Meissonic¶
| Model | Source | Revision | Weight license | Parity (CPU, fp32) |
|---|---|---|---|---|
meissonic |
MeissonFlow/Meissonic | 7ee5068 |
Apache-2.0 | Verified |
Catalog: meissonic ·
Model: meissonic · Task: t2i
Architecture. Meissonic is a masked image-token generator. Its CLIP text tower produces a pooled condition and 77 token states. A 1B-parameter transformer embeds a 64×64 grid of VQ tokens, compresses it to 32×32, applies 14 joint text/image blocks and 28 single-stream blocks with rotary position embeddings, restores the grid, and predicts one of 8192 codebook entries per position. The scheduler repeatedly samples tokens and remasks low-confidence positions on a cosine schedule. A VQ-VAE with 16× spatial compression decodes the completed grid.
from imagehub import pipeline
with pipeline("t2i", backend="meissonic", model="checkpoints/meissonic") as pipe:
result = pipe("a cat, oil painting", seed=42)
result.images[0].save("cat.png")
Settings. Defaults are 1024×1024, 48 steps, guidance 9, aesthetic score 6, crop (0, 0),
and temperature (2, 0). Options are seed, steps, guidance, width, height,
negative_prompt, aesthetic_score, crop_coord, and temperature. Dimensions must be multiples
of 32 for the published transformer. A temperature pair is linearly interpolated across steps;
one number falls to 0.01. Guidance at or below 1 disables the negative branch.
Memory. The full-precision transformer is 4.0 GB, text encoder 1.4 GB, and VQ-VAE 0.6 GB. The 1024px transformer operates on 1,024 compressed image tokens plus 77 text tokens per branch.
Validation and limits. CPU float32 tests compare the tokenizer, text encoder, transformer,
scheduler, VQ decoder, completed token grid, and uint8 image with the pinned official Pipeline.
The real end-to-end validation uses two steps and a narrow valid image to shorten its CPU run; it
is a parity test, not a quality demonstration. Batches, caller-provided embeddings or tokens,
callbacks, LoRA, image-to-image, inpainting, and fp16/bf16 parity are not supported.
License. The weights and implementation declare Apache-2.0. imagehub does not distribute weights. See validation.