discusses short clips around 5 seconds. Shorter segments are easier to iterate and less likely to drift.
Why this works: By editing only the back half of the shot, the original footage and AI-edited footage can meet at a clean transition point. If the AI preserves the speaker and background, the edit feels like a practical effect rather than a full regeneration.
5
Use the WAN2GP settings shown in the video
Inside WAN2GP, the creator uses LTX 2.3 with the Edit Anything LoRA and a control video path. Use this as your baseline before experimenting.
| Area | Setting | Recommended value |
|---|---|---|
| Model | LTX model choice | LTX 2.3 Dist / LTX 2.3 distilled option |
| Generation mode | Main mode | Text prompt only |
| Control video | Control video type | LTX2 Raw Format Control Video for IC LoRA |
| Area process | Frame area | Whole frame |
| Control strength | Control video strength | 1 |
| Output basis | Generation source | Generate video based on control video and audio track and text prompt |
| Resolution | Category / size | 480p for first pass |
| Advanced → MISC | Overwrite frames per second | 30 |
| General settings | Phase count | One phase |
| LoRA | EditAnything checkpoint | ltx23_edit_anything_global_rank128_v1_9000steps_adamw.safetensors or the 9000-step option available in your install |
Critical setting: Use
one phase, not two phases. In the video, two-phase generation caused identity drift, background changes, and creepy replacement subjects. One phase took longer but preserved the source identity and scene much better.Fast setup from existing generated clips
If you have a successful WAN2GP output file with embedded metadata, use WAN2GP’s import path:
- Go to
Import media. - Choose
Import media to gallery. - Select a previously generated output video.
- Use
Extract settings. - Replace only the prompt and control video for your new edit.
This is the easiest way to reproduce the creator’s settings if you download his raw files or build your own library of successful outputs.
6
Prompt in the LoRA’s training format
The video emphasizes that this LoRA was trained around specific prompt patterns. Do not write a vague cinematic prompt first. Start with the operation type: add, remove, replace, or convert/style.
| Operation | Prompt pattern | Example |
|---|---|---|
| Add | Add a/an [subject/object] with [clear visual attributes], [precise location in the scene]. | Add a woman with long dark hair wearing a red dress, standing behind the man on the right side of the room. |
| Remove | Remove the [subject/object] [location or identifying description]. | Remove the black microphone in the foreground near the speaker's chest. |
| Replace | Replace the [original subject/object] [location] with a/an [new subject/object] with [clear visual attributes]. | Replace the phone in the man's left hand with a large steel warrior hammer with a leather handle. |
| Convert / Style | Convert the video into a [style name] style. | Convert the video into a cinematic anime style. |
Use an LLM to enforce the format
The creator suggests giving the format examples to Gemini or another language model, asking it to study the structure, and then writing your intent in plain language. Use a prompt like this:
You are helping me write prompts for the LTX 2.3 Edit Anything LoRA.
Use exactly one of these formats:
1. Add a/an [subject/object] with [clear visual attributes], [precise location in the scene].
2. Remove the [subject/object] [location or identifying description].
3. Replace the [original subject/object] [location] with a/an [new subject/object] with [clear visual attributes].
4. Convert the video into a [style name] style.
My edit request: [describe what I want]
Return one concise prompt only.Be specific about position. For removal or replacement, include foreground/background, left/right, top/bottom, clothing, object color, or another identifier. The model needs a clear target.
When removal is unreliable
The model page notes that removal can be guided by painting the unwanted object in a strong mask color, such as magenta, directly in the guide video and prompting: Remove object masked with the pink color. Use this only when text is not enough to identify the object.
7
Generate, inspect, and refine
- Upload the control video into the WAN2GP control-video field.
- Enter a single structured edit prompt. Keep it direct. Do not combine multiple edit types on the first attempt.
- Confirm the baseline settings. Check
480p,30 fps,one phase,control strength 1, and the EditAnything LoRA. - Generate one short test. Do not queue a batch until one clip preserves identity and timing.
- Review for scene preservation. The best output should keep the original person, camera movement, background, and timing intact while adding or changing only the requested element.
- Iterate on the prompt before changing many settings. Make location and attributes clearer before increasing complexity.
Optional strength tuning: Public model notes suggest starting with a distilled setup around
CFG = 1. If the edit is too weak, try increasing CFG or LoRA strength, with around 1.2 mentioned as a possible LoRA-strength experiment. Treat this as tuning, not the first move.8
Upscale and stitch the edit back into the timeline
- Send the generated segment to your upscaler. In the video, the creator sends the WAN2GP result to ComfyUI RTX Super Resolution to upscale to 4K.
- Import the upscaled segment back into your editor. Place it exactly after the cut point where the original segment was removed.
- Check the transition frame. The first AI-edited frame should align with the prior original frame as closely as possible.
- Color match if needed. The creator noticed a slight tone shift and used CapCut’s color-match adjustment to make the transition less visible.
- Export the final video. Keep the untouched sections untouched and only replace the short effect segment.
Best-case result: The viewer sees the new object/person/style appear naturally, while the original face, pose, background, and camera motion remain stable.
9
Success checks
- The edit appears only where intended.
- The original subject’s identity does not shift.
- The background does not regenerate into a different room or layout.
- The generated segment starts at the correct transition frame.
- The color, contrast, and sharpness match the surrounding footage after upscaling.
- The final clip does not reveal a resolution drop, frame-rate mismatch, or timing jump.
10
Troubleshooting
| Problem | Likely cause | Fix |
|---|---|---|
| Identity changes completely | Two phases selected, or the model is effectively regenerating too much of the clip. | Switch to one phase. Keep the control video strength at 1. Shorten the clip if needed. |
| Background shifts or room changes | Generation is not anchored tightly enough to the control video. | Use one phase, lower ambition, shorten the clip, and make the edit target more localized. |
| Edit is too weak | Prompt is vague or guidance is too low. | Use the exact add/remove/replace/convert format. Add precise visual attributes and location. Then experiment with CFG or LoRA strength. |
| Wrong object removed | Target description is ambiguous. | Add foreground/background and left/right language. Consider a magenta guide mask and prompt Remove object masked with the pink color. |
| Person/object appears random each time | The workflow is text-guided and has no reference image control. | Accept variation, run multiple seeds, or use a reference-based inpainting workflow instead. The video explicitly warns that you cannot precisely control the identity of newly added people. |
| Generation is too slow | Resolution or duration too high. | Use 480p, short clips, and upscale afterward. |
| Seam is visible after stitching | Color tone, exposure, or sharpness changed. | Use your editor’s color match/color correction. Compare the frame before and after the cut. |
11
Sources
- Prompt Mastery — “LTX 2.3: Edit Anything LoRA Works? (WAN2GP Version)”, YouTube, uploaded 2026-05-13. Guide used video metadata and auto-generated English transcript.
- EditAnything - v1.0 | LTX Video LoRA, Civitai model page. Used for public prompt patterns and inference notes.
- Alissonerdx/LTX-LoRAs, Hugging Face model repository. Used for model-file naming and experimental status.
- RunComfy — LTX 2.3 Edit Anything workflow overview. Used as a secondary explanation of prompt-driven video-to-video editing behavior.
Related local guide: LTX 2.3 AI Video Model Guide.