💪TB | Set Image

The first thing you'll notice about the TinyBreaker pack is that it doesn't work like the ComfyUI you're used to. Instead of separate sliders for every knob, most settings travel along one wire called genparams - a bundle of generation parameters that nodes hand to each other. 💪TB | Set Image is one of the writers for that bundle, and it's where you decide the shape and size of the image before anything else happens.

That single design choice is the whole reason this pack exists. TinyBreaker is an experimental base+refiner model built on top of PixArt Sigma (0.6B params) plus a Photon refiner, tuned to generate big images fast on modest GPUs - roughly 1536×1024 in about 12 seconds on an RTX 3080. Squeezing a usable workflow out of it means keeping the graph small, and folding image format into one parameter object is how the author (Martin Rizzo) does that. Set Image isn't glamorous, but it's the node you'll wire first in almost every TinyBreaker graph.

What you actually set

Four inputs, and the first two matter far more than the rest:

• ratio - a dropdown of ten aspect ratios, from 1:1 (square) through 16:9 (hdtv) all the way to 32:9 (s.ultrawide). Because the model was trained with aspect-ratio bucketing, it copes with the extremes instead of collapsing into a blurry mess, which is more than you can say for some older checkpoints.
• orientation - landscape or portrait. This picks which dimension gets the long side; pair it with a ratio and you're done.
• size - small, medium, large. The tooltip is honest about the catch: medium is the size the model was actually trained on, but large is recommended anyway. TinyBreaker was built to punch above its native resolution, so don't be shy about it.
• batch_size - how many images to generate in one run. Default 1, and you rarely need more.

The node also takes the genparams input so it can copy and extend the bundle rather than start from scratch. Its single output is that updated genparams, and the tooltip spells out the intended usage: chain it into other genparams nodes. That's the pattern - Set Image feeds something like 💪TB | Unified Prompt Input, which feeds the sampler.

Installing it

The pack installs the usual way:

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

Or search "tinybreaker" in ComfyUI Manager and click Install. Either way, restart ComfyUI after. There's no requirements.txt - the pack is pure Python, so no pip dependency fights. The real prerequisite is the model files: tinybreaker_prototype1.safetensors (3 GB) into models/checkpoints and t5xxl_fp8_e4m3fn.safetensors (4.9 GB) into models/clip or models/text_encoders.

Where people trip up

The classic gotcha: you fiddle with ratio, hit queue, and nothing changes. That's almost always because the node downstream isn't reading the genparams you think it is - check that the sampler is actually getting the genparams wire, not just a LATENT and a MODEL. Also note this node only writes the format; the seed, prompt, and style live in other genparams nodes. If you're coming from a standard ComfyUI workflow, it takes a minute to get used to the idea that "settings" are now a data type you can route and merge. Once it clicks, it clicks hard.