Ana içeriğe geç

PRX

Model Source Revision Weight license Parity (CPU, fp32)
prx-1024-t2i-beta Photoroom/prx-1024-t2i-beta 0370504 Apache-2.0 Verified
prx-512-t2i-sft Photoroom/prx-512-t2i-sft 2996423 Apache-2.0 Verified

Catalog: prx · Models: prx-1024-t2i-beta, prx-512-t2i-sft · Task: t2i

Architecture. PRX uses the 26-layer, width-2304 T5Gemma encoder with alternating bidirectional sliding-window and full attention. A 16-layer transformer consumes 2×2 patches of the 16-channel latent, attends to the text states, and uses two-axis rotary positions. A FLUX-style 8× VAE decodes the result. Flow-match Euler sampling uses a static shift of 3.0.

Usage.

python -m imagehub.pretrained prx-512-t2i-sft checkpoints/prx-512-t2i-sft
from imagehub import pipeline

with pipeline("t2i", backend="prx-512-t2i-sft", model="checkpoints/prx-512-t2i-sft") as pipe:
    result = pipe("a product photograph of a ceramic lamp", seed=42)
    result.images[0].save("lamp.png")

Settings. Both checkpoints default to 28 steps and guidance 4.0. Their native sizes are 1024×1024 and 512×512 respectively. The supported options are seed, steps, guidance, width, height, negative_prompt, max_sequence_length (default 256), and skip_text_cleaning. Width and height are mapped to the reference aspect-ratio bins by default; disable use_resolution_binning to request dimensions that are multiples of 16 directly. Guidance 1.0 disables the unconditional pass.

Precision and validation. The 1024 checkpoint declares bfloat16 for its text encoder; the 512 checkpoint declares float32. Both repositories publish float32 transformer and VAE weights. Exact tiny-pipeline parity and published-weight parity are covered; the real-checkpoint run took 7 minutes on an M2 Max (CPU) (validation).

Out of scope. Batches, caller-supplied latents or embeddings, callbacks, LoRA, and fp16/bf16 parity are not supported.

License. Both repositories declare Apache-2.0. imagehub does not distribute weights.