💪TB | Load Transcoder

TinyBreaker's base model (PixArt-Σ) and its refiner (Photon/SD1.5) were trained in different latent spaces, so you can't just hand one's output to the other. The thing that translates between them is a transcoder, and 💪TB | Load Transcoder is the standalone loader for it - for when you want to pull a specific transcoder file off disk instead of using the one embedded in the checkpoint.

What a transcoder even is

Think of a VAE as a codec: it turns pixels into a compressed latent and back. A transcoder is a codec-to-codec bridge - it decodes an image out of one latent space and re-encodes it into another, so the two models in a base+refiner pipeline can talk. In this pack the transcoder shows up as its own type, TRANSCODER, and it feeds the sampler's transcoder input. The main checkpoint loader builds one automatically; this node is the explicit, file-based version of the same thing.

The one input, the one output

• transcoder_name - a dropdown listing the transcoder files it can see. Here's the catch worth knowing: it lists files from the same directory as the VAE loader, i.e. ComfyUI/models/vae. A transcoder is just a .safetensors with a transcoder-style state dict, so that's where you drop any you download.
• Output - a single TRANSCODER, ready for the sampler's transcoder input.

That's it. One dropdown, one wire, no knobs. If you're using the pack's reference workflow, you don't need this node at all - the checkpoint loader supplies the transcoder. You'd reach for it only to experiment with an alternative bridge, which is the same use case as Build Custom Transcoder, just from a file instead of from two VAEs.

Install

The pack install, as usual: ComfyUI Manager → search "tinybreaker", or

cd ComfyUI/custom_nodes
git clone https://github.com/martin-rizzo/ComfyUI-TinyBreaker

then restart. Any transcoder file you want to load goes in ComfyUI/models/vae.

Things to know before you go hunting

The name might tempt you to look for standalone "transcoder" files on CivitAI or Hugging Face. They exist, but most people never need them - the checkpoint's built-in transcoder is the one the model was tuned with, and swapping it for an arbitrary file is an experiment, not an upgrade. Also remember a transcoder is only as good as the pair of autoencoders it's built from: if you load a file that doesn't match either of TinyBreaker's latent spaces, you'll get garbage or an error rather than a politely degraded image. If your dropdown is empty, the file isn't in models/vae, or the filename has an extension the loader doesn't filter for - check the folder first.