DEV Community

AI-KSK
AI-KSK

Posted on

Repaint Instead of Cover: LTX 2.3 Text-Pollution Removal 1.0 with a Reference Frame

#comfyui #ai #tutorial

The difference between this and every mosaic-and-blur approach is what happens inside the mask: LTX 2.3 regenerates the region conditioned on a clean reference frame, so the area is rebuilt rather than hidden. The author's claim is a 100% win rate on the narrow case of removing gibberish text that LTX itself produced.

Evidence note: one workflow JSON ships with this episode and was parsed directly; node count, LoRA filenames and strengths, scheduler, sampler, CFG and resize size below come from that file.

When this is useful

  • Removing subtitles, logos and AI gibberish from a clip you want to keep.
  • Repairing text baked into model output on the lower-third or upper-right, where those artefacts usually land.
  • Long single takes: the author reports 480 frames, roughly 20 seconds, without obvious degradation.
  • Any job where cropping the text away would ruin the framing.

How this differs from the previous build

The 1.0 graph is narrower than its predecessors and better for it.

  • It uses the R2V (reference-to-video) 3000-step LoRA rather than the Edit Anything 9000-step build or the older 10000-step masked text-to-video path. The reference frame is the whole point.
  • Concatenation was dropped. Earlier LTX repair graphs stitched the reference in; this one relies on the IC-LoRA working without it, the same reasoning the author applied to the LTX 2 detail LoRA.
  • The model is GGUF-quantised. The graph loads through UnetLoaderGGUF with a GGUF dual CLIP loader, which matches the Q4_K_M quantisation the author describes.
  • The mask is static. One painted region, applied for the whole clip, which is fine because subtitles and watermarks sit in fixed places.

Running it

  1. Load the clip. Most examples are LTX 2.3 output that already carries text artefacts.
  2. Set the frames to process — ten seconds is the tested unit.
  3. Keep the frame rate at 24.
  4. Supply a first-frame reference image with no text in it. If your first frame is already clean, use it directly; if not, repaint the first frame first.
  5. Open the mask editor and paint over where the text appears. This is a static mask.
  6. Run. Keep the positive prompt as shipped; the author pre-tuned it for a high hit rate and says only the worst ~10% of results need an LLM-adjusted prompt.

Parameters, and where the numbers come from

Item Value in the shipped graph
Node count 66
Model loader UnetLoaderGGUF (GGUF dual CLIP loader alongside)
LoRA, repair ltx23_inpaint_masked_r2v_rank32_v1_3000steps.safetensors at 1.0
LoRA, edit ltx23_edit_anything_global_rank128_v1_9000steps_adamw.safetensors at 1.0
Scheduler bong_tangent, steps 8, denoise 1
Sampler euler_ancestral_cfg_pp
CFG 1
Audio latent 24 fps
Resize 512x512

The 512x512 resize applies to the reference and mask path, not to the delivered resolution.

Honest limits

  • The 100% win rate applies to one narrow case: text that LTX 2.3 itself generated. It is not a general watermark-removal rate.
  • The mask is static, so drifting or animated watermarks need a better mask tool than the one built in.
  • If the first frame is itself obscured, you have to repaint it before this graph can use it as a reference.
  • Quantising the model to Q4_K_M trades some quality for duration headroom — the author's own framing.
  • Without a reference image the repaired region is visibly different from its surroundings.

Episode resources

RunningHub workflows

Top comments (0)