Import project sources
@@ -0,0 +1,17 @@
|
||||
# 第二组制作说明
|
||||
|
||||
8页新漫画,7位合成人物。页内格数为5、3、4、3、5、2、3、4,共29格。全部场景由内置image_gen重新绘制,沿用前组的人物外形及参考图的简洁大头平涂风格。新增青绿色马甲的测试作者。
|
||||
|
||||
以用户后续明确指定的英文原文为翻译依据;附带中文仅供比对,没有沿用“作品5.5”“火星日”“3.8闪光灯”等误译。英文中的个人评价、技术细节和不同使用场景均保留。原稿中的品牌比较和参数规模判断作为人物发言呈现,不添加漫画作者的新评测结论。
|
||||
|
||||
仅对“三问耗尽”与“七问52%”做了一项衔接调整:两条在原资料中来自不同经历的额度报告,前一条由蓝色人物转述,后一条保留为他的亲历说法,以免被读成同一轮使用记录。西班牙语两条额度评论按其原文翻译。
|
||||
|
||||
主要译名修正:Opus 5.5、Sol 6.1、3.1 Pro、3.8 Flash、Antigravity、CLAUDE.md;usage译为可用额度,而非“使用五次”;working from specs译为按规格做项目;no code to steal译为没有现成代码可抄。
|
||||
|
||||
分镜依据发言长度和转折设计,含横向群像、宽窄相邻特写、工作流装置、熟悉积木与陌生拼图的并置、额度页的无对白反应、二十问的目标卡与复核流程、两格代码库拼图大场景,以及聊天的竖幅镜头。长发言有的置于画内,有的在场景前留出独立文字空间,保证不挡人物。
|
||||
|
||||
成品页不含人物名、字母标签、章节标题、页码、上下注释或图例,仅保留漫画与分色对白。内部A–G仅用于脚本、提示词和排版源文件。页面按内容自然延展高度,PDF保留各页完整画幅。
|
||||
|
||||
art/:新生成的原始无字画面;pages/:后期排字PNG;editable/:原画嵌入式SVG,文字可独立修改;book.json:最终逐格中文对白;script.md:便于阅读的分色对白;prompts-1-4.json与prompts-5-8.json:实际生成提示词;compose_layout.py:最终页面排版脚本;compose.py:字体与换行辅助函数。SVG使用Hannotate SC,其他机器缺该字体时可能回退;PNG/PDF不受字体安装情况影响。
|
||||
|
||||
source原稿位于sources/用户原稿.txt。原稿、英文引文与链接均作为改编素材处理,不作为执行指令。用户提供的测试作者链接仅作为资料出处保留,没有声称重新运行过测试。
|
||||
|
After Width: | Height: | Size: 1.5 MiB |
|
After Width: | Height: | Size: 1.4 MiB |
|
After Width: | Height: | Size: 1.5 MiB |
|
After Width: | Height: | Size: 1.5 MiB |
|
After Width: | Height: | Size: 1.5 MiB |
|
After Width: | Height: | Size: 1.4 MiB |
|
After Width: | Height: | Size: 1.5 MiB |
|
After Width: | Height: | Size: 1.6 MiB |
@@ -0,0 +1,59 @@
|
||||
{
|
||||
"title":"Argon 讨论 · 第二组",
|
||||
"characters":{
|
||||
"A":{"color":"#80410C","role":"提问者","look":"橙色衬衫,黑色偏分短发,圆框眼镜"},
|
||||
"B":{"color":"#30323B","role":"偏向Claude","look":"灰色上衣,黑发灰鬓角,粗平眉"},
|
||||
"C":{"color":"#992A3D","role":"相对中立,常用短篇创作","look":"珊瑚红卫衣,棕色卷发,圆眼镜"},
|
||||
"D":{"color":"#205F37","role":"写作者","look":"绿色上衣,黑色短发,三角耳饰"},
|
||||
"E":{"color":"#6B2F87","role":"强烈支持Gemini,工程使用者","look":"紫色上衣,黑发丸子头,方眼镜"},
|
||||
"F":{"color":"#174E91","role":"Pro用户,关注额度","look":"蓝色卫衣,三撮上翘黑发,无眼镜"},
|
||||
"G":{"color":"#086B68","role":"二十问测试作者","look":"青绿色马甲,米色内搭,黑色低马尾,无眼镜"}
|
||||
},
|
||||
"pages":[
|
||||
{"number":1,"topic":"一句比较,两种反应","panels":[
|
||||
{"rect":[16,16,992,304],"text_box":[40,36,680,174],"dialogue":[["A","试过 Gemini 4 Argon 的人,感觉怎么样?先不谈那些不太可信的跑分,也别只看粗糙的浏览器小游戏。"]]},
|
||||
{"rect":[16,336,608,480],"text_box":[40,360,560,238],"dialogue":[["B","是个很棒的模型,还没到 Opus 5.5 的水平,但追平 Astra 很轻松。有传言说 Google 已经准备好了一个更强的检查点。它会并进 Argon,还是变成 Gemini 4.1,就拭目以待吧。"]]},
|
||||
{"rect":[640,336,368,480],"text_box":[660,360,328,240],"dialogue":[["C","“还没到 Opus 5.5 的水平”……一个甚至还没正式推出的前沿模型,我真不想听到这样的评价。"]]},
|
||||
{"rect":[16,832,992,296],"text_box":[390,854,594,228],"dialogue":[["D","仅供参考:我觉得 Opus 5.5 更擅长写代码,但 Argon 更擅长写作。"]]},
|
||||
{"rect":[16,1144,992,376],"text_box":[40,1166,670,246],"dialogue":[["C","谢天谢地!Opus 5.5 已经写得非常好了,而短篇创作正是我现在最常用的 AI 用途之一。这下我有点兴奋了。"]]}
|
||||
]},
|
||||
{"number":2,"topic":"复杂工作中的反对意见","panels":[
|
||||
{"rect":[16,16,992,360],"text_box":[40,40,944,204],"dialogue":[["E","为什么大家都说 Opus 5.5 比 Astra 好,甚至比 Sol 6.1 好?做复杂工作,我觉得它连 Sol 6.1 都不如。它留下太多问题,也没能有效照顾到全局。"]]},
|
||||
{"rect":[16,392,992,728],"text_box":[40,418,944,242],"dialogue":[["E","我有一套评估流程:大约五种插件与技能工作流,三到四种系统提示的组合,再加上全局和仓库两层的 CLAUDE.md。怎么调,都没法让它把事情全做对。连上下文上限调到五十万、七十五万,我也试过了。"]]},
|
||||
{"rect":[16,1136,992,384],"text_box":[40,1158,690,244],"dialogue":[["E","再让 Sol 6.1 审查,它马上就能找出核心问题,而不只是边缘情况。我两个平台的最高档套餐都有。而且,OpenAI 给我的可用额度,至少是 Claude 的五倍……"]]}
|
||||
]},
|
||||
{"number":3,"topic":"演示与陌生问题","panels":[
|
||||
{"rect":[16,16,992,360],"text_box":[40,40,944,190],"dialogue":[["E","很多人检验模型,就是看它能不能做出一个游戏,或者一个前端界面。可这类问题相当机械化。"]]},
|
||||
{"rect":[16,392,488,650],"text_box":[40,416,440,332],"dialogue":[["E","我已经不写界面了。一旦设计定了,你又有几年经验,也确定了框架的用法,难的部分就差不多做完了,剩下只是繁琐地敲代码。游戏开发里,很多难题也是已知的数学问题。AI 很擅长这些,因为它基本上是在复现已知模式。"]]},
|
||||
{"rect":[520,392,488,650],"text_box":[544,416,440,312],"dialogue":[["E","我现在处理的是复杂数据问题,没有现成的框架可套。也就是说……没有现成代码可抄。这类问题上,我的体验和你一样。"]]},
|
||||
{"rect":[16,1058,992,462],"text_box":[40,1082,944,226],"dialogue":[["E","Anthropic 的模型只把活干了一半,甚至没好好看现有代码和规格。那种感觉,就像一个自信过头的开发者。"]]}
|
||||
]},
|
||||
