Unverified50% confidenceBenchmarkExact time
On the ImageNet-256 dataset, with NFE of 25 and no guidance, the time-lag-measurement-based gain selection method reduces FID from 28.0 to 12.2.
1
Sources
50%
Confidence
Long-term
Relevance
10/9/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedImage models can maintain INT8 or FP8 precision when offloaded, and Z-Image generation time on a B200 with a 10 GiB VRAM limit dropped from 38.4 seconds to 2.3 seconds65% similarVerifiedMXFP4 mitigates numerical overflow and precision degradation at 4-bit low precision by sharing a single scaling factor within small blocks of data (typically 32 elements), theoretically reducing memory footprint and bandwidth requirements by approximately 4× compared to FP16.61% similarVerifiedQuantization reduces FP32 (32-bit) weights to INT4 (4-bit), achieving a theoretical compression ratio of 8×, but introduces precision loss.61% similarUnverifiedDouble quantization further quantizes the FP32 scaling factors shared among every 64 parameters into FP8 format, reducing the quantization overhead per parameter from 0.5 bits to approximately 0.127 bits.61% similarUnverifiedCPU image encoding consumes zero VRAM and recognizes one image in about 5 seconds, which is an advantage for scenarios where 24 GB of VRAM is already fully occupied by the model.61% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/990728API
curl https://kongchang.com/api/v1/knowledge/claims/990728MCP
get_claim(id=990728)