GitLab

* A1111 import: Set VAE name

This patch sets the VAE name for the `VAELoader` when present in the png
metadata.

* A1111 import: Skip all hashes

When importing from A1111 the parsing assumes that values of a key will
never contain a ":", which is not correct.

There are 2 cases where we can have ":" in the value:

- Inside a string. E.g.:
  Lora hashes: "xl_more_art-full_v1: fe3b4816be83, add-detail-xl: 9c783c8ce46c"

- When the value is a json dictionary. E.g.:
  Hashes: {"vae": "63aeecb90f", "embed:negativeXL_D": "fff5d51ab6"}

This patch changes how we parse the metadata to take those 2 cases into
account and also skips the following additional keys that are present in
some Forge images:

- Version
- VAE hash
- TI hashes
- Lora hashes
- Hashes

* A1111 import: Parse Hires steps

This patch parses the `Hires steps` parameter that is part of the High
Resolution Upscale configuration when it  is present, and fallbacks to
the one from the `samplerNode` (like the code currently does) if it's
not present.

This node enables the creation of nodes to change the guider/denoiser and
the noise algorithm.

calc_cond_batch can take an arbitrary amount of cond inputs.

Added a calc_cond_uncond_batch wrapper with a warning so custom nodes
won't break.

This is the code to load the model and inference it with only a text
prompt. This commit does not contain the nodes to properly use it with an
image input.

This supports both the original SD1 instructpix2pix model and the
diffusers SDXL one.

Lower peak memory usage when changing model.

Fix case where model weights would be unloaded and reloaded.