Nodes/ComfyUI-LocateAnything/LocateAnything Grounding
ComfyUI Node

LocateAnything Grounding

Runs visual grounding frame by frame for an IMAGE batch. Task modes: - ground_multi: locate every matching instance. - ground_single: locate one matching instance. - detect: detect matching categories or descriptions. - ground_text: locate a requested text phrase. - detect_text: detect scene text; query is ignored. - gui_box: locate a GUI region. - gui_point: return a point for a GUI target. - point: return a point for a described target. - custom: use query as the complete prompt. Coordinates are parsed from the model's normalized [0, 1000] format. A batch of video frames is processed independently, with native ComfyUI progress updates.

By alisson-anjos·Created 2 months ago·Updated 2 months ago· 17
LocateAnything Grounding
  • model
  • image
  • answer
  • locations_json
  • annotated_image
  • mask
  • mask_overlay
taskground_multi
queryperson
generation_modehybrid
max_new_tokens2048
temperature0.00
top_p0.90
top_k0
repetition_penalty1.10
point_radius12
mask_grow0
mask_blur0.0
overlay_color#00ff66
overlay_opacity0.45
seed0
verbosetrue
CategoryLocateAnything

Inputs (17)

NameTypeDefaultDescription
modelLOCATEANYTHING_MODEL
imageIMAGE
taskCOMBOground_multiOperation mode. Hover the node help (?) for the complete list. Use custom to send query as the full model prompt.
querySTRINGpersonDescription, text, GUI target, or full prompt for custom mode.
generation_modeCOMBOhybridDecoding strategy: hybrid uses fast decoding with stable fallback; fast prioritizes speed; slow prioritizes the stable path.
max_new_tokensINT20481–8192Maximum generated tokens. Reduce this when shorter answers are sufficient.
temperatureFLOAT0.000–2Sampling temperature. 0.0 is the safest path; the node retries with safe settings if higher temperatures fail.
top_pFLOAT0.900–1Nucleus sampling cutoff. Relevant when temperature is above 0.
top_kINT00–1024Top-k sampling cutoff. 0 disables top-k, matching the official worker.
repetition_penaltyFLOAT1.100–3Penalty for repeated tokens. The official worker uses 1.1.
point_radiusINT121–512Radius in pixels used to draw point results into the output mask.
mask_growINT0-512–512Grow the output mask by this many pixels. Negative values shrink it.
mask_blurFLOAT0.00–100Gaussian blur radius applied after mask grow. Use 0 for hard edges.
overlay_colorSTRING#00ff66Hex RGB color used by the mask overlay preview, for example #00ff66 or #ff0000.
overlay_opacityFLOAT0.450–1Opacity of the colored mask overlay preview.
seedINT00–18446744073709550000Sampling seed. Relevant when temperature is above 0. For IMAGE batches, frame N uses seed + N.
verboseBOOLEANtruePrint the official generation step log in the terminal. Disable for quieter runs.

Outputs (5)

NameTypeDescription
answerSTRINGRaw model response. For batches, this is a JSON array of responses.
locations_jsonSTRINGStructured JSON with prompt, timing, normalized coordinates, pixel coordinates, and batch index.
annotated_imageIMAGEInput image batch annotated with returned boxes and points.
maskMASKPost-processed mask batch after grow and blur: filled boxes and circles centered on returned points.
mask_overlayIMAGEOriginal image batch with the post-processed mask blended using overlay_color and overlay_opacity.