LTEx LTXVAddGuide (Cond Only)
LTEx LTXVAddGuide (Cond Only)
- positive
- negative
- vae
- latent
- image
- attention_mask
- iclora_parameters
- positive
- negative
- latent_passthrough
Every LTX user eventually meets the guide-crop shuffle: LTXVAddGuide appends your guide frames to the working latent, which means a LTXVCropGuides node has to strip them back out downstream, and the latent you pass around isn't the latent you think it is. This node is the pack's experiment in skipping that whole dance - it carries the guide on the conditioning instead, injects it as extra tokens inside the model's forward pass, and strips it before output. No crop node, latent never touched.
Name note first: the page title says "Cond Only," but in ComfyUI you'll find it as LTEx LTXVAddGuide (Cond Only) - the pack's display name. Same thing.
The honest framing
Read the description carefully before you build a production workflow on this: it's a proof of concept. The author's own docs call it "a drop-in alternative to LTXVAddGuide for A/B testing the conditioning-carried path." That's what it's for. If you just want guides working today, use the stock nodes - they're battle-tested. Reach for this one when the crop-node indirection is genuinely annoying you, or when you're curious whether the conditioning path behaves differently (it does: strength semantics and per-token timestep handling differ subtly from the latent-appended route).
It works for LTXV and for LTXAV's video stream - on the audio-video model, the guide targets the video slot and leaves audio alone.
How it works
Where the rest of the pack refuses to touch core, this node deliberately installs a gated monkeypatch of LTXBaseModel._forward. That's because reaching mid-forward - between input processing and timestep preparation, where the guide has to go in - isn't possible through any official wrapper hook. The patch is idempotent and a complete no-op unless the guide_cond_latents conditioning key is present, so normal LTXV/LTXAV sampling stays byte-identical to stock. The trade-off is documented in the code, which is the right way to ship an experiment like this.
There's also a real gotcha here: the path is gated off by default. The node raises RuntimeError telling you to set LTXCONDGUIDE_ENABLED to True in guide_cond.py inside the pack's folder, and restart. That's a deliberate flag on an experimental feature, not a bug - but it will absolutely stop you cold the first time if you don't know it's there.
The inputs
- positive / negative: your conditionings, in and out unchanged in structure.
- vae and latent: the VAE encodes the guide; the latent only sizes that encode. It is not modified - you get it back on the
latent_passthroughoutput. - image: the guide image or video. Must be
8n + 1frames or it's cropped to the nearest. - frame_idx (default 0, in pixel frames): where the guide sits. For 9+ frame guides it must be divisible by 8 (rounded down); negative values count from the end.
- strength (default 1.0): how hard the guide pins its frames.
1.0locks them clean, lower loosens,0ignores the guide. - attention_mask (optional): a pixel-space spatial mask controlling per-region influence, same semantics as
LTXVAddGuide. - iclora_parameters (optional): from a Get IC-LoRA Parameters node, for downscaled IC-LoRA references - the low-res reference gets spread across the full-res grid via dilated RoPE coordinates.
Outputs: positive, negative, and latent_passthrough (the unmodified input latent).
Installation
Standard pack install - Manager search "LTExperiments", or:
cd ComfyUI/custom_nodes
git clone https://github.com/drozbay/LTExperiments
then restart. No extra Python dependencies, no model downloads, and it's by drozbay - the same person keeping RES4LYF maintained through 2026, so the code quality is well above the average custom-node repo. That's partly why the gating flag reads as prudence rather than sloppiness.
Where people get burned
- The
LTXCONDGUIDE_ENABLEDgate. Set it, restart, or the node errors on execute. - Treating it as a permanent replacement for
LTXVAddGuide. It's an A/B experiment; if a workflow breaks after an update, this node is the first suspect. - Forgetting the
8n + 1frame rule on guide videos - the crop-to-nearest behavior can silently shift your guide's timing.
If you're already comfortable with LTX guides, this is a fun one to A/B against the stock path and see which looks better to you. Just don't let the "Cond Only" badge convince you it's the stable default.
Inputs (9)
| Name | Type | Default | Description |
|---|---|---|---|
| positive | CONDITIONING | — | |
| negative | CONDITIONING | — | |
| vae | VAE | — | |
| latent | LATENT | The latent being sampled. Used only to size the guide encode; not modified. | |
| image | IMAGE | Guide image or video. Must be 8*n + 1 frames or it is cropped to the nearest. | |
| frame_idx | INT | 0-16384–16384 | Pixel-frame index to place the guide at. For 9+ frame guides, frame_idx must be divisible by 8 (rounded down). Negative values count from the end. |
| strength | FLOAT | 1.000–1 | How strongly the guide pins its frames. 1.0 locks them (clean), lower loosens; 0 ignores the guide. |
| attention_maskopt | MASK | Optional pixel-space spatial mask. Controls per-region guide influence via self-attention, multiplied by strength. Same semantics as LTXVAddGuide. | |
| iclora_parametersopt | IC_LORA_PARAMETERS | Optional IC-LoRA parameters (e.g. reference_downscale_factor > 1) from a Get IC-LoRA Parameters node. The low-res reference is spread across the full-res grid via dilated RoPE coordinates; the latent is still untouched. |
Outputs (3)
| Name | Type | Description |
|---|---|---|
| positive | CONDITIONING | — |
| negative | CONDITIONING | — |
| latent_passthrough | LATENT | Input latent, returned unmodified |