Comparing two model generations is usually done across releases, with different prompts, different graphs and different settings. This release does it properly: the same three-stage digital-human graph, built twice, once per model generation.
The two graphs
| Workflow | Nodes | Checkpoint | LoRA |
|---|---|---|---|
| LTX 2.3 three-stage digital human | 64 | ltx-2.3-22b-dev.safetensors |
ltx-2.3-22b-distilled-lora-384 at 0.5 |
| LTX 2 three-stage digital human | 69 | ltx-2-19b-dev.safetensors |
ltx-2-19b-lora-camera-control-static at 0.5 |
Both use EmptyLTXVLatentVideo at 320×288 / 97 frames, LTXVEmptyLatentAudio at 97/25/1, and KSamplerSelect → euler_ancestral_cfg_pp.
The shared structure is the point. Same graph shape, same latent geometry, same sampler. The only variables are the model generation and its matching LoRA.
The generational difference is 19B versus 22B
-
LTX 2:
ltx-2-19b-dev.safetensors— 19B parameters -
LTX 2.3:
ltx-2.3-22b-dev.safetensors— 22B parameters
Both are the dev (non-distilled) builds, and both pair with a distillation LoRA at 0.5.
That consistency is what makes the comparison valid. If one used the distilled checkpoint and the other the dev build, a difference in output could be attributable to distillation rather than to the model generation.
Note the LoRA's purpose changed
This is easy to miss and it matters:
- LTX 2.3 graph:
ltx-2.3-22b-distilled-lora-384— a distillation LoRA - LTX 2 graph:
ltx-2-19b-lora-camera-control-static— a camera control LoRA, at 0.5
Two different LoRA families at the same weight. The comparison is therefore not purely "which model is better" — the LTX 2 graph is also applying camera control that the 2.3 graph is not.
If you want a clean generation comparison, you would need to remove the camera control LoRA from the LTX 2 graph or add an equivalent to the 2.3 graph. As published, this is a comparison of two production configurations, not two bare models.
That is a legitimate thing to compare — it just is not the same question as "which model is better".
Audio VAE consistency
Both graphs deserve credit here:
- LTX 2.3: checkpoint
ltx-2.3-22b-dev/ audio VAEltx-2.3-22b-dev— match - LTX 2: checkpoint
ltx-2-19b-dev/ audio VAEltx-2-19b-dev— match
No distilled/dev mismatch. This author consistently pairs them correctly, which is worth knowing if you are using their workflows as a reference — the mismatches you see elsewhere in this family are not universal.
Where this comparison is weak
- Different LoRA families (distillation vs camera control) mean the comparison conflates two variables.
- Node count differs by five (64 vs 69), so the graphs are not structurally identical either.
- One prompt set. Conclusions are specific to the prompts used.
- Camera control on one side only means motion will differ for reasons unrelated to model quality.
What to actually take from it
The useful output is a sense of how much the 19B-to-22B jump changes things, holding the graph shape constant. If the difference is large, moving to 2.3 is worth the disk space. If it is marginal, the older model is cheaper to run and its LoRAs are already built.
To answer that properly, strip the camera control LoRA from the LTX 2 graph and re-render. That is one node change and it turns an interesting comparison into a clean one.
Workflows
Two graphs, published on RunningHub with invite code rh-v1111:
- LTX 2.3 three-stage "1+3" digital human, full performance
- LTX 2 three-stage "1+3" digital human, full performance
Files
https://huggingface.co/AI-KSK/ksk-video-207-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 (1)
A same-graph comparison is strongest when the seed, sampler settings, precision mode, and hardware are held constant alongside the visual result. I’d also include peak VRAM, wall-clock latency, and failed-run rate, because a model upgrade that improves a few showcase frames but changes operational reliability needs a different adoption decision.