GuidanceNFT checkpoints

Fine-tuned checkpoints reported in GuidanceNFT: NFT-Based Reward Fine-Tuning of Guided Diffusion Models (arXiv:2610.15856, project page). Code, configs and evaluation scripts: github.com/Teresa-l23/GuidanceNFT. Every folder holds a checkpoint.json with the training config, the update count and the SHA-256 of each file.

SiT-XL/2 and Model Guidance on ImageNet 256 (class-name PickScore)

model.safetensors holds the policy weights that were evaluated in the paper; config.json is the training config. Evaluate with sit_imagenet/evaluate_released.py.

Folder	Method	Update	PickScore (paper protocol)
sit_nft_cfg1_step200	NFT, CFG-1 rollouts	200	20.16 (CFG 1), 20.47 (CFG 1.5)
sit_guidancenft_cfg1p5_step400	GuidanceNFT, CFG-1.5 rollouts	400	21.47 (CFG 1), 21.58 (CFG 1.5)
mg_nft_step200	NFT on Model Guidance	200	20.01
mg_guidancenft_step400	GuidanceNFT on Model Guidance	400	21.58

Base models: SiT-XL/2 and Model Guidance SiT-XL/2.

SD3.5-Medium LoRA adapters

Each lora/ folder is a PEFT adapter for the transformer of stabilityai/stable-diffusion-3.5-medium (revision b940f670). GuidanceNFT adapters also contain v5/, the learned center, and old/, the old policy at the same update.

Task	NFT (CFG 1)	NFT, both branches (CFG 1)	GuidanceNFT (CFG 4 rollouts)
GenEval	sd35_geneval_nft_cfg1_step200	sd35_geneval_nft_cfg1_dual_step200	sd35_geneval_guidancenft_cfg4_step200
Learned3	sd35_learned3_nft_cfg1_step200, _step300	sd35_learned3_nft_cfg1_dual_step200	sd35_learned3_guidancenft_cfg4_step200, _step300
PickScore	sd35_pickscore_nft_cfg1_step100, _step200	sd35_pickscore_nft_cfg1_dual_step100	sd35_pickscore_guidancenft_cfg4_step100, _step200
OCR	sd35_ocr_nft_cfg1_step50, _step200	sd35_ocr_nft_cfg1_dual_step50	sd35_ocr_guidancenft_cfg4_step50, _step200

sd35_pickscore_center_init is the PickScore run after center initialization (100 fitting epochs, policy unchanged); the center-refresh ablation starts from it.

ELF-B on OpenWebText

Training checkpoints (Flax msgpack) at update 100, loaded by the elf_owt code: elf_owt_nft_step100 (NFT, SCCFG 1), elf_owt_guidancenft_step100 (GuidanceNFT, SCCFG 3), and elf_owt_lfrm_reconstruct_step100 (LFRM-style reconstruction, SCCFG 1).

Qwen-based ELF-B on GSM8K

CEDR inference packages (ELF and prompt encoder at EMA 0.99), loaded by cedr generate with the gsm8k code: gsm8k_nft_step200 (NFT, update 200) and gsm8k_guidancenft_step500 (GuidanceNFT, update 500), both trained with SCCFG-2 rollouts. Each package needs the shared tokenizer and embedding of the CEDR gsm8k-b packages in shared/; gsm8k/scripts/download_checkpoints.sh adds them and verifies every file.

Folder	Method	Update	Pass@1	Pass@8
gsm8k_nft_step200	NFT	200	41.39	67.17
gsm8k_guidancenft_step500	GuidanceNFT	500	43.04	68.61

Pass@k (%) on the 1,319 GSM8K test questions over eight seeds, with 64 steps and SCCFG 2.

License

The adapters and weights are derivatives of their base models and follow their licenses: the Stability AI Community License for SD3.5-Medium, MIT for SiT-XL/2 and ELF, and Apache-2.0 for the CEDR ELF-B models, which retain the attribution of Qwen3-4B-Instruct-2507. Our contributions are released under MIT.