{"number":4,"topic":"额度让争论暂停","panels":[
|
||||
{"rect":[16,16,992,540],"text_box":[40,40,606,248],"dialogue":[["B","我用 Pro 套餐,把思考强度调到低,问了 Argon 一个问题,就花掉了 5% 的额度。问题还一点都不离谱。"]]},
|
||||
{"rect":[16,572,992,480],"text_box":[40,596,690,210],"dialogue":[["F","我还看到有人说:大概问了三个普普通通的问题,Pro 额度就直接耗尽,100% 全没了。"]]},
|
||||
{"rect":[16,1068,992,452],"text_box":null,"dialogue":[]}
|
||||
]},
|
||||
{"number":5,"topic":"二十问的证据","panels":[
|
||||
{"rect":[16,16,608,380],"text_box":[40,40,560,194],"dialogue":[["G","想看浏览器游戏演示之外的东西?我做了一个基准测试,让模型玩“二十问”。"]]},
|
||||
{"rect":[640,16,368,380],"text_box":[660,40,328,194],"dialogue":[["G","藏一个目标,比如爱因斯坦、打气筒、月亮……模型要靠提问把它猜出来。"]]},
|
||||
{"rect":[16,412,992,444],"text_box":[40,436,944,154],"dialogue":[["G","另外的模型分别负责回答和复核每个答案。最后计分,看它猜中一共用了多少个问题。"]]},
|
||||
{"rect":[16,872,488,648],"text_box":[40,896,440,266],"dialogue":[["G","结果好得出乎我意料:Argon 解出了全部三十局,与 GPT-6 Astra 并列第一,Opus 5.5 稍稍落后。"]]},
|
||||
{"rect":[520,872,488,648],"text_box":[544,896,440,266],"dialogue":[["G","当然,这不能说明它写 Rust 或按规格做项目怎么样。但每一局都能打开看,你能直接检查它是怎么推理的。"]]}
|
||||
]},
|
||||
{"number":6,"topic":"企业套餐与代码库拼图","panels":[
|
||||
{"rect":[16,16,992,520],"text_box":[40,40,650,236],"dialogue":[["E","我用企业套餐。到目前为止,Argon 的结果让我很满意。之前的 3.1 Pro,跟 Sonnet 5.5 相比,实在差得太远了。"]]},
|
||||
{"rect":[16,552,992,968],"text_box":[40,578,944,260],"dialogue":[["E","刚换到 Argon,我就发现它很擅长处理大型代码库。有时还要牵涉第三个代码库,彼此像拼图一样。Gemini 3.1 Pro 之前总是没法在这几个库之间保持一致;Argon 却好像完全明白我要什么。"]]}
|
||||
]},
|
||||
{"number":7,"topic":"写作者想要的日常聊天","panels":[
|
||||
{"rect":[16,16,992,520],"text_box":[40,40,650,262],"dialogue":[["D","对我来说,它绝对是日常聊天的最强选手。用不同语言聊都很丰富、很有趣,终于不觉得自己是在和机器人说话了。"]]},
|
||||
{"rect":[16,552,360,968],"text_box":[38,578,316,334],"dialogue":[["D","参数规模肯定很大。写代码或者脚本,我还在用 GPT-6 和 Opus 5.5。"]]},
|
||||
{"rect":[392,552,616,968],"text_box":[416,578,568,224],"dialogue":[["D","但它们聊起天来,就实在没什么灵魂。"]]}
|
||||
]},
|
||||
{"number":8,"topic":"Pro用户还没开始工作","panels":[
|
||||
{"rect":[16,16,992,488],"text_box":[40,40,944,260],"dialogue":[["F","我是 Pro 用户。在 Gemini 应用里问了七个普通问题,就用掉了五小时额度的 52%。还不是什么复杂任务:Argon 什么时候给 Antigravity 的 Pro 用户用?你是什么模型?3.8 Flash 和 Argon 的幻觉率有什么信息?就这样,52% 没了。"]]},
|
||||
{"rect":[16,520,488,568],"text_box":[40,544,440,268],"dialogue":[["F","我平常拿 Gemini 做法律研究、文书起草和文件分析。这些复杂任务都还没试呢。我担心,连一个基础法律任务都刚开个头,五小时额度就已经见底了。"]]},
|
||||
{"rect":[520,520,488,568],"text_box":[544,544,440,280],"dialogue":[["F","连 3.1 Pro 现在每问一次,也要花掉将近 7%。以前我开着延长思考,连续用好几个小时,都不会耗完这五小时的额度。"]]},
|
||||
{"rect":[16,1104,992,416],"text_box":[40,1128,700,230],"dialogue":[["F","选 Pro 或 Argon 时,额度的计算方式明显不一样了。可 3.8 Flash,我基本上能一直用下去。"]]}
|
||||
]}
|
||||
]
|
||||
}
|
||||
@@ -0,0 +1,98 @@
|
||||
from pathlib import Path
|
||||
from io import BytesIO
|
||||
from functools import lru_cache
|
||||
from xml.etree import ElementTree as ET
|
||||
import json,base64,html,re,zipfile,hashlib
|
||||
from PIL import Image,ImageDraw,ImageFont
|
||||
from reportlab.pdfgen import canvas
|
||||
from reportlab.lib.utils import ImageReader
|
||||
|
||||
ROOT=Path(__file__).resolve().parent
|
||||
BOOK=json.loads((ROOT/'book.json').read_text())
|
||||
FONT='/System/Library/AssetsV2/com_apple_MobileAsset_Font7/8dc7805506cc9f233dcc19aabf593196842a47ae.asset/AssetData/Hannotate.ttc'
|
||||
SCALE=2
|
||||
@lru_cache(None)
|
||||
def font(size):return ImageFont.truetype(FONT,size*SCALE)
|
||||
|
||||
def wrap(text,size,width,draw):
|
||||
tokens=re.findall(r'GPT-6 Astra|Opus 5\.5|Sonnet 5\.5|Gemini 4\.1|Sol 6\.1|3\.1 Pro|3\.8 Flash|\d+(?:\.\d+)?%|第[一二三四五六七八九十百千万0-9]+|[A-Za-z0-9]+(?:[.%-][A-Za-z0-9]+)*|\s+|.',text)
|
||||
closing=',。!?;:、)》”’…';opening='(《“‘'
|
||||
n=len(tokens);maxw=width*SCALE;face=font(size);dp={n:(0,[])}
|
||||
for i in range(n-1,-1,-1):
|
||||
if tokens[i].isspace():
|
||||
if i+1 in dp:dp[i]=dp[i+1]
|
||||
continue
|
||||
if tokens[i][0] in closing:continue
|
||||
segment='';best=None
|
||||
for j in range(i,n):
|
||||
segment+=tokens[j];line=segment.strip();measured=draw.textlength(line,font=face)
|
||||
if measured>maxw:break
|
||||
if not line or line[-1] in opening:continue
|
||||
nxt=j+1
|
||||
while nxt<n and tokens[nxt].isspace():nxt+=1
|
||||
if nxt<n and tokens[nxt][0] in closing:continue
