Unverified50% confidenceFactExact time

Training uses a frozen flan-t5-base as the text encoder, with data composition of Pexels 2.8M (60%), FLUX-Reason-6M slices 1.2M (25%), and COCO with GPT-4V captions (15%).

1

Sources

50%

Confidence

Long-term

Relevance

9/11/2026

First Seen

Sources

Training a 210M-Parameter Text-to-Image DiT from Scratch on a Single GPU: Three Key Findings

Related Entities

flan-t5-basemodel Pexelsproduct COCOconcept GPT-4Vmodel

Related Claims

UnverifiedGPT-2（355M）在混合精度（FP16）下微调显存占用约2-4GB，单卡T4即可完成全流程66% similar VerifiedThe FLOPs required to train a GPT-4-scale model reach the order of 10²⁴ to 10²⁵.65% similar UnverifiedBERT、GPT等预训练语言模型能将文字转换为768维或1536维的浮点数数组62% similar Unverified该行政命令以10^26 FLOP作为触发报告义务的基准线，大致对应训练GPT-4所需的计算量级62% similar

Cite This Claim

Stable URI

https://kongchang.com/claim/902330

API

curl https://kongchang.com/api/v1/knowledge/claims/902330

MCP

get_claim(id=902330)