stella221125/HPSv3-PlusPlus-bnb-NF4

HPSv3++ — bitsandbytes NF4

Unofficial prequantized derivative for local image/prompt preference scoring with hpsv3-4bit. Download only 6.454 GB of weights, instead of the BF16 source. All model config, reward settings, tokenizer and processor files are included. The BF16 checkpoint and base-model weight download are unnecessary for inference.

Usage

Requires an NVIDIA CUDA GPU, approximately 9 GB free VRAM, and the corresponding project environment (Python 3.12; see its uv.lock for exact dependencies).

git clone --recurse-submodules https://github.com/Stella2211/hpsv3-4bit
cd hpsv3-4bit
uv sync --project hpsv3pp
uv run --project hpsv3pp hpsv3pp/scripts/score_batch.py --model stella221125/HPSv3-PlusPlus-bnb-NF4 --input records.json --output scores.json

records.json is a list of objects with id, image (local path) and prompt. --input also accepts a directory of images with matching .txt prompt files. The project downloads the full model repository, including README, LICENSE and NOTICE, from the latest main by default. Use --revision COMMIT to select a specific version when needed. Cached inference supports HF_HUB_OFFLINE=1 and --local-files-only.

Use the project's custom reward-model loader. A generic AutoModel or hosted text-generation pipeline is not a supported scorer. No remote Python code is included; HPSv3++ still requires the project's pinned upstream submodule.

Conversion and provenance

• Official weights: Junjun2333/HPSv3-PlusPlus.
• Architecture and tokenizer: Qwen/Qwen3-VL-8B-Instruct.
• bitsandbytes NF4, double quantization, BF16 compute; no additional training or calibration. Source revisions and SHA-256 hashes are in release_manifest.json.

Limitations

NF4 scores are not guaranteed to match full BF16 inference or human preferences. Different devices, kernels, dependency versions, prompts and batching may change scores. Use iter_step=0.0 for plain preference scoring; batch composition can affect conditioning. This model inherits limitations and biases of the upstream reward and base models.

License and privacy

Model weights are distributed under Apache-2.0; see LICENSE and NOTICE for attribution and modifications. Repository code has separate licensing; the upstream Python implementation is not included in this model repo.

Only model assets and public documentation are distributed. Local paths, private evaluation inputs, per-image scores, logs and credentials are not included. Images and prompts used for local scoring are not uploaded by this project. Initial model retrieval contacts Hugging Face; offline cached inference does not require contacting it.