DEV Community

AI-KSK
AI-KSK

Posted on Edited on Originally published at huggingface.co

A Transition LoRA That Stops First/Last-Frame Video From Snapping

First/last-frame video has an obvious failure: the model hits the first frame, hits the last frame, and gets from one to the other with a visible snap. The subject does not move, it jumps. A transition LoRA exists to fix that, and this workflow is built around it.

The graph

Node Value
CheckpointLoaderSimple ltx-2.3-22b-dev.safetensors
LTXVAudioVAELoader ltx-2.3-22b-dev.safetensors
LoraLoaderModelOnly ltx-2.3-22b-distilled-lora-384.safetensors at 0.5
EmptyLTXVLatentVideo 320×288, 97 frames
LTXVEmptyLatentAudio 97, 25, 1
KSamplerSelect euler_ancestral_cfg_pp

81 nodes total.

The checkpoint pairing is finally consistent

Check the two loaders:

  • Checkpoint: ltx-2.3-22b-dev.safetensors
  • Audio VAE: ltx-2.3-22b-dev.safetensors

The same file. If you have been following the LTX 2.3 workflows covered in adjacent releases, you will know how unusual that is — most of them pair a distilled checkpoint with a dev audio VAE, loading two different weight sets in one graph.

This workflow does it right. Whatever the reason, it means the audio and video paths here are running coherent weights, and that is the configuration to use as a baseline when comparing against the mismatched ones.

The dev checkpoint plus a 0.5 distillation LoRA

There is a second difference worth noting:

Most workflows load the distilled checkpoint directly. This one loads the dev (non-distilled) checkpoint and adds the distillation LoRA at 0.5.

Those are two different ways to arrive at "fast sampling":

  • Distilled checkpoint — the distillation is baked in, you get the distilled behaviour at full strength
  • Dev checkpoint + distillation LoRA at 0.5 — you get half the distillation, mixed with the dev model's own behaviour

The second gives the model more of its own judgement back. For a transition task that matters: a fully distilled model produces the most probable interpolation, which is often a straight line or a snap. A partially distilled model is more willing to take an interesting path between the two frames.

That is the mechanism behind "the transition no longer looks stiff", and it is a more satisfying explanation than "the LoRA fixes transitions".

What 81 nodes are doing

For a first/last-frame graph, 81 nodes is substantial. The bulk is the multi-image conditioning: multiple images entering, being assigned to slots, and being combined with the temporal constraint that one is first and one is last.

The prompt rules referenced in the release title live in this structure — the way you describe the transition has to name what changes and what holds, and the graph's conditioning reflects that.

Where this breaks

  1. Two frames is a weak constraint. The model knows the endpoints and nothing about the path. If the two frames are very different (different lighting, different angle), the model has to invent the whole transition.
  2. 0.5 distillation is a trade. You pay in step count. At very low budgets this graph will be worse than a fully distilled one.
  3. 320×288 base. The upscale path is doing real work; the base alone is not deliverable.
  4. No audio driving. The audio latent is structured but the graph is video-focused — [97, 25, 1] exists to satisfy the AV sampler, not because the output has meaningful audio.

Practical approach

Pick two frames that are the same shot with a change in pose or object, not a change in scene. The LoRA helps the model interpolate; it cannot invent a plausible path between two frames that do not belong together. Once the transition reads cleanly on an easy pair, try a harder one.

Workflows

One graph, published on RunningHub with invite code rh-v1111: LTX 2.3 multi-image first/last frame high-consistency smooth transition.

Files

https://huggingface.co/AI-KSK/ksk-video-215-comfyui-workflows

Notes

  • Workflow file only. Model weights are not included; each keeps its own licence.
  • Tutorial export, not a product.
  • Node count and loader values quoted above are read from the published JSON file.
  • Independent AI-KSK release. Not affiliated with 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 (1)

Collapse
 
alexshev profile image
Alex Shev •

A transition LoRA should be evaluated on the middle frames, not only the endpoints. I’d test motion continuity, subject identity, camera path, and editability under several prompt changes, with a few deliberately incompatible first and last frames. That distinguishes a smoother interpolation from a model that merely hides the cut by losing the scene’s constraints.