Finally, a Domestic Model Supports “Code from Images”: Hands‑On Test of Doubao‑Seed‑Code for 9.9 RMB/Month

# Doubao‑Seed‑Code: Breaking Free from Claude Code

## Overview

Hello everyone,  

Over the past three months, I've felt that we’re getting closer to **fully breaking free from Claude Code**.  

ByteDance’s newly released **Doubao‑Seed‑Code model** has surpassed my expectations — not only in raw coding ability, but by introducing **a low‑cost, multi‑modal subscription model with image‑based coding support**.  

Previously, domestic programming models either lacked practical *"from screenshot to code"* capabilities or only offered demo‑quality results.  

This time, **Doubao‑Seed‑Code treats image‑to‑code as a core feature**:  
- Transforms **UI screenshots, Figma exports, or hand‑drawn sketches** into working code.  
- Performs **visual comparisons with iterative fine‑tuning fixes**.  

This closes a **long-missing gap in domestic coding models** — making development smoother and faster.

---

## TL;DR

1. **Multi‑modal support** (including image‑to‑code) — unique among current domestic AI coding models.
2. **Affordable subscription** — first month just ¥9.9, renewal ¥40 RMB.
3. **Strong coding skills** — first‑tier domestic ranking, slightly behind Claude Sonnet 4.5 / Opus 4.1 / Codex.
4. **256K context length**.
5. **Compatible** with Claude Code, Cursor, Cline, Codex CLI, etc.
6. **Native integration in TRAE (China)** — now with image‑coding capability.

---

## Live Testing

Our main evaluation: Using Doubao‑Seed‑Code inside Claude Code, focusing on **image‑to‑code performance** — currently unmatched across domestic models.

### Technical Benchmarks

According to ByteDance’s data, Doubao‑Seed‑Code scored strongly on:  
- **Terminal Bench**  
- **SWE‑Bench‑Verified‑Openhands**  
- **Multi‑SWE‑Bench‑Flash‑Openhands**  
Ranking just behind Claude Sonnet 4.5.

![image](https://blog.aitoearn.ai/content/images/2025/11/img_001-343.jpg)

It integrates deeply with the TRAE environment — **faster and more accurate problem‑solving**, topping the SWE‑Bench Verified leaderboard.

![image](https://blog.aitoearn.ai/content/images/2025/11/img_002-324.jpg)

---

## Getting Started

**Purchase:**  
Scan the QR code and try for **¥9.9** — about the cost of one coffee.  

![image](https://blog.aitoearn.ai/content/images/2025/11/img_003-306.jpg)  

If scanning fails, click “Read Full Article” to access.

**Setup in Claude Code:**

export ANTHROPIC_BASE_URL=https://ark.cn-beijing.volces.com/api/compatible

export ANTHROPIC_AUTH_TOKEN=

export ANTHROPIC_MODEL=doubao-seed-code-preview-latest

claude

After configuration, you can send an image to verify the model’s identity.

![image](https://blog.aitoearn.ai/content/images/2025/11/img_004-290.jpg)

---

# Test Cases

## Test 1 — Building a *Monument Valley*-Style 3D Scene

**Challenge:**  
Recreate a *Monument Valley*‑style isometric scene — bright, richly textured, animated, with camera controls (`OrbitControls`).

**Prompt requirements:**
- Isometric scene from reference screenshot.
- Continuous animation (rotating structure).
- Mouse‑based rotation/zoom.
- High‑quality lighting/shadows.
- **All implemented in a single code pass**.

![image](https://blog.aitoearn.ai/content/images/2025/11/img_005-261.jpg)

**Execution:**

![image](https://blog.aitoearn.ai/content/images/2025/11/img_006-237.jpg)

**Evaluation:**
- **Strengths:** Solid structure, correct 3D modeling, good texture quality.
- **Weakness:** Fidelity to the original reference not perfect.
- **Score:** **75/100**  
  *(Claude Sonnet 4.5 benchmark ~90/100)*

---

## Test 2 — From Hand Sketch to Interface

**Scenario:** Whiteboard brainstorming → instant UI.  

**Sketch input:**
![image](https://blog.aitoearn.ai/content/images/2025/11/img_007-227.jpg)  
![image](https://blog.aitoearn.ai/content/images/2025/11/img_008-211.jpg)

**Prompt:**
> Create a WeChat mini‑program coffee ordering interface — homepage and menu, based on my sketch. Dark tones, fine details, optional SVG fills. Two screens: homepage, then menu page on click. Focus on visuals, no backend functionality.

**Execution:**
![image](images/img_009.jpg)

**Evaluation:**
- Excellent multi-image understanding.
- Accurately interprets hand notes, implements **all specified elements**.
- Supplements missing details intelligently.
- Draws SVGs for richer UI.

**Score:** **80/100**  
- Deduction for aesthetics — design wasn’t fully modern.

---

## Test 3 — Pixel‑Perfect Product Recreation

**Scenario:** Reference an existing product’s design for integration.

**Input:** Grammarly homepage screenshot.

![image](https://blog.aitoearn.ai/content/images/2025/11/img_010-183.jpg)

**Prompt:**
> Replicate this page 1:1, including all elements. Missing images should be recreated as SVG.

**SVG recreation example:**
![image](https://blog.aitoearn.ai/content/images/2025/11/img_011-175.jpg)

**Evaluation:**
- SVG logo fidelity ~70%.
- Overall page recreation: **75/100**.

---

## Summary & Insights

Across all tests, Doubao‑Seed‑Code delivered:
- **Strong image understanding and code generation**.
- Multi‑image reasoning and sketch interpretation.
- Intelligent supplementation for missing assets.

**Areas for improvement:**
- Visual aesthetic polish to match global top-tier design.

For creators progressing from **sketch → prototype → polished UI**, this model is a significant productivity boost.  

Pairing tools like **[AiToEarn](https://aitoearn.ai/)** — for global AI content generation, publishing, and monetization — can take such prototypes to profitable deployment across Douyin, Kwai, WeChat, Bilibili, Xiaohongshu, Facebook, Instagram, LinkedIn, Threads, YouTube, Pinterest, and X (Twitter).  
With features like analytics and **model rankings** ([AI模型排名](https://rank.aitoearn.ai)), you can seamlessly move from code to revenue.

---

**Question:**  
Would you rate Doubao-Seed-Code’s performance **higher** or **lower** than my scores?