LongCat-Image¶
| Model | Source | Revision | Weight license | Parity (CPU, fp32) |
|---|---|---|---|---|
longcat-image |
meituan-longcat/LongCat-Image | d2ea50b |
Apache-2.0 | Verified |
Catalog: longcat-image ·
Models: longcat-image · Task: t2i
Architecture. A Qwen2.5-VL 7B text model with 28 layers, width 3584, 28 query heads, 4 key-value heads and 16/24/24 multimodal RoPE sections first rewrites the prompt deterministically with the published generation configuration and up to 512 new tokens, then encodes it. A FLUX-style transformer has 10 double-stream blocks, 20 single-stream blocks and 24 attention heads of width 128, with no guidance embedding. The 16-channel FLUX VAE decodes the latents. The flow-match Euler scheduler uses dynamic shifting, and classifier-free guidance is renormalized by default.
Usage.
from imagehub import pipeline
with pipeline(
"t2i", backend="longcat-image", model="checkpoints/longcat-image", device="mps"
) as pipe:
result = pipe(
"a cat, oil painting",
seed=42,
enable_prompt_rewrite=True,
enable_cfg_renorm=True,
)
result.images[0].save("cat.png")
Settings. steps defaults to 50, guidance defaults to 4.5, enable_prompt_rewrite and
enable_cfg_renorm default to true, and width/height default to 1024. The other supported
options are seed, latents, negative_prompt and cfg_renorm_min. Unrecognized settings raise
an error.
Seed compatibility. On CPU, the same seed and settings give the same image as diffusers with
generator=torch.Generator("cpu").manual_seed(seed) (validation). MPS and CUDA
runs are smoke tests only; there is no claim of exact equality on these devices.
Out of scope. LongCat-Image-Edit and fp16/bf16 parity are not supported. The published weights are bfloat16; validated parity is in float32 after upcasting.
License. The weights use the Apache-2.0 license. imagehub does not distribute weights.
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.

referenceimagehub