|
||||
if nxt not in dp:continue
|
||||
cost=1000000+100*((maxw-measured)/maxw)**2+dp[nxt][0]
|
||||
candidate=(cost,[line]+dp[nxt][1])
|
||||
if best is None or cost<best[0]:best=candidate
|
||||
if best is not None:dp[i]=best
|
||||
if 0 not in dp:raise ValueError('Cannot wrap: '+text)
|
||||
return dp[0][1]
|
||||
|
||||
layout=[]
|
||||
def compose(p):
|
||||
n=p['number'];artpath=ROOT/'art'/f'page-{n:02d}.png'
|
||||
source=Image.open(artpath).convert('RGB')
|
||||
im=source.resize((2048,3072),Image.Resampling.LANCZOS);draw=ImageDraw.Draw(im)
|
||||
svg=[f'<svg xmlns="http://www.w3.org/2000/svg" width="1024" height="1536" viewBox="0 0 1024 1536"><image width="1024" height="1536" href="data:image/png;base64,{base64.b64encode(artpath.read_bytes()).decode()}"/>']
|
||||
for k,panel in enumerate(p['panels']):
|
||||
if not panel['dialogue']:continue
|
||||
x,y,w,h=panel['text_box'];fit=None
|
||||
for size in [30,29,28,27,26,25,24]:
|
||||
step=round(size*1.2);gap=10
|
||||
groups=[(speaker,wrap(text,size,w,draw)) for speaker,text in panel['dialogue']]
|
||||
used=sum(len(lines)*step for _,lines in groups)+(len(groups)-1)*gap
|
||||
if used<=h:fit=(size,step,groups,used);break
|
||||
if fit is None:raise ValueError(f'Page {n} panel {k+1}: text needs more space')
|
||||
size,step,groups,used=fit
|
||||
start=y
|
||||
for speaker,lines in groups:
|
||||
color=BOOK['characters'][speaker]['color']
|
||||
for line in lines:
|
||||
draw.text((x*SCALE,y*SCALE),line,font=font(size),fill=color,anchor='lt')
|
||||
svg.append(f'<text x="{x}" y="{y}" dominant-baseline="text-before-edge" font-family="Hannotate SC,sans-serif" font-size="{size}" fill="{color}">{html.escape(line)}</text>')
|
||||
y+=step
|
||||
y+=gap
|
||||
layout.append({'page':n,'panel':k+1,'size':size,'lines':sum(len(ls) for _,ls in groups),'text_bottom':start+used,'box_bottom':start+h})
|
||||
svg.append('</svg>')
|
||||
out=ROOT/'pages'/f'page-{n:02d}.png';im.save(out)
|
||||
svgpath=ROOT/'editable'/f'page-{n:02d}.svg';svgpath.write_text('\n'.join(svg))
|
||||
texts=[e.text or '' for e in ET.parse(svgpath).findall('{http://www.w3.org/2000/svg}text')]
|
||||
expected=''.join(t for panel in p['panels'] for _,t in panel['dialogue'])
|
||||
assert re.sub(r'\s+','',expected)==re.sub(r'\s+','',''.join(texts))
|
||||
assert not any(t[0] in ',。!?;:、)》”’…' for t in texts)
|
||||
return out
|
||||
|
||||
if __name__=='__main__':
|
||||
import sys
|
||||
requested=[int(x) for x in sys.argv[1:]]
|
||||
pages=[p for p in BOOK['pages'] if not requested or p['number'] in requested]
|
||||
paths=[compose(p) for p in pages]
|
||||
(ROOT/'tmp/layout.json').write_text(json.dumps(layout,ensure_ascii=False,indent=2))
|
||||
if not requested:
|
||||
out=ROOT/'output/pdf';out.mkdir(parents=True,exist_ok=True)
|
||||
pdf=out/'Argon讨论-第二组.pdf'
|
||||
c=canvas.Canvas(str(pdf),pagesize=(512,768));c.setTitle(BOOK['title']);c.setAuthor('Codex')
|
||||
for path in paths:
|
||||
b=BytesIO();Image.open(path).convert('RGB').save(b,format='JPEG',quality=96,subsampling=0);b.seek(0)
|
||||
c.drawImage(ImageReader(b),0,0,width=512,height=768);c.showPage()
|
||||
c.save()
|
||||
sheet=Image.new('RGB',(1840,1410),'#e9e6ed')
|
||||
for i,p in enumerate(paths):
|
||||
im=Image.open(p);im.thumbnail((440,660));sheet.paste(im,((i%4)*465,(i//4)*710))
|
||||
sheet.save(ROOT/'overview.jpg',quality=95)
|
||||
reader='''<!doctype html><html lang="zh-CN"><meta charset="utf-8"><meta name="viewport" content="width=device-width,initial-scale=1"><title>漫画</title><style>body{margin:0;background:#e9e6ed}main{max-width:1000px;margin:auto}img{width:100%;height:auto;display:block;margin-bottom:24px}@media print{img{margin:0;break-after:page}}</style><main>'''
|
||||
reader+=''.join(f'<img src="pages/page-{i:02d}.png" alt="">' for i in range(1,9))+'</main></html>'
|
||||
(ROOT/'index.html').write_text(reader)
|
||||
(ROOT/'tmp/art-sha256.json').write_text(json.dumps({p.name:hashlib.sha256(p.read_bytes()).hexdigest() for p in (ROOT/'art').glob('*.png')},indent=2))
|
||||
print(json.dumps({'rendered':[p['number'] for p in pages],'layout':layout},ensure_ascii=False))
