Ana içeriğe geç

Latent Diffusion Model

Model Source Revision Weight license Parity (CPU, fp32)
ldm-text2im-large-256 CompVis/ldm-text2im-large-256 30de525 Apache-2.0 Verified

Catalog: ldm · Models: ldm-text2im-large-256 · Task: t2i

Architecture. A 32×32×4 latent U-Net conditioned by a 32-layer LDMBert encoder and decoded by an f=8 KL autoencoder. The published sampler is DDIM.

Usage. python -m imagehub.pretrained ldm-text2im-large-256 checkpoints/ldm

Settings. steps (50), guidance (1.0), width/height (256, multiples of 8), negative_prompt, num_images, seed, latents, and eta (0.0 only).

Validation. CPU float32 parity is established against diffusers on CPU; all 5 real-checkpoint tests passed with exactly 0 difference. Safety checkers, watermarking, img2img, and stochastic DDIM sampling are out of scope.

Reference and imagehub, side by side

Drag the slider. Left is the recorded reference, right is imagehub, from the same seed, on CPU in float32.