nemo_automodel.components.models.hunyuan_image3.release

The release’s VAE and prompt format, loaded from the checkpoint’s remote code.

The release ships its VAE, tokenizer wrapper and image processor as remote code inside the checkpoint (trust_remote_code); they are not vendored here. Preprocessing and sampling both go through this module, so the latents and token sequences a model is trained on are the ones it is sampled with.

Module Contents

Classes

Name	Description
`HunyuanImage3PromptTokenizer`	Token ids of the release’s text-to-image sequence around the image span.

Functions

Name	Description
`load_release_vae`	Build the release VAE (`AutoencoderKLConv3D`) and load the checkpoint’s `vae.*` weights.

API

class nemo_automodel.components.models.hunyuan_image3.release.HunyuanImage3PromptTokenizer(
    wrapper: typing.Any,
    image_processor: typing.Any,
    image_base_size: int,
    sequence_template: str
)

Token ids of the release’s text-to-image sequence around the image span.

Parameters:

wrapper

Any

The release TokenizerWrapper.

image_processor

Any

The release HunyuanImage3ImageProcessor (resolution group and image token grid).

image_base_size

int

image_base_size of the checkpoint config.

sequence_template

str

sequence_template of the checkpoint’s generation config.

nemo_automodel.components.models.hunyuan_image3.release.HunyuanImage3PromptTokenizer.__call__(
    prompt: str,
    height: int,
    width: int
) -> dict[str, torch.Tensor]

Return the token ids around the image span for one prompt and image size.

The sequence is the release’s gen_image chat template with classifier-free guidance (bot_task auto, no system prompt), which is also what its generate_image samples with.

Returns: dict[str, torch.Tensor]

prompt_input_ids / uncond_prompt_input_ids: 1D long ids before the image span, ending in

nemo_automodel.components.models.hunyuan_image3.release.HunyuanImage3PromptTokenizer.from_pretrained(
    model_dir: str,
    config: typing.Any
) -> nemo_automodel.components.models.hunyuan_image3.release.HunyuanImage3PromptTokenizer

classmethod

Load the release tokenizer wrapper and image processor from model_dir.

Parameters:

model_dir

str

Local checkpoint directory.

config

Any

The checkpoint config (needs image_base_size).

nemo_automodel.components.models.hunyuan_image3.release.HunyuanImage3PromptTokenizer.target_size(
    width: int,
    height: int
) -> tuple[int, int]

Snap width x height to the release’s resolution group (33 ratios around image_base_size).

Returns: tuple[int, int]

(width, height) of the closest size the release generates; its <img_ratio_*> token names it.

nemo_automodel.components.models.hunyuan_image3.release.load_release_vae(
    model_dir: str,
    vae_config: dict[str, typing.Any],
    device: str | torch.device
) -> torch.nn.Module

Build the release VAE (AutoencoderKLConv3D) and load the checkpoint’s vae.* weights.

Parameters:

model_dir

str

Local checkpoint directory.

vae_config

dict[str, Any]

The vae section of the checkpoint config.

device

str | torch.device

Device for the returned VAE.

Returns: torch.nn.Module

The VAE in fp32 and eval mode.