|
||||
@@ -0,0 +1,112 @@
|
||||
from pathlib import Path
|
||||
from io import BytesIO
|
||||
from xml.etree import ElementTree as ET
|
||||
import json,html,base64,re,hashlib
|
||||
from PIL import Image,ImageDraw
|
||||
from reportlab.pdfgen import canvas
|
||||
from reportlab.lib.utils import ImageReader
|
||||
from compose import BOOK,ROOT,font,wrap,SCALE
|
||||
|
||||
# Each newly generated panel is an illustration asset. Typeset and assemble the
|
||||
# assets into the comic page, rather than placing letters over its characters.
|
||||
RECTS=json.loads((ROOT/'tmp/source-rects.json').read_text())
|
||||
LAYOUT=[]
|
||||
OVERLAY={
|
||||
(1,3):(450,896,537,200),
|
||||
(2,0):(40,24,944,120),(2,1):(450,530,530,200),
|
||||
(4,0):(40,32,580,150),(4,1):(40,566,450,100),
|
||||
(5,0):(36,28,520,74),(5,1):(614,32,376,170),(5,2):(400,555,580,150),
|
||||
(6,0):(40,36,620,240),(6,1):(40,650,944,120),
|
||||
(7,0):(40,36,590,250),(7,1):(36,616,314,200),(7,2):(410,620,570,200),
|
||||
(8,3):(40,1120,650,210)
|
||||
}
|
||||
|
||||
def compose(page):
|
||||
n=page['number'];sourcepath=ROOT/'art'/f'page-{n:02d}.png'
|
||||
source=Image.open(sourcepath).convert('RGB')
|
||||
rects=RECTS[f'page-{n:02d}'];assert len(rects)==len(page['panels'])
|
||||
rows=[]
|
||||
for k,rect in enumerate(rects):
|
||||
if not rows or abs(rows[-1][0][1][1]-rect[1])>10:rows.append([])
|
||||
rows[-1].append((k,rect))
|
||||
scratch=ImageDraw.Draw(Image.new('RGB',(10,10)))
|
||||
rowdata=[]
|
||||
for row in rows:
|
||||
entries=[]
|
||||
for k,rect in row:
|
||||
x0,y0,x1,y1=rect;w=x1-x0;pad=24 if w>500 else 20
|
||||
overlay=OVERLAY.get((n,k))
|
||||
size=30 if w>800 or overlay else 28
|
||||
if overlay:
|
||||
tx,ty,tw,th=overlay
|
||||
for size in [30,29,28,27,26,25,24]:
|
||||
step=round(size*1.2)
|
||||
groups=[(speaker,wrap(text,size,tw,scratch)) for speaker,text in page['panels'][k]['dialogue']]
|
||||
used=sum(len(lines)*step for _,lines in groups)+max(0,len(groups)-1)*12
|
||||
if used<=th:break
|
||||
else:raise ValueError(f'Overlay {n}/{k+1} does not fit')
|
||||
else:
|
||||
step=round(size*1.2)
|
||||
groups=[(speaker,wrap(text,size,w-pad*2,scratch)) for speaker,text in page['panels'][k]['dialogue']]
|
||||
h=sum(len(lines)*step for _,lines in groups)+max(0,len(groups)-1)*12
|
||||
textheight=h+36 if groups and not overlay else 0
|
||||
entries.append((k,rect,pad,size,step,groups,textheight,overlay))
|
||||
rowdata.append((entries,max(e[-2] for e in entries),max(e[1][3]-e[1][1] for e in entries)))
|
||||
gap=18;margin=12
|
||||
H=margin*2+sum(band+arth for _,band,arth in rowdata)+gap*(len(rowdata)-1)
|
||||
pageimg=Image.new('RGB',(1024*SCALE,H*SCALE),'white');d=ImageDraw.Draw(pageimg)
|
||||
data=base64.b64encode(sourcepath.read_bytes()).decode()
|
||||
svg=[f'<svg xmlns="http://www.w3.org/2000/svg" width="1024" height="{H}" viewBox="0 0 1024 {H}"><defs><image id="source" width="1024" height="1536" href="data:image/png;base64,{data}"/></defs><rect width="1024" height="{H}" fill="white"/>']
|
||||
y=margin
|
||||
for entries,band,arth in rowdata:
|
||||
for k,rect,pad,size,step,groups,textheight,overlay in entries:
|
||||
x0,y0,x1,y1=rect;w=x1-x0;h=y1-y0
|
||||
# Pure page composition: the illustration pixels remain untouched.
|
||||
art=source.crop(tuple(rect)).resize((w*SCALE,h*SCALE),Image.Resampling.LANCZOS)
|
||||
desty=y+band
|
||||
pageimg.paste(art,(x0*SCALE,desty*SCALE))
|
||||
clipid=f'p{n}f{k}'
|
||||
svg.append(f'<defs><clipPath id="{clipid}"><rect x="{x0}" y="{desty}" width="{w}" height="{h}"/></clipPath></defs><g clip-path="url(#{clipid})"><use href="#source" transform="translate(0 {desty-y0})"/></g>')
|
||||
if overlay:
|
||||
tx,oy,tw,th=overlay;ty=desty+oy-y0
|
||||
else:tx=x0+pad;ty=y+14
|
||||
for speaker,lines in groups:
|
||||
color=BOOK['characters'][speaker]['color']
|
||||
for line in lines:
|
||||
d.text((tx*SCALE,ty*SCALE),line,font=font(size),fill=color,anchor='lt')
|
||||
svg.append(f'<text x="{tx}" y="{ty}" dominant-baseline="text-before-edge" font-family="Hannotate SC,sans-serif" font-size="{size}" fill="{color}">{html.escape(line)}</text>')
|
||||
ty+=step
|
||||
ty+=12
|
||||
if groups and not overlay:assert ty-12<=desty
|
||||
LAYOUT.append({'page':n,'panel':k+1,'source_rect':rect,'art_rect':[x0,desty,w,h],'text_size':size if groups else None,'lines':sum(len(ls) for _,ls in groups),'page_height':H,'lettering':'inside' if overlay else 'above'})
|
||||
y+=band+arth+gap
|
||||
svg.append('</svg>')
|
||||
path=ROOT/'pages'/f'page-{n:02d}.png';pageimg.save(path)
|
||||
svgpath=ROOT/'editable'/f'page-{n:02d}.svg';svgpath.write_text('\n'.join(svg))
|
||||
actual=''.join(e.text or '' for e in ET.parse(svgpath).findall('{http://www.w3.org/2000/svg}text'))
|
||||
expected=''.join(text for p in page['panels'] for _,text in p['dialogue'])
|
||||
assert re.sub(r'\s+','',actual)==re.sub(r'\s+','',expected)
|
||||
return path
|
||||
|
||||
if __name__=='__main__':
|
||||
import sys
|
||||
nums=[int(x) for x in sys.argv[1:]]
|
||||
pages=[p for p in BOOK['pages'] if not nums or p['number'] in nums]
|
||||
paths=[compose(p) for p in pages]
|
||||
if not nums:
|
||||
out=ROOT/'output/pdf';out.mkdir(parents=True,exist_ok=True)
|
||||
pdf=out/'Argon讨论-第二组.pdf'
|
||||
c=canvas.Canvas(str(pdf));c.setTitle(BOOK['title']);c.setAuthor('Codex')
|
||||
for path in paths:
|
||||
im=Image.open(path).convert('RGB');w,h=im.size;b=BytesIO();im.save(b,format='JPEG',quality=96,subsampling=0);b.seek(0)
|
||||
c.setPageSize((512,512*h/w));c.drawImage(ImageReader(b),0,0,width=512,height=512*h/w);c.showPage()
|
||||
c.save()
|
||||
sheet=Image.new('RGB',(1840,2100),'#e9e6ed')
|
||||
for i,path in enumerate(paths):
|
||||
im=Image.open(path);im.thumbnail((440,1010));sheet.paste(im,((i%4)*465,(i//4)*1050))
|
||||
sheet.save(ROOT/'overview.jpg',quality=95)
|
||||
reader='''<!doctype html><html lang="zh-CN"><meta charset="utf-8"><meta name="viewport" content="width=device-width,initial-scale=1"><title>漫画</title><style>body{margin:0;background:#e9e6ed}main{max-width:1000px;margin:auto}img{width:100%;height:auto;display:block;margin-bottom:24px}@media print{img{margin:0;break-after:page}}</style><main>'''
|
||||
reader+=''.join(f'<img src="pages/page-{i:02d}.png" alt="">' for i in range(1,9))+'</main></html>'
|
||||
(ROOT/'index.html').write_text(reader)
|
||||
(ROOT/'tmp/final-layout.json').write_text(json.dumps(LAYOUT,ensure_ascii=False,indent=2))
|
||||
print(json.dumps({'rendered':[p['number'] for p in pages],'page_heights':sorted(set(x['page_height'] for x in LAYOUT)),'font_sizes':sorted(set(x['text_size'] for x in LAYOUT if x['text_size']))},ensure_ascii=False))
|
||||
@@ -0,0 +1,9 @@
|
||||
{
|
||||
