LTX 2.3 Suite 2.5: Getting Clean Plates Out of the Model Instead of Cleaning Up After It
LTX 2.3 Suite 2.5: Getting Clean Plates Out of the Model Instead of Cleaning Up After It Subtitles burned into generated video are a problem with no good downstream fix. OCR-based removal guesses at the glyphs; masking
LTX 2.3 Suite 2.5: Getting Clean Plates Out of the Model Instead of Cleaning Up After It
Subtitles burned into generated video are a problem with no good downstream fix. OCR-based removal guesses at the glyphs; masking leaves a soft rectangle where the text used to be. The 2.5 suite takes a different route: it makes the model edit the text out during generation, so the plate is clean at the source.
That change is the headline of this release, and it is worth understanding what it costs structurally.
What "IC Edit" actually does in these graphs
Every one of the five main workflows in this suite carries the same editing conditioning path. The mechanism is a paired LoRA stack rather than a post-process:
| Workflow | Nodes | Links |
|---|---|---|
| Single digital human | 109 | 172 |
| Dual digital human | 108 | 169 |
| Image-to-video | 273 | 438 |
| Multi-image reference | 121 | 213 |
| Text-to-video | 90 | 143 |
The image-to-video graph being roughly three times the size of text-to-video is the interesting part. Reference conditioning is not a single node — it is a large plumbing network, and that is the honest cost of "give the model a picture and let it make video from it".
The model stack
Across the suite:
-
ltx-2.3-22b-dev-dare-ties-distilled-1.1.safetensors— the AV checkpoint, loaded into both the video VAE path and theLTXVAudioVAELoader. The audio half is not an afterthought in these graphs; it gets its own VAE. -
ltx-2.3-spatial-upscaler-x2-1.1.safetensors— latent 2× upscaler. Note that the dual-character graph is pinned to 1.0 while everything else uses 1.1. That is not accidental — mixing versions in a multi-character render changes how the two subjects resolve at the upscale boundary. -
Ltx2.3-Licon-VBVR-I2V-390K-R32.safetensors— the likeness LoRA, at the 390K checkpoint rather than the older 240K. This is the file that carries identity through the render. -
ltx-2.3-22b-ic-lora-motion-track-control-ref0.5.safetensors— motion track control, weight 0.5, used in the single digital-human graph. -
LTX2.3-IC-LORA-Dual-Character.safetensors— the dual-character LoRA at weight 0.75. Lower than 1.0 deliberately: at full weight the two identities start competing for the same facial structure. -
MelBandRoFormer_comfy/MelBandRoformer_fp16.safetensors— vocal separation. This is what makes clean speech possible: the voice is isolated before it becomes an audio condition, so the instrumental bed cannot bleed into the mouth timing.
Sampler choice tells you which graph you are in
Two samplers appear:
-
euler_ancestral_cfg_pp— used in all five main generation graphs. -
euler_ancestral— used in the two IC Edit utility graphs and the image-to-video graph.
The cfg_pp variant matters when guidance actually matters, which is generation from a prompt. For pure editing passes the prompt is secondary to the reference, so the plainer sampler is the right call — and it is measurably cheaper per step.
Latent geometry: the numbers that repeat everywhere
Every graph starts from the same latent shape:
-
EmptyLTXVLatentVideoat 320×288, 97 frames, batch 1 -
LTXVEmptyLatentAudioat 97, 25, 1
97 frames is the working unit. Everything else — upscaling, editing, extension — works outward from that. If you are planning a workflow of your own, treat 97 frames at 320×288 as the base unit and let the spatial upscaler handle resolution, because that is what the suite itself does.
The two IC Edit utilities
These are the smallest graphs in the set and the ones you will reach for most:
IC Edit 去水印字幕 (35 nodes) — removes burned-in subtitles and watermarks. Because it runs inside the model, the pixels under the text are regenerated rather than blurred over.
IC Edit 高清修复 (37 nodes) — repair pass. Useful when a generated clip is slightly soft but compositionally correct; it re-renders detail without re-deciding the scene.
Both run euler_ancestral and keep the same AV latent structure. Neither needs a prompt.
The Relay prompt director
The odd one out: a 6-node graph that produces a still image for prompt-director use. It is a helper, not a video workflow, and it exists because writing the dual-character prompt format by hand is error-prone.
Where this suite is strong, and where it is not
Strong:
- Any pipeline where burned-in text is a problem. Clean at the source beats cleanup.
- Dual-character dialogue, because the vocal separation and the 0.75-weight dual LoRA are both tuned for it.
- Identity hold across a 97-frame take at the 390K likeness checkpoint.
Not strong:
- Anything past the single 97-frame unit without an explicit extension step. The suite does not hide that — it is why the image-to-video graph is so large.
- Fast iteration. A 273-node graph at this resolution is not a quick loop; use the text-to-video graph (90 nodes) as the sandbox and move to image-to-video only when the shot matters.
Troubleshooting notes
-
Subtitles survive the IC Edit pass — check that the editing LoRA is actually loaded in the active model path. A missing
LoraLoaderModelOnlyon the model branch is silent: the graph runs and produces video, just without the edit. -
Two characters converge on the same face — drop
LTX2.3-IC-LORA-Dual-Characterbelow 0.75. The identity competition is a weight problem, not a prompt problem. - Seam visible between subjects after upscale — check the upscaler version. The dual-character graph expects 1.0; substituting 1.1 changes the boundary behaviour.
- Speech timing drifts against mouth movement — the vocal separation path has to run before the audio VAE encode. Skipping MelBandRoFormer and encoding the mixed track directly produces an audio latent whose timing does not match what the graph conditioned on.
- OOM — 97 frames at 320×288 is the base unit for a reason. Reduce frame count before reducing resolution; the graph's upscale path assumes a fixed latent shape.
Workflows
Eight graphs ship with this release. All are published on RunningHub with invite code rh-v1111:
- Suite 2.5 — single digital human, IC Edit, subtitle-free speech
- Suite 2.5 — text-to-video, IC Edit, subtitle-free generation
- Suite 2.5 — dual digital human, IC Edit, subtitle-free dialogue
- Suite 2.5 — image-to-video, IC Edit, subtitle-free generation
- Suite 2.5 — multi-image reference, IC Edit, subtitle-free generation
- IC Edit HD repair
- IC Edit watermark and subtitle removal
- Relay story + dual-character LoRA prompt director
Files
The ComfyUI workflow exports are in the companion repository:
https://huggingface.co/AI-KSK/ksk-video-239-comfyui-workflows
Video walkthrough (Chinese): https://youtu.be/ESU7isCNZww
Bilibili: https://www.bilibili.com/video/BV1za5y6FE7r/
Notes
- Workflow files only. Model weights are not included; each model keeps its own licence.
- These are tutorial exports, not a packaged product. Node parameters match the walkthrough.
- Node and link counts quoted above are measured from the published JSON files, not estimated.
- Where a behaviour is described as a consequence of graph structure, that is analysis — measured throughput numbers are not claimed here.
- Independent AI-KSK release. Not affiliated with or endorsed by Lightricks or RunningHub.
Originally published by Dev.to AI. Aggregated on AIWithGhost for educational purposes — full credit and traffic to the original publisher.