Nodes/comfyui_toriigate/ToriiGate Captioner
ComfyUI Node

ToriiGate Captioner

A ComfyUI node in ToriiGate with 12 inputs and 1 output.

By litch230·Created 3 months ago·Updated 3 months ago· 20
ToriiGate Captioner
  • image
  • caption
model_pathMinthy/ToriiGate-0.5
devicecuda
dtypebfloat16
max_pixels_mp1.0
keep_model_alivefalse
prompt
max_new_tokens512
temperature0.50
decodinggreedy_fast
show_generation_progressfalse
seed0
CategoryToriiGate

Inputs (12)

NameTypeDefaultDescription
imageIMAGEComfyUI image tensor. The node uses only the first image in the batch and converts it to RGB before captioning.
model_pathoptSTRINGMinthy/ToriiGate-0.5Model path. Use Minthy/ToriiGate-0.5 to download from Hugging Face, or a local path for an already downloaded copy.
deviceoptCOMBOcudaWhere to run the model. auto uses CUDA when available; cuda is recommended for speed; cpu works, but will be very slow.
dtypeoptCOMBObfloat16Weight precision. bfloat16 is the native/recommended format; float16 may save VRAM on some GPUs; float32 uses more memory and is usually slower.
max_pixels_mpoptFLOAT1.00.1–8Resolution limit sent to the model, in megapixels. 1.0 MP is recommended; higher values may improve detail, but increase VRAM use and runtime.
keep_model_aliveoptBOOLEANfalseKeeps the model loaded in VRAM/cache after generation. Enable for repeated captions; disable to free VRAM after each run.
promptoptSTRINGOptional prompt. You can connect the text output from the ToriiGate Grounding Builder here, or type your own.
max_new_tokensoptINT51264–4096Maximum generated tokens. Lower this for speed: 256-512 for quick captions; 1024+ for fuller long_thoughts formats.
temperatureoptFLOAT0.500.01–2Generation randomness in sample mode. Lower values are more consistent; higher values are more creative, but may invent details.
decodingoptCOMBOgreedy_fastsample uses temperature-based sampling and can vary more; greedy_fast chooses deterministic tokens and is usually faster and more stable.
show_generation_progressoptBOOLEANfalseShows a progress bar and tokens/s in the console during generation. Disable to reduce overhead and gain a little speed.
seedoptINT00–18446744073709550000Seed for reproducibility. Use 0 for a random seed; use a fixed value to repeat results in sample mode.

Outputs (1)

NameTypeDescription
captionSTRING