BlehTAEVideoEncode

This is the mirror image of BlehTAEVideoDecode: instead of turning a video latent into frames, it turns frames into a video latent, using the same family of fast approximate TAE models (from madebyollin, the person behind TAESD) built specifically for Wan, Hunyuan, Mochi, and LTX latent formats. Where the decoder is useful for quick previews and exports, the encoder is what you reach for when you need to get existing video frames into one of these models' latent space fast - say, a video-to-video workflow where you're feeding real or pre-generated frames back in for further processing, and a full, precise VAE encode is more cost than you want to pay for that step.

Same core tradeoff as its decode sibling: real VAE encoding is accurate but comparatively expensive, especially across a whole clip's worth of frames. The TAE version is dramatically cheaper and a bit less faithful - the right call when speed matters more than squeezing out the last bit of encode fidelity.

Inputs and outputs

• latent_type - which TAE model to encode with, matched to the target latent format: hunyuanvideo, ltxv, ltxv23, ltxv23wide, mochi, wan21, or wan22. Same tooltip note as the decode node applies here too: use ltxv for "LTX-2 AV" specifically, since the LTX variants don't map one-to-one onto obvious names.
• parallel_mode - boolean, default false. Faster when enabled, at the cost of using more memory - the tooltip states this plainly, so treat it as a real tradeoff rather than a free win.
• image - the IMAGE input: your source frames, as an image batch.
• Output: LATENT, in the target model's latent space, ready to feed into further sampling or processing for that model.

Installing it - same models as the decode node

Install the pack once and you get both encode and decode:

• ComfyUI Manager - search "ComfyUI-bleh", install, restart.
• Manual - cd ComfyUI/custom_nodes && git clone https://github.com/blepping/ComfyUI-bleh, restart ComfyUI.

The model files are shared between encode and decode - download once, both nodes work. Drop the matching .pth into models/vae_approx, filename unchanged:

• Wan 2.2: taew2_2.pth - only for the 5B Wan 2.2 model; the 14B Wan 2.2 models actually use the Wan 2.1 latent format under the hood.
• Wan 2.1: taew2_1.pth
• Hunyuan: taehv.pth
• Mochi: taem1.pth
• LTX 2.0: laeltx_2.pth
• LTX 2.3: laeltx2_3.pth
• LTX 2.3 wide: taeltx2_3_wide.pth

Check the pack's README for the exact download links - they come from a couple of different madebyollin repos and one lives on a specific non-main branch, so the filename alone isn't enough to find the right source.

Common issues

Node errors immediately on run. Almost certainly the model file - these aren't bundled, and if the right .pth isn't sitting in models/vae_approx under the exact expected name, the node has nothing to load.

Encoded latent doesn't behave right in downstream sampling. Check latent_type against what you actually intend to feed this latent into next. Encoding with the wrong TAE variant produces a latent shaped like the wrong model's expectations - remember the Wan 2.2 14B/2.1-format gotcha specifically, since it's the one exception to "match the model name" that isn't obvious from the option list alone.

Running out of memory with parallel_mode on. That's the documented tradeoff, not a malfunction - turn it off if you're VRAM-constrained, particularly with longer clips where the memory cost compounds across more frames.

Result looks noticeably different from a real VAE encode. Expected, same as any TAE-family shortcut - it's approximate by design. If you're doing a final, quality-critical pass rather than a fast iteration or preview step, use the model's real VAE encode instead and save this node for where speed is the priority.