Two LTX-2 image-to-video workflows, same node count, same geometry, same sampler — and one of them uses a model you may not want in your pipeline. Understanding the difference is more useful than the consistency claim.
The two graphs
Both 54 nodes, both EmptyLTXVLatentVideo at 768×512 / 97 frames, both KSamplerSelect → euler, both LTXVEmptyLatentAudio at 97/25/1.
| Workflow | UNET | LoRA |
|---|---|---|
| De-distilled image-to-video | ltx-2-19b-dev-fp8_transformer_only |
ltx-2-19b-distilled-lora_resized_dynamic_fro09_avg_rank_... |
| Distilled image-to-video | ltx2-phr00tmerge-nsfw-v62 |
LTX-2-Image2Vid-Adapter at 1.0 |
The fork
The second graph's checkpoint is ltx2-phr00tmerge-nsfw-v62.safetensors — a community merge with nsfw in the name.
That matters in two ways. First, a merge changes model behaviour in ways that are not documented by a version number; v62 tells you it has been iterated a lot. Second, an NSFW-tuned merge is not the same model as the official build, and its content filters are different.
The release title says "非审查蒸馏" (uncensored distillation) explicitly, so this is disclosed, not hidden. But when you see two workflows presented as a quality comparison and one uses a differently-trained merge, the comparison is not measuring what it appears to measure.
If your pipeline needs the official model's behaviour, use the first graph. If you need the second graph's output, know that you are on a fork.
The Adapter LoRA
LTX-2-Image2Vid-Adapter.safetensors at 1.0 in the distilled graph.
An adapter LoRA adapts the model to a mode — here, image-to-video. It appears at full weight in the distilled path and is absent from the de-distilled path, which instead uses the distillation LoRA itself.
So the two graphs are not variations on one configuration. They are two different answers to "how do I get image-to-video working on a 19B LTX-2 base":
- Load the dev model and add distillation
- Load a distilled merge and add an adapter
The transformer_only checkpoint
ltx-2-19b-dev-fp8_transformer_only.safetensors — no VAE bundled. The graph loads LTX2_video_vae_bf16 and LTX2_audio_vae_bf16 separately.
Note the naming: LTX2_* (no dot) for the LTX-2 generation VAE files, versus LTX23_* for LTX 2.3. Easy to mix up when you have both families on disk.
The path-with-backslashes issue
The LoRA is referenced as LTX2\ltx-2-19b-distilled-lora_resized_dynamic_fro09_avg_rank_... — a Windows-style subdirectory path embedded in the model name.
On Linux or in a container this will not resolve. The model is expected in a LTX/v2/ subdirectory of the loras folder. If the workflow fails to find the LoRA, check whether that directory exists.
Where this breaks
- The fork. One workflow is a community NSFW merge; the other is the official dev build.
- Backslash paths in model references.
-
Two generations of VAE naming (
LTX2_*vsLTX23_*). - 97 frames at 768×512 is the unit; the title claims 60fps 1080p output, which must come from a post step not in these graphs.
-
eulerrather than an ancestral sampler — deterministic, which is right for reference-driven work but less exploratory.
Practical approach
Run the de-distilled graph. It uses the official dev model, and the distillation LoRA is a standard component. If output does not move enough, that is the distillation weight to adjust — not a reason to switch to the fork.
Workflows
Two graphs, published on RunningHub with invite code rh-v1111:
- LTX-2 de-distilled image-to-video, consistency 60fps 1080p
- LTX-2 distilled image-to-video, consistency 60fps 1080p
Files
https://huggingface.co/AI-KSK/ksk-video-173-comfyui-workflows
Notes
- Workflow files only. Model weights are not included; each keeps its own licence.
- Tutorial exports, not a product.
- Node counts and loader values 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)