Making a music video with a video model runs into a hard wall: the model generates roughly 97 frames at a time, and a song is three minutes. The gap between those two numbers is the entire problem.
This workflow closes it with a for-loop: generate segments in sequence, carry state across iterations, and stitch the result into a continuous piece.
What the graph loads
| Node | Value |
|---|---|
CheckpointLoaderSimple |
ltx-2.3-22b-dev-dare-ties-distilled-1.1.safetensors |
LTXVAudioVAELoader |
ltx-2.3-22b-dev-dare-ties-distilled-1.1.safetensors |
LoraLoaderModelOnly |
ltx-2.3-22b-ic-lora-motion-track-control-ref0.5.safetensors at 1.0
|
EmptyLTXVLatentVideo |
320×288, 97 frames |
LTXVEmptyLatentAudio |
97, 25, 1 |
KSamplerSelect |
euler_ancestral_cfg_pp |
The motion-track-control LoRA at weight 1.0 is the continuity mechanism. Note that in other workflows the same LoRA family is used at 0.5; here it is at full weight because a loop needs stronger carry-forward than a single shot does.
The 97-frame unit appears in both the video and audio latent. That is not decoration — it means the segment boundary falls at a known audio timestamp, which is what makes the loop legal.
The loop is the design
A for-loop over segments has to answer four questions, and the graph's node count (133) is mostly those answers:
- Where does segment N start? The tail of segment N-1 supplies the anchor. Without it, each segment is an independent short clip and the cut is visible.
- How does audio stay aligned? The audio latent is segmented on the same 97-frame grid, so segment N's audio is the next slice of the track rather than a re-generation.
- How does the digital human stay consistent? Audio-driven mouth movement plus the motion-track LoRA. Identity is carried by the anchor, motion by the audio.
- How do the pieces join? Frames are concatenated, but the seam is only clean if consecutive segments share enough conditioning that the model's idea of the scene did not reset.
That fourth point is where for-loop pipelines live or die. A loop with a weak anchor produces a series of clips that each look right and cut wrong.
The honest limitations
- Drift compounds. Each segment inherits the previous one's errors. Over ten segments, small identity drift becomes visible.
- Audio must be decided first. Because the audio is sliced by frame count, changing the song means re-running the whole loop. There is no partial re-render.
- The seam is still a seam. Even with strong continuity, a segment boundary is a place where a viewer can sense a reset. Picking boundaries that land on a beat masks it — which is a musical decision, not an engineering one.
-
Output is an image sequence.
SaveImagerather thanSaveVideomeans assembly is a separate step; the graph produces frames, not a file.
What this changes about music-video production
Normally the visual is cut to the music. Here the music drives the generation, and the loop follows it. The result has a quality that cut-to-music does not: motion that was generated against the audio keeps a relationship to it that survives the edit.
That is checkable and worth checking: if the motion still reads as musical with the audio muted, the conditioning worked. If it only feels right with sound, you have cut-to-music wearing a reactive pipeline's clothes.
Workflows
One graph, published on RunningHub with invite code rh-v1111: the For-loop LTX 2.3 long MV auto-generation workflow.
Files
The ComfyUI workflow export is in the companion repository:
https://huggingface.co/AI-KSK/ksk-video-242-comfyui-workflows
Video walkthrough (Chinese): https://youtu.be/eqTjKOURnF0
Bilibili: https://www.bilibili.com/video/BV1CSLJ6vEMQ/
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 count 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.
Optional API platform - 88API (affiliate link)
Explore 88API for access to multiple AI model APIs. Available models, routes and prices are listed on the platform; Alipay top-ups are supported.
https://88api.ai/sign-up?aff=usx3
Disclosure: I may earn a commission from eligible purchases after you register through this referral link.
Top comments (0)