💪TB | Load TinyBreaker Checkpoint (advanced)

The standard 💪TB | Load TinyBreaker Checkpoint auto-detects just about everything inside the checkpoint. 💪TB | Load TinyBreaker Checkpoint (advanced) is the version that hands you the manual override for each sub-component instead. Same eight outputs, same hybrid model, but now you can force which VAE, which transcoder, which refiner, and even the base resolution the model assumes - useful when you're mixing components across checkpoints or squeezing a specific VRAM budget. If the basic loader "just works" for you, you don't need this one. If it doesn't, this is where the levers are.

What you can override

• vae - auto / fast / high_quality. The pixel↔latent codec for generation. high_quality looks better, costs more VRAM and time.
• transcoder - auto / embedded. The latent-space bridge between base and refiner. auto picks the best embedded option; embedded forces the one baked into the checkpoint.
• refiner - auto / embedded. The second-stage model. auto selects the best available refiner; embedded uses the checkpoint's own.
• resolution - auto / 512 / 1024 / 2048 / 4096. This tells the model the base resolution it should assume. Getting this wrong is a classic quality-killer: run a model tuned for 1024 at 512 and you get smeary mush; the dropdown exists so you can force what the metadata guessed wrong.
• upscaler_vae - auto / fast / high_quality. The VAE used during the upscale pass; high_quality here is a VRAM hog, per the tooltip, so auto/fast are the recommendation.

The outputs

Identical to the standard loader: MODEL, CLIP, VAE, TRANSCODER, REFINER_MODEL, REFINER_CLIP, UPSCALER_VAE, and METADATA (the checkpoint's embedded generation parameters as a genparams blob). Same wiring: base MODEL/CLIP to the first denoise, VAE to decode, TRANSCODER into the sampler, and METADATA can seed your genparams chain.

Install

The pack install, nothing special - ComfyUI Manager search "tinybreaker", or:

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

Restart. You still need tinybreaker_prototype1.safetensors in ComfyUI/models/checkpoints and the fp8 T5 encoder in models/clip or text_encoders.

The honest take

For 95% of runs, auto everywhere is right and this node is just the verbose sibling of the standard loader. Reach for it when you're diagnosing - a weirdly soft image often traces to the wrong assumed resolution, and an OOM trace to a high_quality VAE you can now downgrade without rebuilding the graph. One caution: these overrides fight the checkpoint's own metadata, so if you force resolution: 2048 on a model that was never trained there, you get bigger images, not better ones. The advanced loader gives you the controls, but the author's own defaults are still the best first guess.