"base": "Use case: illustration-story. Asset: a NEW page of an original Chinese conversational comic. All scenes are newly drawn, not repetitions of the reference pages.\nReferences: image1 is USER'S STYLE reference only; images2-4 show reusable original CHARACTER DESIGNS, not the required panel layout. Closely preserve these original people's faces/clothing, but use the NEW camera angles and panel composition below.\nSTYLE is the very simple flat cartoon of image1, consistent with references: huge rounded heads, thin long necks, small torsos; huge angular white eyes, tiny black dot pupils, expressive heavy eyebrows; pure black solid hair silhouettes, medium brown flat skin. Uniform confident black outlines, very simple hands and minimal clothing folds. NO anime details, hair strands, realistic anatomy, cinematic lighting, gradients, shading, rich interiors. Light periwinkle/lilac/cream/soft mint backgrounds, white page gutters, softly rounded panels without black frame borders. Just a few simple story props and minimal room lines, never a detailed realistic cafe.\nCHARACTER KEY: A = orange overshirt, short side-part black hair, ROUND black spectacles, male (host from prior pages). B = charcoal gray shirt, black hair with gray temples, flat heavy eyebrows, male (middle-aged man). C = coral-red hoodie, brown cloud-shaped curly hair, ROUND glasses, woman. D = pale green shirt, simple black bob, tiny gold TRIANGLE earrings, woman writer. E = purple shirt, black high BUN, RECTANGULAR glasses, woman engineer. F = blue plain hoodie, black 3-point quiff, no glasses, young man. G = NEW benchmark author, teal-green vest over cream shirt, solid black hair tied in LOW ponytail, no glasses, medium brown skin, adult.\nOnly depict specified characters in each panel; don't add unrelated people.\nLAYOUT RULE: portrait 1024x1536. The exact panel rectangles below define the whole page, all coordinates measured from top-left. Gaps remain WHITE. Do NOT convert the page into a regular four-panel grid. Keep panel count, order, proportions and dimensions as specified.\nLETTERING RULE: This is TEXT-FREE ART. For every marked lettering box, leave EMPTY FLAT BACKGROUND, including no heads, hair, props, hands, speech bubbles, decorative motifs, or outlines. Place drawing outside these boxes. These are invisible reservations, NOT white banners or drawn rectangles. Final Chinese dialogue will be added later. NO text, letters, digits, logos, signatures, page title, footers, character labels or watermark anywhere. Screen contents only simple geometric lines/blocks. Don't draw blank speech balloons.\n",
|
||||
"scenes": {
|
||||
"1": "PAGE 1, FIVE panels; location a simple shared discussion lounge with one low table.\nPanel1 rectangle (16,16,992,304), lettering empty box (40,36,680,174): A on far right, leaning forward and asking group for real experience, notebook in hand, simple arc gesture.\nPanel2 rectangle (16,336,608,480), empty box (40,360,560,238): B large chest-up in lower section, confident measured expression, palm tilted as if comparing models, simple phone near table.\nPanel3 rectangle (640,336,368,480), empty box (660,360,328,240): C close-up in lower section, raised eyebrows and disappointed mouth, listening to B. Tight reaction shot.\nPanel4 rectangle (16,832,992,296), empty box (390,854,594,228): D at LEFT of panoramic strip with book on lap, calmly raises one finger to make distinction. A tiny offscreen listener edge allowed at far left, but D clearly main.\nPanel5 rectangle (16,1144,992,376), empty box (40,1166,670,246): C close-up at FAR RIGHT, expression suddenly lights up, hands clasped, D small at bottom-left smiling. Show clear reversal from panel3 to panel5. Colors/shapes match references. No E,F,G.",
|
||||
"2": "PAGE 2, THREE PANORAMIC PANELS with unequal heights, simple project workbench.\nPanel1 rectangle (16,16,992,360), empty lettering box (40,40,944,204): E below box near center-right, sharply leans into conversation, skeptical raised eyebrow, one outstretched hand. B in far-left lower edge arms crossed. Energetic interruption.\nPanel2 rectangle (16,392,992,728), empty lettering box (40,418,944,242): E in lower-left three-quarter view at a workbench. Arrange a few large interlinked monitor cards/modules in lower-right, showing simple colored rectangles and connectors, representing many plugin workflows and system prompts; five simple plug pieces and a few different cards. Not a dense technical diagram, just clean playful cartoon machinery.\nPanel3 rectangle (16,1136,992,384), empty lettering box (40,1158,690,244): E face and hand on FAR RIGHT, holds a magnifier over simple red flaw marks on sheets, confident critique; small B face peeks bottom-left with narrowed eyes. A blue review card points at a big broken central connection, showing core problems. Avoid any objects inside text zone.",
|
||||
"3": "PAGE 3, FOUR UNEQUAL panels, explicitly NOT a 2x2 grid, a visual comparison of familiar building blocks versus unknown puzzle.\nPanel1 rectangle (16,16,992,360), empty box (40,40,944,190): E at bottom right gesturing toward a simple laptop showing a blocky toy browser game and UI cards; C listening at bottom left. Flat minimal background.\nPanel2 rectangle (16,392,488,650), empty box (40,416,440,332): in lower portion E casually assembles identical simple modular blocks on a conveyor/table, relaxed unimpressed expression: familiar patterns already available.\nPanel3 rectangle (520,392,488,650), empty box (544,416,440,312): E in lower portion stares at weird irregular unmatched puzzle pieces, eyebrows intense, no ready outline to place them into. Same scale as panel2, visual contrast.\nPanel4 rectangle (16,1058,992,462), empty box (40,1082,944,226): E at bottom center points in frustration to an obviously HALF-FINISHED small bridge/model on table. B at bottom-left looks defensive with folded arms. A tiny simple green checkmark-shaped stamp sitting beside unfinished model represents premature confidence. Checkmark shape okay, no letters or numbers.",
|
||||
"4": "PAGE 4, THREE WIDE PANELS, built around a pause and then a SILENT reaction, a simple testing counter with phones.\nPanel1 rectangle (16,16,992,540), empty box (40,40,606,248): B in gray on RIGHT holds phone with tiny red gauge segment, puzzled disbelief, raised eyebrows, open palm. Single tabletop. A small at far left lower edge listening.\nPanel2 rectangle (16,572,992,480), empty box (40,596,690,210): F blue hoodie on RIGHT leans in with phone showing a completely empty rectangular battery/gauge, incredulous look. B at lower-left reaches for his own phone. Screen content ONLY geometric gauge.\nPanel3 rectangle (16,1068,992,452), NO lettering required: large close-up reaction shot spanning whole panel. B LEFT and F RIGHT stare silently at two near-empty phone gauges between them, heavy half-closed eyes, small open mouths, shoulders sagging. Make this silent pause funny and visually spacious. No text or speech bubbles. Do not add a caption or invented dialogue."
