New Release: Ultralytics v8.3.219

Discussion
, ,
1

Ultralytics v8.3.219 — FastSAM prompt reliability upgrade :rocket:

Quick summary

  • FastSAM gets a robust mask-resizing fix and a cleaner, built-in CLIP integration for prompt-based segmentation. Expect more accurate masks, fewer type-related issues, and simpler setup. Version bump to v8.3.219.

New features

  • Built-in CLIP for text prompts
    • FastSAM now uses the internal ultralytics.nn.text_model.CLIP for text-image similarity, removing the external dependency. This standardizes behavior across environments and simplifies offline installs.
    • You can review the change in the PR titled FastSAM masks .float() upsampling fix (PR #22460) by Y-T-G.

Improvements

  • More reliable mask resizing in FastSAM
    • Masks are converted to float before upsampling and thresholded back to byte. This prevents artifacts and type-related errors during prompt workflows, improving accuracy and stability for a wide range of image sizes and PyTorch versions.
    • Details are available in the same PR #22460 implementing the mask and prompt improvements by Y-T-G.
  • Version bump
    • Package version updated to 8.3.219 for distribution consistency.

Bug fixes

  • Addressed incorrect or unstable mask upsampling paths that could cause visual artifacts or dtype errors in specific configurations. The fix is included in the FastSAM mask upsampling PR credited to Y-T-G.

Why this matters

  • Accurate text prompts: Cleaner, more reliable masks improve downstream prompt-based segmentation results.
  • Fewer environment issues: The internal CLIP integration reduces ad-hoc installs and runtime imports.
  • No action needed: Existing FastSAM workflows benefit automatically after upgrading.

Getting started or upgrading

  • Install or upgrade with pip install -U ultralytics.
  • Try FastSAM text prompts with built-in CLIP:
from ultralytics import FastSAM
from ultralytics.models.fastsam import FastSAMPrompt

# Load model
model = FastSAM('FastSAM-x.pt')  # or 'FastSAM-s.pt'
image = 'path/to/image.jpg'

# Run inference
results = model(image, imgsz=1024, conf=0.4, retina_masks=True)

# Prompt-based segmentation (uses built-in CLIP under the hood)
prompt = FastSAMPrompt(image, results)
ann = prompt.text_prompt(text='cat')
prompt.plot(annotations=ann, output='fastsam_text_prompt.jpg')

Recommendation

  • YOLO11 remains the latest stable and recommended Ultralytics YOLO model for all use cases. You can learn more in the YOLO11 model overview.

Changelog and release links

Call for feedback

  • We’d love to hear how the new prompt flow performs for you. Share your experience or questions in Ultralytics GitHub Discussions, or report issues via the issue tracker. Your feedback helps the community and the Ultralytics team keep improving.