Seven workflows for one job — image-plus-audio-driven digital human — each slightly different. That is not redundancy, it is a search for the sweet spot, and the range of settings between them is the useful output.
The seven graphs
| Workflow | Nodes | Geometry | Key setting |
|---|---|---|---|
| Image-to-video single-stage dual-track | 42 | 736×1280, 121 | Adapter LoRA at 1.0 |
| Image-to-video distilled two-stage upscale | 40 | 544×960, 121 | distil. LoRA at 0.5 |
| Text-to-video single-stage dual-track | 42 | 1920×1088, 121 | Adapter LoRA at 1.0 |
| Text-to-video distilled two-stage upscale | 40 | 960×544, 121 | Adapter LoRA at 1.0 |
| Image-to-video single-stage optimal | 32 | 736×1280, 121 | distil. LoRA at 0.2 |
| Image-to-video single-stage distilled optimal | 30 | 736×1280, 121 | Adapter LoRA at 1.0 |
| Text-to-video one-stage non-accelerated + two-stage upscale | 43 | — | — |
The 0.2 weight is the interesting one
ltx-2.3-22b-distilled-lora-384.safetensors at 0.2 — the lowest distillation weight covered anywhere in this series.
Elsewhere it appears at 0.5 (prompt-driven), 0.6 and 0.7 (larger geometries), and 1.0 (conservative reference-driven). At 0.2 you are barely distilling at all, which means:
- You pay close to full step cost
- You get almost all of the dev model's own motion judgement
For a digital human that is the right direction. The face and mouth movement are the hard part, and a heavily distilled model produces the most average mouth shapes. At 0.2 the model is willing to commit to more specific articulation.
The cost is speed, and the graph is named "optimal" rather than "fast", which is a fair label.
The Adapter LoRA
LTX-2-Image2Vid-Adapter.safetensors at weight 1.0 appears in five of the seven graphs.
An "adapter" LoRA is a different mechanism from a likeness LoRA: it adapts the model to a mode of operation (image-to-video) rather than carrying a specific identity. At full weight it is claiming to be necessary for the mode, not decorative.
Note the graph named 图生视频单阶段蒸馏最优 — "distilled optimal" — uses the Adapter LoRA, not the distillation LoRA. The naming is inconsistent with its sibling, which uses the distillation LoRA at 0.2. Worth reading the loaders rather than trusting the titles.
The geometries
Three distinct sizes across seven graphs:
- 736 × 1280 — portrait, ~0.94 MP. Roughly 9:16. This is the one to use for vertical output.
- 544 × 960 — portrait, ~0.52 MP. Lower res, used with the two-stage upscale path.
- 1920 × 1088 — landscape, ~2.09 MP. Full HD, used only in the text-to-video dual-track graph.
- 960 × 544 — landscape, ~0.52 MP.
121 frames in every graph — consistent.
That consistency matters: 121 frames at 24 fps is about five seconds, and every graph here is built for that unit. Nothing needs a sigma schedule port.
The dual-track structure
"单阶段双轨" (single-stage dual-track) in the names refers to the audio path running in parallel with the video path rather than sequentially. That is a structural difference from the audio-then-video chaining seen elsewhere.
The practical difference: dual-track lets the model resolve picture and sound together; chained approaches let you iterate on one without re-rendering the other.
Where this release is weak
- Seven graphs with overlapping names — "optimal" appears three times, and the loaders do not match the titles.
- The 0.2 weight is only in one graph. No comparison at 0.2 versus 0.5 is provided, so you have to run both to know what it costs.
- Only one graph is 1920×1088, so full-HD output is limited to text-to-video.
-
Both VAE loaders point at
LTX23_*_vae_bf16files while the UNET istransformer_fp8— so the VAE is bf16 and the transformer is fp8. That is a mixed-precision setup, which is fine but means the "fp8" in the name does not describe the whole pipeline.
Practical approach
Start with the 32-node graph (distillation 0.2). It is the smallest, and it exposes the lowest-distillation configuration. If it is too slow, move to the 30-node graph (Adapter LoRA) which should be faster at the cost of adherence. Only then try the two-stage upscale paths.
For vertical output use the 736×1280 geometry; for landscape use 1920×1088.
Workflows
Seven graphs, published on RunningHub with invite code rh-v1111:
- LTX 2.3 image-to-video single-stage dual-track
- LTX 2.3 image-to-video distilled two-stage upscale
- LTX 2.3 text-to-video single-stage dual-track
- LTX 2.3 text-to-video distilled two-stage upscale
- LTX 2.3 image-to-video single-stage optimal
- LTX 2.3 image-to-video single-stage distilled optimal
- LTX 2.3 text-to-video one-stage non-accelerated + two-stage upscale
Files
https://huggingface.co/AI-KSK/ksk-video-198-comfyui-workflows
Notes
- Workflow files only. Model weights are not included; each keeps its own licence.
- Tutorial exports, not a product.
- Node counts, geometries and weights quoted above are read from the published JSON files.
- 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 (0)