💪TB | Load T5 Encoder (Experimental)
Squeezing a 5 GB T5 encoder onto a mid-range GPU, experimentally
- CLIP
TinyBreaker's base model is genuinely tiny - 0.6B parameters, fast on a 3080. Its text encoder is the opposite. The T5-XXL it leans on is a 4.9 GB fp8 download and the biggest single thing in the whole setup. 💪TB | Load T5 Encoder (Experimental) is Martin Rizzo's answer to that asymmetry: a loader that swaps how the T5 runs so mid-range and low-end GPUs don't choke on it. The "experimental" in the name is doing a lot of work - treat this as a sharp tool, not a set-and-forget one.
What it actually does
ComfyUI's default T5 handling works, but it's not always kind to small VRAM budgets. This node gives you three independent dials:
inference_mode- the interesting one.comfyui nativeis the stock path.cpu (slow)pins the encoder to system RAM - slow per prompt, but leaves the GPU free for the denoising stages.gpu (high vram usage)is the fast path for cards with headroom.dynamic loading(the default when you pickauto) loads parts on demand and is the mode to try first on a 6–8 GB card. The author's framing for the whole pack is "run on mid and low-end hardware by offloading models to RAM", and this node is where that philosophy is most visible.inference_dtype-bfloat16orfloat32. Note what's not there: fp16. The code disables it because T5 doesn't work properly in fp16 - a real footgun if you habitually set every dtype widget to fp16.autoresolves to float32.type-sd3orpixart; this tells ComfyUI how to interpret the embeddings the T5 produces.autopickssd3.
All three default to auto, which in practice means dynamic loading + float32 + sd3. That's the right starting point for a 6–8 GB card.
Output
One output: CLIP. It's the loaded T5 encoder ready to plug into the pack's base-stage CLIP input (the same wire the main checkpoint loader's CLIP output feeds). So you have a genuine choice in a workflow: let the checkpoint loader supply the stock T5 CLIP, or override it with this node when VRAM is tight.
Install and the file it wants
Same pack install as everything else here (Manager → search "tinybreaker", or git clone https://github.com/martin-rizzo/ComfyUI-TinyBreaker into custom_nodes, then restart). The encoder file is the one from the README: t5xxl_fp8_e4m3fn.safetensors (4.9 GB), placed in ComfyUI/models/clip or text_encoders. It's the Comfy-Org fp8 build shared with FLUX and SD3.5, so you may already have it sitting in your models folder from another project - the loader will pick it up.
Where people get burned
The obvious trap is expecting a speedup and getting the opposite: cpu (slow) exists to make things fit, not to make them fast, and gpu (high vram usage) can push a small card into OOM that dynamic loading handles fine. There's also a plain reality about the underlying model: even when the T5 runs smoothly, TinyBreaker cannot render legible text in images, because its PixArt base wasn't trained for it. A fancy text encoder can't fix that.
And remember it's experimental. If auto gives you weird behavior, the pragmatic fallback is comfyui native mode - it's more VRAM-hungry but it's the code path ComfyUI upstream actually maintains, and sometimes boring is the most reliable setting on this node.
Inputs (4)
| Name | Type | Default | Description |
|---|---|---|---|
| t5_name | COMBO | The name of the T5 encoder checkpoint to load. | |
| type | COMBO | auto | Specifies the model format in which ComfyUI processes embeddings generated by the T5 encoder. |
| inference_mode | COMBO | auto | Choose the code that will be used to perform inference. |
| inference_dtype | COMBO | auto | Data type used for inference. |
Outputs (1)
| Name | Type | Description |
|---|---|---|
| CLIP | CLIP | The loaded T5 Encoder ready for use as a CLIP connection. |