💪TB | Load TinyBreaker Checkpoint

This is the node the whole pack is built around, and the reason it looks so different from a normal checkpoint loader is that a TinyBreaker checkpoint isn't one model. It's a stack: a PixArt-Σ base model (0.6B parameters), a Photon/SD1.5 refiner, the Tiny AutoEncoders that convert latents between those two stages, and a couple of VAE variants. 💪TB | Load TinyBreaker Checkpoint unpacks all of it from a single .safetensors file and hands you eight outputs instead of the usual three. First time you see it sprawled across the canvas you'll wonder if you grabbed the wrong node - you didn't.

Why the eight outputs

TinyBreaker's whole trick is doing a fast base pass then a refiner pass, with tiny autoencoders (the TAESD family) shuttling the latent between the two models so they don't have to speak the same latent dialect. That means one loader outputting:

• MODEL and CLIP - the base model and its text encoder, for the first denoising stage.
• VAE - for encoding/decoding pixels to and from latent space at the end.
• TRANSCODER - the bridge that converts the base stage's latents into the refiner's latent space. This is the pack's own custom type; nothing in vanilla ComfyUI produces or consumes it.
• REFINER_MODEL and REFINER_CLIP - the second-stage model and its encoder.
• UPSCALER_VAE - a separate VAE used during the tiny upscaler pass, only active since v0.3 with the prototype1 checkpoint.
• METADATA - the generation parameters embedded in the checkpoint's metadata, output as a GENPARAMS blob you can feed straight into the rest of the chain.

Wiring it the "usual" way means MODEL and CLIP into the base sampler, the decoded latent into the refiner, and VAE out to the final VAE Decode. But if that sounds like assembly work, don't worry - the pack's reference workflow (drag the workflow image into ComfyUI, or grab one from the workflows folder) does it all for you, and it's the fastest way to see how the pieces connect.

The inputs you'll actually touch

• ckpt_name - the TinyBreaker checkpoint to load. It reads from ComfyUI/models/checkpoints, and the one you need is tinybreaker_prototype1.safetensors (3.0 GB, from CivitAI). The "prototype1" version matters: the tiny upscaler only works with it.
• vae_type - auto / fast / high_quality. high_quality gives the nicer decode but eats noticeably more VRAM and time; on a mid-range card, auto or fast is the sane default.
• upscaler_vae_type - same trio, but for the upscaling VAE. The README is blunt here: high_quality is available but its VRAM appetite makes auto/fast the recommendation.

Install and the two files you must not skip

Install via ComfyUI Manager (search "tinybreaker") or:

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

Restart, then place both of these or nothing works:

• tinybreaker_prototype1.safetensors (3.0 GB) → ComfyUI/models/checkpoints
• t5xxl_fp8_e4m3fn.safetensors (4.9 GB) → ComfyUI/models/clip or text_encoders

That second file is the T5-XXL text encoder - the same one FLUX and SD3.5 use - and it's the pack's real heavyweight dependency. The Python side is light: the pack ships no extra pip requirements, just ComfyUI's own.

Gotchas

The most common failure is the empty ckpt_name dropdown, which just means the checkpoint isn't in models/checkpoints. If the loader errors on missing keys, you've got a partial or wrong file - verify the 3.0 GB download, it's easy to end up with a truncated one. And don't expect legible text from this model: the author is upfront that the PixArt base was never trained for it. If you're planning a workflow full of signage, this pack will disappoint you.

For anyone on this page because they saw the "❌ Load TinyBreaker Checkpoint [Deprecated]" node in an old workflow: that's the pre-2025 version with fewer outputs and no CLIP/upscaler split. This loader is the current one, and the advanced variant (Load TinyBreaker Checkpoint (advanced)) is the same idea with per-sub-component override controls if you ever need them.