Tiny FLUX.2 VAE Loader (CRT)

FLUX.2's full VAE is a heavyweight - fine at the end of a render, annoying when all you want is a quick preview or a fast latent decode. TinyFlux2VAELoader is the CRT pack's answer: it loads fal/FLUX.2-Tiny-AutoEncoder, a tiny autoencoder built on diffusers' AutoencoderTiny, and hands it to you as a TINY_FLUX2_VAE object that Tiny FLUX.2 VAE Decode (CRT) (and Encode) understand.

The clever bit is that this isn't just a fast VAE that happens to be close enough. The tiny model pads a 32-channel AutoencoderTiny with an extra stride-2 convolution to reach 128 channels at 1/16 spatial scale - which is exactly the latent format the full FLUX.2 (Klein) VAE produces. Same latent shape, same channel count, a fraction of the compute. So it's a drop-in stand-in for workflows that need the FLUX.2 latent layout but don't want to pay for the full decode on every pass.

The one input

• model_file - a single-file dropdown. If the weights aren't downloaded yet, the dropdown literally reads [NOT FOUND] diffusion_pytorch_model.safetensors, which is the node's honest way of telling you to fetch them.

The weights go in ComfyUI/models/vae_approx/FLUX.2-Tiny-AutoEncoder/ and come from the fal/FLUX.2-Tiny-AutoEncoder repo. This node doesn't auto-download them (unlike the CRT AutoDL family) - the file is the one part you fetch by hand.

The honest trade-off

"Tiny" means small and fast, and it also means lower fidelity. Use it for previews, fast iterations, latent-space experiments, or anywhere a near-enough decode beats a slow exact one. Don't ship a final image through it and then wonder why fine detail softened - that's what the real FLUX.2 VAE is for, and the pack gives you that too.

Gotchas

• diffusers must be installed. The node imports AutoencoderTiny from diffusers and won't register/load without it. The pack's requirements.txt includes it, so a clean pip install -r requirements.txt covers you; a stale install that predates these nodes may not have it.
• The [NOT FOUND] label is literal. A fresh install shows it in the dropdown until you drop the safetensors into vae_approx/FLUX.2-Tiny-AutoEncoder/.
• Keep the pair together. This loader emits the custom TINY_FLUX2_VAE type, which only CRT's Tiny FLUX.2 encode/decode nodes consume. Standard VAE sockets won't accept it.

Install CRT-Nodes through ComfyUI Manager (CRT-Nodes) or git clone https://github.com/plugcrypt/CRT-Nodes.git into custom_nodes, install requirements, restart - then grab the weights from HuggingFace and drop them where the node's [NOT FOUND] message points.