# Dongxi NLP revises an LLM diagram through 30 Seedream 5.5 edits

> Original title: 用 Seedream 5.5 画 LLM 原理图：同一张图连续改 30 轮

- Source: Dongxi NLP
- Published: 2026-11-10T05:49:55.000Z
- Bad Influence: https://badinfluence.news/items/x52ih6bp8jzldbgio8bm24m8r
- Original: https://x.com/dongxi_nlp/status/2120031178006695998

## Summary

Dongxi NLP says it used Doubao's Seedream 5.5 image model to draw LLM inference and post-training diagrams. It says one diagram went through 30 consecutive edits in the same chat, with the largest single-round change covering 1.02% of the image and a median of 0.20%. The author describes the workflow as rounds of text instructions, region marking, and erasing, with the Qwen3-0.6B example holding the earlier content in place.

## Post

https://x.com/i/article/2119890649227120640