csig-track1-fromild-s600

csig-track1-fromild-s600

One-step VOSR-0.5B restoration model used for CSIG Track-1 submissions W12–W15 (W14 official score 4.2512; W15 = same pixels at JPEG q100).

clean_weights/model.safetensors — full weights with the LoRA already merged (1948337520 bytes, md5 0371a50f0f51f6ebccc5974586689921). args.json is the training run's config as read by the inference script.

Recipe

Base = csig-track1-vosr05b-ckpt4000

• 600 steps LoRA (rank 4) with two changes (train_config.yml):

• fr_on_infer: true — the fidelity loss is applied to the one-step inference output rather than to x̂₀ at a random t (whose optimum is a mean prediction ⇒ blur).
• params_aigc_mild.yml — degradation strength re-calibrated to the real test-set severity (LPIPS(LQ,HQ) ≈ 0.33; the previous recipe was 2.3× heavier).

Use

Code: https://github.com/woodlingbombardier-dev/csig (run/pipeline.sh, default = W15).

huggingface-cli download wuhisbajsi/csig-track1-fromild-s600 --local-dir fromild
export CSIG_CKPT=$PWD/fromild
cd vosr && python inference_vosr_onestep.py -c $CSIG_CKPT -i inputs -o out \
  -u 1 --align_method wavelet --tile_size 512 --tile_overlap 64 --infer_steps 1 --seed 42 --cond_strength 0.90

Delivery chain W15 then applies track1/pipeline/op_tto_joint.py (50 steps, JPEG q100). ⚠ That TTO step is direct gradient ascent on the evaluation metrics; see docs/FINDINGS.md §7/§10.

Also needs the upstream VOSR presets: stable-diffusion-2-1-base VAE, sd21_lwdecoder.pth, DINOv2 ViT-B/14.