LTX 2.3 的兩種影片後期修復流程:清除浮水印與字幕雜字,再擴展至電影比例
Two Rescue Workflows: Erase Watermarks and Subtitle Garbage, Then Outpaint to Cinema Ratio
兩種以 LTX 2.3 為基礎的影片後期流程,分別用於重繪浮水印、字幕及亂碼文字,以及向外擴展畫面以調整比例。作者估計清理流程成功率約九成,但高動態畫面效果不佳,建議片段約十秒,超過 15 秒效果會迅速下降。本集沒有提供流程 JSON 或流程連結。
Two workflows aimed at the last mile of post, not at generating anything new: one repaints over watermarks, subtitles and gibberish text; the other extends the frame in any direction so a 3:4 clip can become 21:9. Both lean on LTX 2.3, and both are imaginative reconstruction rather than cloning.
Evidence note: this episode ships no workflow JSON and no workflow or mirror links — 说明.txt lists neither. All settings below are the author's spoken values, not values read out of a graph.
When this is useful
- Cleaning a clip exported from a tool whose watermark you cannot remove by subscription.
- Deleting subtitles and AI gibberish from a take you would otherwise shoot again.
- Reframing a vertical or square clip to 16:9, 21:9 or 9:16 without letterboxing.
- Cropping away something unfixable and extending the frame back to the target ratio instead.
How this differs from the previous build
Both workflows are derivations, and the author is explicit about the ancestry.
- The cleanup graph extends the earlier dual-dialogue repair flow (episode 230) and hard-codes a positive prompt that has worked well, so the prompt usually needs no editing.
- The outpainting graph is close to the official IC-LoRA example, with a prompt the author insists you should actually write.
- The mask is dynamic here, not a single static rectangle: the author samples the mask over time, which is why a drifting watermark can be handled.
Running it
Watermark / subtitle removal
- Load the clip and lock the frame rate at 24.
- Set the frames to process — 121 in the demo.
- Set the longest-edge length for the processed video: 1800 in the demo, with about 2000 covering most material.
- Supply a reference image only if you want the IF branch on. Toggle the IF node to
falseto run pure R2V imaginative repair with no reference; toggletrueand a reference frame becomes mandatory. - Paint the mask in the mask editor. For output from tools like the one demonstrated, expect three regions: top-left, upper-third left, lower-third right.
- Run. The first pass gives a preview with the masked areas visible in pink.
Outpainting
- Load the clip, keep the frame rate at 24 (the author calls this the LTX 2.3 default recommendation), and pick the frame range — 121 frames starting 100 frames in, in the demo.
- Choose a mode: extend by target megapixels and ratio (for example 2M pixels at 3:2, 16:9 or 9:16), or use the KJ expand node to push up, down, left and right with colour
000. - Write the prompt. If you bypass the image-to-video branch, your reference image does not act, and the prompt carries the whole scene.
- Keep the image-to-video branch bypassed unless you specifically want the reference to constrain the result.
Honest limits
- The author puts the cleanup workflow at about a nine-in-ten success rate — a self-estimate, not a measured rate.
- Reference-guided repaint works on low-motion footage; the author says high-motion scenes such as sport will not come out well.
- Keep segments around ten seconds. The author reports LTX 2.3 degrades quickly past 15 seconds, so longer clips are worse, not more efficient.
- Outpainting is imaginative: seams disappear because the model invents content, not because it reconstructs truth.
- No workflow links and no mirror link are recorded in 说明.txt for this episode — there is nothing to click through to.
Episode resources
- Video tutorial (YouTube): https://www.youtube.com/watch?v=uz3ha3fDxCo
- Chinese original (Bilibili): https://www.bilibili.com/video/BV1Qa5s6tESM
來源:DEV Community · dev.to