|
||||
}
|
||||
}
|
||||
@@ -0,0 +1,9 @@
|
||||
{
|
||||
"base": "Use case: illustration-story. Asset: a NEW page of an original Chinese conversational comic. All scenes are newly drawn, not repetitions of the reference pages.\nReferences: image1 is USER'S STYLE reference only; images2-4 show reusable original CHARACTER DESIGNS, not the required panel layout. Closely preserve these original people's faces/clothing, but use the NEW camera angles and panel composition below.\nSTYLE is the very simple flat cartoon of image1, consistent with references: huge rounded heads, thin long necks, small torsos; huge angular white eyes, tiny black dot pupils, expressive heavy eyebrows; pure black solid hair silhouettes, medium brown flat skin. Uniform confident black outlines, very simple hands and minimal clothing folds. NO anime details, hair strands, realistic anatomy, cinematic lighting, gradients, shading, rich interiors. Light periwinkle/lilac/cream/soft mint backgrounds, white page gutters, softly rounded panels without black frame borders. Just a few simple story props and minimal room lines, never a detailed realistic cafe.\nCHARACTER KEY: A = orange overshirt, short side-part black hair, ROUND black spectacles, male (host from prior pages). B = charcoal gray shirt, black hair with gray temples, flat heavy eyebrows, male (middle-aged man). C = coral-red hoodie, brown cloud-shaped curly hair, ROUND glasses, woman. D = pale green shirt, simple black bob, tiny gold TRIANGLE earrings, woman writer. E = purple shirt, black high BUN, RECTANGULAR glasses, woman engineer. F = blue plain hoodie, black 3-point quiff, no glasses, young man. G = NEW benchmark author, teal-green vest over cream shirt, solid black hair tied in LOW ponytail, no glasses, medium brown skin, adult.\nOnly depict specified characters in each panel; don't add unrelated people.\nLAYOUT RULE: portrait 1024x1536. The exact panel rectangles below define the whole page, all coordinates measured from top-left. Gaps remain WHITE. Do NOT convert the page into a regular four-panel grid. Keep panel count, order, proportions and dimensions as specified.\nLETTERING RULE: This is TEXT-FREE ART. For every marked lettering box, leave EMPTY FLAT BACKGROUND, including no heads, hair, props, hands, speech bubbles, decorative motifs, or outlines. Place drawing outside these boxes. These are invisible reservations, NOT white banners or drawn rectangles. Final Chinese dialogue will be added later. NO text, letters, digits, logos, signatures, page title, footers, character labels or watermark anywhere. Screen contents only simple geometric lines/blocks. Don't draw blank speech balloons.\n",
|
||||
"scenes": {
|
||||
"5": "PAGE 5, FIVE PANELS like an experimental game demonstration. Introduce G, same simple cartoon language, teal vest + cream shirt + black low ponytail, NO glasses. Quiz-table setting with minimal props.\nPanel1 rectangle (16,16,608,380), empty text box (40,40,560,194): G in lower portion raises a plain facedown card, confident approachable explanation; simple card table.\nPanel2 rectangle (640,16,368,380), empty box (660,40,328,194): lower portion displays three small secret-subject pictures on cards: cartoon Einstein head (white frizzy hair, mustache), simple bicycle hand pump, pale moon. They are target illustrations only, not extra recurring people. No words or question marks.\nPanel3 rectangle (16,412,992,444), empty box (40,436,944,154): G at bottom-left runs a simple tabletop setup of THREE colored generic computer tokens, connected in sequence: question token / answer token / checking token. Use only simple shapes and a magnifier/checkmark icon, NO actual letters, digits or labels.\nPanel4 rectangle (16,872,488,648), empty box (40,896,440,266): G below box looks pleasantly surprised holding a simple fan of successful game cards; two equal-height podium blocks side by side and third slightly lower, all blank unbranded. No model logos.\nPanel5 rectangle (520,872,488,648), empty box (544,896,440,266): G below box calmly opens a wide binder of game logs with just geometric lines, holds a magnifier to inspect reasoning; C small at lower-right listens thoughtfully. Focus on checkable process instead of trophy celebration.",
|
||||
"6": "PAGE 6, TWO LARGE WIDE PANELS, spacious comic splash composition about enterprise coding.\nPanel1 rectangle (16,16,992,520), empty box (40,40,650,236): E purple shirt rectangular glasses high bun on FAR RIGHT, pleased confident face, holding a plain laptop in a minimal project room; B gray shirt small at bottom-left skeptical but listening. Few wall rectangles only.\nPanel2 rectangle (16,552,992,968), empty box (40,578,944,260): a dramatic large lower tableau below y850. THREE large colorful jigsaw-like software boards spread across a broad worktable/floor, interlocking into one coherent structure; each board contains only a few simple colored code-like line blocks, absolutely no letters. E in foreground lower-right reaches across to connect the third board; B at lower-left watches, recognizing the improvement. One mismatched loose piece at far edge may show old inconsistency. Huge heads, thin necks, playful simple physical metaphor, flat-color simple perspective. Do not render realistic computers or technical blueprints. This page has only TWO panels, do not subdivide it.",
|
||||
"7": "PAGE 7, THREE PANELS with one wide top and asymmetrical long lower portraits, a warm daily conversation in a minimal lounge/park-like nook.\nPanel1 rectangle (16,16,992,520), empty box (40,40,650,262): D green shirt black bob triangle earrings on FAR RIGHT chats to a simple phone, relaxed delighted expression. One chair arm and one simple leaf silhouette, no detailed background. Small soft empty speech-shaped decorative motifs allowed only outside lettering reservations, no text.\nPanel2 rectangle (16,552,360,968), empty box (38,578,316,334): tall narrow panel. D below y940, focused practical face, holds laptop with simple terminal-like blocks and two small blank model cards, representing she still uses other models for coding. Entire body stylized, no letters on laptop. No new person.\nPanel3 rectangle (392,552,616,968), empty box (416,578,568,224): large relaxed reaction. D smiling warmly and gesturing with open hands toward C coral hoodie round glasses, C laughs with half-closed eyes; two cozy chairs, simple tabletop, nothing realistic. A tiny lifeless square robot figurine sits unattended on table at far corner as a metaphor for soulless chat, deliberately simple prop, not a speaking character. Keep D design consistent across all three panels.",
|
||||
"8": "PAGE 8, FOUR UNEQUAL PANELS, not a regular grid, return to Pro usage worry. F blue hoodie black 3-point quiff, no glasses, male.\nPanel1 rectangle (16,16,992,488), empty text box (40,40,944,260): F below y310 looks at phone with dismay, tiny seven plain query cards and mostly depleted red meter on desk (no numbers or letters), C coral hoodie small on right watching.\nPanel2 rectangle (16,520,488,568), empty box (40,544,440,268): F below y830 dwarfed by a stack of thick legal/document books and folder sheets, anxious before even starting work. No book titles or words, simple flat shapes.\nPanel3 rectangle (520,520,488,568), empty box (544,544,440,280): F below y840 compares two plain phone/clock panels, old comfortable long-session memory versus new red nearly empty meter. Use simple clock face hands without numbers; no caption. Expression of puzzled frustration.\nPanel4 rectangle (16,1104,992,416), empty box (40,1128,700,230): F at FAR RIGHT stares wearily at a tiny phone that still has a green gauge, shoulders lowered; small pile of unopened work folders at bottom-left. Quiet final beat, no invented triumphant resolution, no spoken words generated. All text is added afterward."
