π LTX Reference Probe
Check your LTX2 reference wiring before you waste a generation
- model
- model
- report
LTX2 generations aren't free - a single video clip can be minutes of GPU time. So when the reference-token family isn't doing anything and you can't tell why, the answer is LTX Reference Probe, the debug node that tells you whether the wiring actually took, before you run a sampler.
The reference system is two hops: LTX Reference Enable patches the model class, then LTX Reference Conditioning attaches the actual reference latent. There are three separate places that state can live, and a lot of "why is my reference being ignored" moments are really "the reference never made it into the model." Probe walks all of them and reports what it finds.
What it reports
Wire a MODEL in and you get two outputs: the model unchanged (passthrough, so you can leave it inline) and a report STRING. It also prints the same report to the console. It checks:
- Patch status - whether
LTXAVModel._process_inputand_prepare_timestepare actually the patched versions (i.e. Enable ran) or the stock originals. - Patchifier state - whether the
unpatchifywrapper is active, and whether it's the LTX reference wrap or the legacy Echo memory wrap. - Reference presence - whether a
reference_latent(or legacymemory_video) exists inmodel_options.transformer_options, its shape and dtype, and which position mode is set. - Attribute side-channel - the second place the latent can live, on the diffusion model itself.
If Enable never ran, the patch-status section shows β and you've instantly found the problem. If Enable ran but no latent is attached, the reference-presence section shows β. That two-line diagnosis is the entire value of the node - it splits your debugging into "didn't patch" versus "didn't attach," which are different fixes.
Why it exists
Probe is from the same author as the reference nodes and was clearly built while debugging them. The README puts it plainly: use it before a sampler to verify wiring without running a generation. It also handles the Echo-legacy naming (memory_video), so if you're migrating a workflow from the JoyAI-Echo memory system to the LTX reference system, Probe is how you confirm which state is actually live on the model.
Installing it
Same TenStrip pack, same ritual:
cd ComfyUI/custom_nodes
git clone https://github.com/TenStrip/10S-Comfy-nodes.git 10S_Nodes
or ComfyUI Manager β "10S". Restart. No extra dependencies.
Gotchas
One, the report STRING output exists but there's no built-in display node that makes it pretty - you'll mostly read the console print, or wire it to a Show Text / display node if you want it in the UI. Two, remember probe is a snapshot: it inspects the model you hand it at that moment in the graph, so put it after both Enable and Conditioning, immediately before your sampler, to get the real picture.
Inputs (1)
| Name | Type | Default | Description |
|---|---|---|---|
| model | MODEL | β |
Outputs (2)
| Name | Type | Description |
|---|---|---|
| model | MODEL | β |
| report | STRING | β |