Priming Models

ImageNet-256 class-conditional B/L checkpoints after 80/160 epochs of reconstruction pretraining, before and after 25,000 distributional post-training updates.

Files under checkpoints/pretraining/ contain original full pretraining checkpoints with online and EMA 0.9999 weights. Files directly under checkpoints/ contain full post-training checkpoints; evaluation uses their EMA 0.999 weights. Post-training initializes from the pretraining online weights. These are original PyTorch training checkpoints, including optimizer state, not standalone Diffusers pipelines or safetensors exports.

The two JSON manifests identify all eight files with sizes and SHA256 hashes. See the code and setup guide for model definitions, evaluation, training recipes, and reproduction records.

Diffusers inference

Verified post-training EMA exports of B-160 and L-160 are available in diffusers/priming_b_160 and diffusers/priming_l_160. Each includes FP32 Safetensors, configuration, self-contained custom pipeline, transformer/scheduler code, Apache-2.0 license, attribution, and conversion checks. These exports contain inference weights only (no optimizer or training state).

Install PyTorch with CUDA support, diffusers==0.37.1, huggingface_hub, and Pillow. No Priming repository installation is needed:

from pathlib import Path
import torch
from huggingface_hub import snapshot_download
from diffusers import DiffusionPipeline

variant = "priming_b_160"  # or priming_l_160
snapshot = snapshot_download(
    "jaykim1102/priming", revision="main",
    allow_patterns=[f"diffusers/{variant}/*"],
)
folder = Path(snapshot) / "diffusers" / variant
pipe = DiffusionPipeline.from_pretrained(
    str(folder), custom_pipeline=str(folder / "pipeline.py"),
    local_files_only=True,
).to("cuda")
with torch.autocast("cuda", dtype=torch.bfloat16):
    image = pipe(
        class_labels=207,
        generator=torch.Generator(device="cuda").manual_seed(1),
    ).images[0]
image.save("sample.png")

This loads and executes the included custom Python components; inspect their source and pin revision to a Hub commit for reproducible use. The snapshot filter downloads only the selected export. Native output is 256Γ—256, one step, with guidance_scale=1; class IDs 0–999 are supported.

Keep FP32 weights and use BF16 autocast to match the original sampling mode. Passing torch_dtype=torch.bfloat16 changes the weights and produces different numerical outputs. B/L fixed-input FP32/BF16 comparisons against the original implementation have zero error in matching precision modes. This is conversion validation, not a new FID evaluation.

Adapted from BiliSakura's community implementation, with retained attribution. This custom pipeline is not upstream Diffusers support.

Evaluation reference assets

assets/fid_stats/ contains the frozen FDr5 reference statistics and the extra ConvNeXt statistic used during training diagnostics. The precomputed Inception precision/recall reference features are under data/prc/. The code's scripts/download_assets.py downloads the required subset, pins upstream evaluation encoders, and fetches the separate official JiT FID reference. Users evaluating Priming do not need to recompute real-image statistics.

The processed ImageNet training release is in the separate dataset repository, with ImageNet access terms and matching class indices/OT calibration. Neither sampling nor final FID/FDr5 evaluation requires that training dataset.

Text-to-image checkpoints

Native MM-JiT-B/16 checkpoints at 512Γ—512 are available before and after distributional OT post-training: 250k + 40k reconstruction updates, then 20k OT updates. Both online and EMA weights are included.

See the T2I release guide for downloads, public-code sampling commands, exact stage definitions, and compatibility checks. The post-OT online model achieves GenEval 0.8280 and DPG-Bench 82.71 with one network evaluation and CFG=1. These T2I checkpoints use the native Priming code, separately from the ImageNet Diffusers pipelines above.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support

Space using jaykim1102/priming 1