generate_video() missing hires params — cannot use LTX 2.3 Spatial Latent Upscaler

[martinbu69]

Bug Description

The C API (stable-diffusion.h) defines sd_hires_params_t inside sd_vid_gen_params_t for the trained Spatial Latent Upscaler (e.g. LTX 2.3's ltx-2.3-spatial-upscaler-x2-1.1.safetensors). The ctypes struct in the Python bindings already has this field, but generate_video() does not accept or pass through the hires parameters.
Note: The constructor's upscaler_path is for ESRGAN pixel upscalers only and causes a segfault when used with the LTX spatial upscaler.

Patch

Three changes to stable_diffusion.py:

1. New parameters in generate_video() signature

hires: bool = False,
hires_path: str = "",
hires_upscaler: Union[str, SDHiresUpscaler, int, float] = "model",
hires_scale: float = 2.0,
hires_width: int = 0,
hires_height: int = 0,
hires_steps: int = 0,
hires_denoising_strength: float = 0.7,
hires_upscale_tile_size: int = 128,

2. Build hires struct before params

hires_upscaler = self._validate_and_set_input(
    hires_upscaler, SD_HIRES_UPSCALER_MAP, "hires_upscaler"
)
_hires_params = sd_cpp.sd_hires_params_t(
    enabled=hires,
    path=hires_path.encode("utf-8"),
    upscaler=hires_upscaler,
    scale=hires_scale,
    width=hires_width,
    height=hires_height,
    steps=hires_steps,
    denoising_strength=hires_denoising_strength,
    upscale_tile_size=hires_upscale_tile_size,
)

3. Pass to sd_vid_gen_params_t

hires=_hires_params

Usage

pipe.generate_video(
    prompt="A sunset over the ocean",
    hires=True,
    hires_path="/path/to/ltx-2.3-spatial-upscaler-x2-1.1.safetensors",
    hires_upscaler="model",
    hires_scale=2.0,
    hires_steps=15,
)

Tested with LTX 2.3 — produces x2 upscaled output (768×512 → 1536×1024).