import torch
from diffusers import DiffusionPipeline
# switch to "mps" for apple devices
pipe = DiffusionPipeline.from_pretrained("BreakpointAI/socknetq", dtype=torch.bfloat16, device_map="cuda")
prompt = "Astronaut in a jungle, cold color palette, muted colors, detailed, 8k"
image = pipe(prompt).images[0]socknetq
Joint image + bounding-box diffusion model with a LoRA adapter, a confidence stream, and Nash-MTL multi-task loss weighting.
Released by Breakpoint AI as part of open-sourcing the company's research artifacts.
| Training data | BreakpointAI/breakpoint-grounding-55m |
| Checkpoint step | 3,120,000 |
| Training run | W&B |
Contents
| Path | Size | What it is |
|---|---|---|
pytorch_model/ |
67.6 GB | |
boxnet/ |
20.1 GB | Joint image + bounding-box diffusion backbone |
pytorch_lora_weights.safetensors |
1.0 GB | LoRA adapter weights |
conf_weighter.pth |
2.0 KB | Confidence-stream loss-weighting head |
weighter.pth |
1.9 KB | Loss-weighting head |
nash_mtl_weights_conf_ema.safetensors |
120 B | Nash-MTL task-weighting coefficients |
nash_mtl_weights_ema.safetensors |
112 B | Nash-MTL task-weighting coefficients |
global_conf_step.txt |
7 B | Training step counter |
global_nonull_step.txt |
7 B | Training step counter |
weighter_steps.txt |
4 B | Loss-weighter step counter |
Inference weights only. Optimizer, LR scheduler, RNG and dataloader state were not uploaded, so this checkpoint cannot be used to resume training.
Citation
@misc{breakpoint_socknetq,
title = {socknetq},
author = {Wang, Franklin and Zhong, Desmond and Murdoch, Jamie},
year = {2026},
url = {https://huggingface.co/BreakpointAI/socknetq}
}
- Downloads last month
- -