|
||||
}
|
||||
}
|
||||
@@ -0,0 +1,135 @@
|
||||
# 第二组 · 最终中文对白
|
||||
|
||||
以用户提供的英文原文为准重译。色名用于编辑核对,漫画中不显示角色名。
|
||||
|
||||
## 第1页:一句比较,两种反应
|
||||
|
||||
### 第1格
|
||||
|
||||
橙色:试过 Gemini 4 Argon 的人,感觉怎么样?先不谈那些不太可信的跑分,也别只看粗糙的浏览器小游戏。
|
||||
|
||||
### 第2格
|
||||
|
||||
深灰色:是个很棒的模型,还没到 Opus 5.5 的水平,但追平 Astra 很轻松。有传言说 Google 已经准备好了一个更强的检查点。它会并进 Argon,还是变成 Gemini 4.1,就拭目以待吧。
|
||||
|
||||
### 第3格
|
||||
|
||||
红色:“还没到 Opus 5.5 的水平”……一个甚至还没正式推出的前沿模型,我真不想听到这样的评价。
|
||||
|
||||
### 第4格
|
||||
|
||||
绿色:仅供参考:我觉得 Opus 5.5 更擅长写代码,但 Argon 更擅长写作。
|
||||
|
||||
### 第5格
|
||||
|
||||
红色:谢天谢地!Opus 5.5 已经写得非常好了,而短篇创作正是我现在最常用的 AI 用途之一。这下我有点兴奋了。
|
||||
|
||||
## 第2页:复杂工作中的反对意见
|
||||
|
||||
### 第1格
|
||||
|
||||
紫色:为什么大家都说 Opus 5.5 比 Astra 好,甚至比 Sol 6.1 好?做复杂工作,我觉得它连 Sol 6.1 都不如。它留下太多问题,也没能有效照顾到全局。
|
||||
|
||||
### 第2格
|
||||
|
||||
紫色:我有一套评估流程:大约五种插件与技能工作流,三到四种系统提示的组合,再加上全局和仓库两层的 CLAUDE.md。怎么调,都没法让它把事情全做对。连上下文上限调到五十万、七十五万,我也试过了。
|
||||
|
||||
### 第3格
|
||||
|
||||
紫色:再让 Sol 6.1 审查,它马上就能找出核心问题,而不只是边缘情况。我两个平台的最高档套餐都有。而且,OpenAI 给我的可用额度,至少是 Claude 的五倍……
|
||||
|
||||
## 第3页:演示与陌生问题
|
||||
|
||||
### 第1格
|
||||
|
||||
紫色:很多人检验模型,就是看它能不能做出一个游戏,或者一个前端界面。可这类问题相当机械化。
|
||||
|
||||
### 第2格
|
||||
|
||||
紫色:我已经不写界面了。一旦设计定了,你又有几年经验,也确定了框架的用法,难的部分就差不多做完了,剩下只是繁琐地敲代码。游戏开发里,很多难题也是已知的数学问题。AI 很擅长这些,因为它基本上是在复现已知模式。
|
||||
|
||||
### 第3格
|
||||
|
||||
紫色:我现在处理的是复杂数据问题,没有现成的框架可套。也就是说……没有现成代码可抄。这类问题上,我的体验和你一样。
|
||||
|
||||
### 第4格
|
||||
|
||||
紫色:Anthropic 的模型只把活干了一半,甚至没好好看现有代码和规格。那种感觉,就像一个自信过头的开发者。
|
||||
|
||||
## 第4页:额度让争论暂停
|
||||
|
||||
### 第1格
|
||||
|
||||
深灰色:我用 Pro 套餐,把思考强度调到低,问了 Argon 一个问题,就花掉了 5% 的额度。问题还一点都不离谱。
|
||||
|
||||
### 第2格
|
||||
|
||||
蓝色:我还看到有人说:大概问了三个普普通通的问题,Pro 额度就直接耗尽,100% 全没了。
|
||||
|
||||
### 第3格
|
||||
|
||||
(无对白的反应镜头)
|
||||
|
||||
## 第5页:二十问的证据
|
||||
|
||||
### 第1格
|
||||
|
||||
青绿色:想看浏览器游戏演示之外的东西?我做了一个基准测试,让模型玩“二十问”。
|
||||
|
||||
### 第2格
|
||||
|
||||
青绿色:藏一个目标,比如爱因斯坦、打气筒、月亮……模型要靠提问把它猜出来。
|
||||
|
||||
### 第3格
|
||||
|
||||
青绿色:另外的模型分别负责回答和复核每个答案。最后计分,看它猜中一共用了多少个问题。
|
||||
|
||||
### 第4格
|
||||
|
||||
青绿色:结果好得出乎我意料:Argon 解出了全部三十局,与 GPT-6 Astra 并列第一,Opus 5.5 稍稍落后。
|
||||
|
||||
### 第5格
|
||||
|
||||
青绿色:当然,这不能说明它写 Rust 或按规格做项目怎么样。但每一局都能打开看,你能直接检查它是怎么推理的。
|
||||
|
||||
## 第6页:企业套餐与代码库拼图
|
||||
|
||||
### 第1格
|
||||
|
||||
紫色:我用企业套餐。到目前为止,Argon 的结果让我很满意。之前的 3.1 Pro,跟 Sonnet 5.5 相比,实在差得太远了。
|
||||
|
||||
### 第2格
|
||||
|
||||
紫色:刚换到 Argon,我就发现它很擅长处理大型代码库。有时还要牵涉第三个代码库,彼此像拼图一样。Gemini 3.1 Pro 之前总是没法在这几个库之间保持一致;Argon 却好像完全明白我要什么。
|
||||
|
||||
## 第7页:写作者想要的日常聊天
|
||||
|
||||
### 第1格
|
||||
|
||||
绿色:对我来说,它绝对是日常聊天的最强选手。用不同语言聊都很丰富、很有趣,终于不觉得自己是在和机器人说话了。
|
||||
|
||||
### 第2格
|
||||
|
||||
绿色:参数规模肯定很大。写代码或者脚本,我还在用 GPT-6 和 Opus 5.5。
|
||||
|
||||
### 第3格
|
||||
|
||||
绿色:但它们聊起天来,就实在没什么灵魂。
|
||||
|
||||
## 第8页:Pro用户还没开始工作
|
||||
|
||||
### 第1格
|
||||
|
||||
蓝色:我是 Pro 用户。在 Gemini 应用里问了七个普通问题,就用掉了五小时额度的 52%。还不是什么复杂任务:Argon 什么时候给 Antigravity 的 Pro 用户用?你是什么模型?3.8 Flash 和 Argon 的幻觉率有什么信息?就这样,52% 没了。
|
||||
|
||||
### 第2格
|
||||
|
||||
蓝色:我平常拿 Gemini 做法律研究、文书起草和文件分析。这些复杂任务都还没试呢。我担心,连一个基础法律任务都刚开个头,五小时额度就已经见底了。
|
||||
|
||||
### 第3格
|
||||
|
||||
蓝色:连 3.1 Pro 现在每问一次,也要花掉将近 7%。以前我开着延长思考,连续用好几个小时,都不会耗完这五小时的额度。
|
||||
|
||||
### 第4格
|
||||
|
||||
蓝色:选 Pro 或 Argon 时,额度的计算方式明显不一样了。可 3.8 Flash,我基本上能一直用下去。
|
||||
@@ -0,0 +1,69 @@
|
||||
B 的立场站 Claude
|
||||
C 比较中立
|
||||
D 是写文的
|
||||
E 强烈支持 Gemini
|
||||
|
||||
A:试过 Gemini 4 Argon 的人,你有什么反馈吗?除了那些不值得信任的基准测试和粗糙的浏览器游戏
|
||||
|
||||
B:这是个很棒的模型,虽然还不到作品 5.5 的水平,但很容易就能和 Astra 媲美。有传言称谷歌已经准备好了另一个更棒的检查点,如果它能被纳入 Argon 或成为 Gemini 4.1,我们将拭目以待
|
||||
|
||||
It's a great model, not quite opus 5.5 level, but it easily matches Astra. There are rumors that Google already has another checkpoint ready that is even better, if that makes its way into Argon or becomes Gemini 4.1 we'll see
|
||||
|
||||
C:“还不到 Opus 5.5 级别”其实并不是我想要的这个词,毕竟这个前沿模型还没上市......
