HeyHomees Lab Notes · No. 13 · inclusionAI · Published October 30, 2026
By HeyHomees Newsroom ·
In 30 seconds
- Full-parameter Z-Image runs in 0.30 s per image at 2 NFE on a single A100 (BF16)【s2】.
- Throughput is 5.41 images/s at 1 NFE on the same hardware【s2】.
What is it?
Intrinsic Distribution Matching (IDM) combines internal distribution matching distillation with explicit moment matching in the generator’s EMA feature space, requiring no external encoder【s1】【s2】.
What does it need to run?
Benchmarks were measured on a single A100 GPU using BF16 precision; LoRA weights are merged before timing【s2】.
By the numbers
- Chinese LongText-Bench score: 94.25%
- English LongText-Bench score: 90.20%
- Latency at 2 NFE (Z-Image): 0.30 s
- Throughput at 1 NFE (Z-Image): 5.41 images/s
- FID at 2 NFE (Z-Image): 6.71
The honest read
The reported 94.25% and 90.20% scores are from the authors' evaluation on the LongText-CN and LongText-EN benchmarks; they reflect the model’s ability to render long prompts but do not measure other aspects such as diversity or safety. Latency and throughput numbers are measured on a single A100 GPU in BF16 precision and include merged LoRA weights, so they may differ on other hardware or with different precision settings. The FID value shows image quality at 2 NFE but lower is better, and the figure is presented without a direct comparison to baseline scores beyond the paper’s table.
What we don't know yet
- Training data and total compute budget are not specified.
- Performance on other benchmarks beyond the five reported metrics is not provided.
- Exact model download size is not mentioned.
Sources
- Paper (arXiv abstract) (primary)
- Code, GitHub
Researched and drafted with AI from the sources above; every figure is checked against them by code before it is published, and a line that fails is dropped.