LTX 2.3.1: What Actually Changed Between the Distilled and Distilled-1.1 Checkpoints
A quiet checkpoint update is easy to ignore. This release tests whether the LTX 2.3 1.1 distilled build is meaningfully better than the plain distilled build, and the file names tell you most of what you need to know bef
A quiet checkpoint update is easy to ignore. This release tests whether the LTX 2.3 1.1 distilled build is meaningfully better than the plain distilled build, and the file names tell you most of what you need to know before you render anything.
The comparison that matters
Compare the checkpoints across two consecutive releases in the same family:
| Older workflows | 1.1 workflows |
|---|---|
ltx-2.3-22b-distilled.safetensors |
ltx-2.3-22b-distilled-1.1.safetensors |
Same 22B distilled lineage. The 1.1 suffix is a revision of the distilled build, not a new model family. That distinction matters: everything else in the graph — the LoRAs, the latent geometry, the sampler — carries over unchanged, which is why the upgrade is a filename swap rather than a rework.
The three graphs in this release
| Workflow | Nodes | Likeness LoRA |
|---|---|---|
| 1.1 + VBVR multi-image first/last frame | 78 |
Ltx2.3-Licon-VBVR-I2V-96000-R32 at 1.0 |
| 1.1 + VBVR digital human | 71 |
Ltx2.3-Licon-VBVR-I2V-96000-R32 at 1.0 |
| 1.1 + VBVR digital human (variant) | 15 |
ltx-2.3-22b-distilled-lora-384 at 0.5
|
The two large graphs share the same likeness LoRA at full weight. The small 15-node graph is a different configuration entirely — it uses the distillation LoRA at 0.5 rather than the likeness LoRA.
That split is a useful pattern to internalise: 78/71-node graphs are reference-driven (likeness LoRA at 1.0), 15-node graphs are prompt-driven (distillation LoRA at 0.5). The node count tells you which kind of graph you are looking at.
Why 0.5 distillation appears again
The 15-node graph loads ltx-2.3-22b-distilled-lora-384 at 0.5, matching what the previous release did.
The reasoning is consistent across releases: full-weight distillation makes the model conservative, and conservative means static. Half weight restores motion at the cost of step efficiency. This is not a one-off tuning choice — it is a pattern in how the author works.
If you see a 15-node LTX graph with the distillation LoRA at anything other than 1.0, that is probably deliberate.
The audio VAE is still mismatched
Both large graphs:
-
CheckpointLoaderSimple→ltx-2.3-22b-distilled-1.1.safetensors -
LTXVAudioVAELoader→ltx-2.3-22b-dev.safetensors
Distilled video checkpoint, dev (non-distilled) audio VAE. This mismatch persists release after release in this family.
It works — the audio is generated, muxed, plays. But the audio path is running different weights than the video path, and nothing in ComfyUI validates that. If you are chasing audio quality problems, aligning these two is the first experiment to run, and it costs one filename change.
Where this release is weak
- The 1.1 claim is qualitative in the video, not measured. The workflows themselves carry no benchmark — you get whatever improvement the revision provides, with no way to quantify it from the files alone.
- Audio VAE mismatch across every large graph.
- 78 vs 71 nodes for multi-image versus single-image digital human is a smaller gap than you would expect from the capability difference; most of the shared machinery is identical, which means switching between them is cheap but also means a mistake in the shared part breaks both.
- The 15-node variant is undocumented. It is present but the release does not say what it is for.
Practical migration
If you already have working LTX 2.3 workflows on the plain distilled checkpoint, migrating to 1.1 is a single CheckpointLoaderSimple filename change. Do that first, re-render one clip, and compare. If the difference does not justify a re-render of your back catalogue, keep the old renders.
Workflows
Three graphs, published on RunningHub with invite code rh-v1111:
- LTX 2.3 1.1 + VBVR multi-image first/last frame
- LTX 2.3 1.1 + VBVR digital human
- LTX 2.3 1.1 + VBVR digital human (15-node variant)
Files
https://huggingface.co/AI-KSK/ksk-video-223-comfyui-workflows
Notes
- Workflow files only. Model weights are not included; each keeps its own licence.
- Tutorial exports, not a product.
- Node counts, LoRA values and checkpoint names quoted above are read from the published JSON files.
- Independent AI-KSK release. Not affiliated with Lightricks or RunningHub.
Originally published by Dev.to AI. Aggregated on AIWithGhost for educational purposes — full credit and traffic to the original publisher.