Repainting Text Pollution in Video: Why Regeneration Beats Masking Subtitles

2026-11-17来源 dev.toEnglish

Removing subtitles from video has historically meant two bad options: blur or mosaic over the text, or crop it out. Both leave evidence. A third option exists — regenerate the pixels underneath — and it produces a fundamentally different result.

This workflow is built for that third option, and it handles more than subtitles: watermarks, garbled text, and any obstruction that sits on top of the picture.

What the graph loads

Node Value
VAELoaderKJ LTX23_video_vae_bf16.safetensors, main_device, bf16
DualCLIPLoaderGGUF gemma_3_12B_it_fp8_e4m3fn.safetensors + ltx-2.3_text_projection_bf16.safetensors
LoraLoaderModelOnly ltx23_edit_anything_global_rank128_v1_9000steps_adamw.safetensors at 1.0
LTXVEmptyLatentAudio 97, 24, 1
EmptyAudio 60, 44100, 2
EmptyImage 512, 512, 1, 16711935
KSamplerSelect euler_ancestral_cfg_pp

Four details are worth unpacking.

ltx23_edit_anything_global_rank128 — a rank-128 Edit Anything LoRA trained for 9000 steps with AdamW. This is the component doing the actual work: it conditions the model to regenerate a region according to a reference rather than to invent freely.

The audio latent frame count is 24, not 25. Every other workflow in this LTX 2.3 family uses [97, 25, 1]. This one uses [97, 24, 1]. If you are porting nodes between workflows, that is a difference that will not raise an error and will produce subtly wrong audio timing.

EmptyAudio at 60 frames, 44100 Hz, stereo. This graph produces an audio latent that is not derived from the video — it is a placeholder to satisfy the AV sampler. That means the audio is effectively silent, and the graph is doing video-only work through a pipeline designed for audio and video together.

EmptyImage with colour 16711935 — that is 0x00FF00FF, magenta. A magenta placeholder is the classic "this region is undefined" marker. It also means if the edit fails to cover a region, you will see magenta rather than a plausible-looking wrong result — which is a better failure mode than silent corruption.

Why regeneration beats masking

Blur and mosaic destroy information and replace it with a visual lie. The viewer sees a rectangle that does not match its surroundings, and the eye goes straight to it.

Regenerating means the model produces pixels consistent with the surrounding scene — lighting direction, grain, texture, perspective. If it works, there is no artifact to notice.

The catch is that regeneration is a guess. On a face, a guess can be wrong in a way blur never is. So the technique is strongest on:

  • Flat or textured backgrounds behind subtitles
  • Watermarks over sky, walls, foliage
  • Garbled text over midground scenery

And weakest on:

  • Text over a face
  • Text over a hand holding something
  • Text over complex small detail

Knowing which side of that line you are on determines whether this workflow saves you time or creates a subtle problem.

The reference requirement

The (参考) in the workflow name is meaningful — this is a reference-conditioned repair. You are not just saying "remove the text"; you are supplying what should be there instead. For a plain background that reference can come from the surrounding frames. For structured content it has to come from somewhere real.

Where this breaks

  1. Audio frame count divergence. 24 vs 25 is a real inconsistency in this family.
  2. Rank-128 LoRA at full weight. Localised edits can be aggressive — the model may regenerate more than you asked. Lower weight narrows the region.
  3. The magenta placeholder. Useful as a failure signal, but it means any uncovered region is glaring. Check the output rather than trusting a clean render.
  4. Only 66 nodes. That is small for a repair graph, which means less internal validation. The graph trusts you to supply a sane reference.

Practical use

Extract the frames, identify the text region, supply a reference for what belongs there, and run the graph. Compare the repaired output against the original at 4× on the region — a repair that looks clean at 1× is often soft at 4×, and softness in a single region is more noticeable than uniform softness.

Workflows

One graph, published on RunningHub with invite code rh-v1111: the LTX 2.3 text-pollution repair workflow (reference-conditioned).

Files

The ComfyUI workflow export is in the companion repository:

https://huggingface.co/AI-KSK/ksk-video-232-comfyui-workflows

Video walkthrough (Chinese): see episode description
Bilibili: see episode description

Notes

  • Workflow file only. Model weights are not included; each model keeps its own licence.
  • This is a tutorial export, not a packaged product.
  • Node counts and widget values quoted above are read from the published JSON file.
  • Independent AI-KSK release. Not affiliated with or endorsed by Lightricks or RunningHub.

原文:https://www.a8q9.cn/en/article/repainting-text-pollution-in-video-why-regeneration-beats-masking-subtitles-bff902.html

相关文章