---

name: python-api-and-config description: "Use SUPIR's Python APIs, model/config objects, image/tensor helpers, and LLaVA caption bridge without running batch scripts or demos." metadata: disco-role: operating disable-model-invocation: true license: NOASSERTION

Python API and Config

Use this sub-skill when the task asks how to import SUPIR modules, construct a SUPIR model, inspect YAML config behavior, call batchify_sample, use the LLaVA caption bridge, convert images/tensors, validate dtype choices, or debug API-level import/config errors.

Do not use this sub-skill for folder CLI orchestration or browser UI launch. Route those to ../batch-restoration/SKILL.md and ../interactive-demos/SKILL.md.

Read these bundled files

• references/api-and-config.md for signatures, parameter behavior, config loading, and API gotchas.
• scripts/supir_api_probe.py for a safe import, signature, CUDA, and optional config probe.
• ../../references/checkpoints-and-environment.md before running anything that loads checkpoints.
• ../../references/troubleshooting.md for shared dependency, CUDA, Transformers/LLaVA, and checkpoint failures.

Quick decision path

1. If the user only needs names/signatures, run the probe with --signatures. Add --skip-llava when captioning is intentionally out of scope.
2. If the user has a YAML config, run the probe with --config <path> and the root checkpoint validator before create_SUPIR_model.
3. If the user wants to generate captions, verify LLavaAgent import and checkpoint variables first; do not instantiate the agent without the model path and CUDA plan.
4. If the user wants restoration output, use this sub-skill to validate API parameters, then hand off to batch or demo workflow routes.

Core APIs

• SUPIR.util.create_SUPIR_model(config_path, SUPIR_sign=None, load_default_setting=False) loads YAML, SDXL, optional SUPIR base, and Q/F checkpoints.
• SUPIRModel.batchify_denoise(x, is_stage1=False) denoises RGB tensors in [-1, 1].
• SUPIRModel.batchify_sample(...) runs the restoration sampler with prompt, CFG, control, color-fix, and seed controls.
• SUPIRModel.init_tile_vae(encoder_tile_size=512, decoder_tile_size=64) adds tiled VAE hooks.
• LLavaAgent(model_path, device='cuda', conv_mode='vicuna_v1', load_8bit=False, load_4bit=False) wraps local LLaVA captioning.
• Image helpers expect RGB inputs and generally round spatial sizes to multiples of 64 for model compatibility.

Boundary reminders

• This sub-skill describes API surfaces and validation; it does not run native end-to-end inference by default.
• CPU import checks are useful for signatures but do not validate SUPIR output.
• If multiple samples are requested through batchify_sample, the source model asserts batch size one.
• Local prompt lists are a special tiled-mode path; route detailed usage to ../interactive-demos/SKILL.md.

Safe examples

python sub-skills/python-api-and-config/scripts/supir_api_probe.py --signatures --check-cuda
python sub-skills/python-api-and-config/scripts/supir_api_probe.py --signatures --skip-llava
python sub-skills/python-api-and-config/scripts/supir_api_probe.py --config path/to/SUPIR_v0.yaml

These commands inspect imports, signatures, and config metadata only. They do not call create_SUPIR_model or load checkpoints.