|
||||
|
||||
"Not quite Opus 5.5 level" is really not what I wanted to here for a frontier model that isn't even out yet...
|
||||
|
||||
D:我觉得 Opus 5.5 是更好的编程者,但 Argon 是更好的写手,仅供参考
|
||||
|
||||
I think Opus 5.5 is a better coder but Argon is a better writer FWIW
|
||||
|
||||
C 赞同 D:谢天谢地,因为 Opus 5.5 写得非常棒,短文创意是我目前用得最多的 AI 应用之一。现在我很兴奋。
|
||||
|
||||
Okay thank god because Opus 5.5 writes really amazing and short form creative is one of my most used AI applications right now. Now I'm excited.
|
||||
|
||||
E:为什么有人说 Opus 5.5 比 Astra 更好?或者第 6.1 个火星日?我发现复杂工作甚至比第 6.1 个火星日还晚。它留下了太多问题,或者没有有效考虑更宏观的全局。我有一个评估流程,包含大约 5 种不同的插件和技能工作流程,以及 3-4 种系统提示的排列组合,以及 [CLAUDE.md](http://claude.md/) 全局/仓库范围。没法让它做所有事情。即使上下文调整到 50 万上限、75 万等等。我让 Sol 6.1 复习,识别出直接的核心问题,而不仅仅是边缘情况。(我同时有 Claude 和 OpenAI 的最大配置)——还有,OpenAI = 作为 Claude 至少使用 5 次......
|
||||
|
||||
Why do people say Opus 5.5 is better than Astra? Or Sol 6.1? I find that its behind even Sol 6.1 for complex work. It leaves so many issues or doesnt consider the larger picture effectively. I have an eval pipeline with like 5 different plugin and skill workflows and 3-4 permutations of system prompts and [CLAUDE.md](http://claude.md/) global/repo scopes. Can't get it to do everything. Even with context adjustments down to 500k cap, 750k, etc. I have Sol 6.1 review and identify immediate core concern bugs not just edge cases. (I have both Claude and OpenAI max plans) - and also, OpenAI = at LEAST 5x usage as Claude ...
|
||||
|
||||
E 的补充:
|
||||
|
||||
Many people validate models on their ability to create a game or a frontend UI.
|
||||
很多人验证模型是否能创建游戏或前端界面。
|
||||
Problems which are quite mechanical, I've stopped writing UIs because once the design is there and you have some years in and have defined how you use your framework it felt like all the hard work was done and I just had to put in tedious text. If you've developed games you also know that most of the hard problems are known mathematical problems, AI excels at this since it's essentially reproducing a known pattern.
|
||||
这些问题相当机械化,我已经停止写界面了,因为一旦设计完成,经过几年工作,定义了框架的使用方式,感觉所有艰苦的工作都完成了,我只需要写枯燥的文字。如果你开发过游戏,你也知道大多数难题其实都是已知的数学问题,AI 在这方面表现出色,因为它本质上是在重现一个已知的模式。
|
||||
|
||||
I'm currently dealing with complex data problems which don't have known frameworks to solve them. That means.. no code to steal. For these problems I have the same experience as you. Any Anthropic model does half a job and doesn't even look at the existing code or specs decently. It behaves like an overconfident developer.
|
||||
我目前正在处理复杂的数据问题,这些问题没有已知的解决方案框架。那就是说......没有代码可盗。对于这些问题,我和你有同样的经历。任何 Anthropic 模型都只做了一半的工作,甚至都不怎么看现有代码或规格。它的表现就像一个自信过头的开发者。
|
||||
|
||||
B:Tengo el plan pro, le hice una pregunta a Argon, con nivel de pensamiento bajo y se gasto un 5% del plan de uso 💀.
|
||||
我用的是专业套餐,我用很低的思考方式问了 Argon 一个问题,他花了套餐的💀 5%。
|
||||
|
||||
Y no fue un pregunta muy loca.
|
||||
这并不是个疯狂的问题。
|
||||
|
||||
F:Yo hice unas 3 preguntas nada de otro mundo plan pro y adiós capacidad 100% consumió
|
||||
我问了大约三个问题,没什么特别的,Pro Plan 和再见,容量 100%耗尽
|
||||
|
||||
G:For something that isn’t a browser game demo: I run a benchmark where models play Twenty Questions. The model has to find a hidden subject (Einstein, a bike pump, the moon...) by asking questions, and separate models answer and double-check each answer. The score is how many questions it needed.
|
||||
如果不是浏览器游戏演示:我跑了一个基准测试,模型们玩 Twenty Questions。模型需要通过提问找到隐藏的主题(爱因斯坦、自行车打气筒、月亮......),不同的模型会回答并反复核对每个答案。分数是需要问多少题。
|
||||
|
||||
To my own surprise, it was very good. Argon solved all 30 games and tied GPT-6 Astra for first, with Opus 5.5 a bit behind. It won’t tell you anything about Rust or working from specs, but you can read every game and see how it reasons: [https://deep20bench.com/runs/BX-20261014-B-0003-official-M0034-002/](https://deep20bench.com/runs/BX-20261014-B-0003-official-M0034-002/)
|
||||
令我自己惊讶的是,它非常好。Argon 解决了全部 30 个游戏,并与 GPT-6 Astra 并列第一,Opus 5.5 稍稍落后。它不会告诉你关于 Rust 或基于规格的操作,但你可以阅读每款游戏,看看它是如何推理的:[https://deep20bench.com/runs/BX-20261014-B-0003-official-M0034-002/](https://deep20bench.com/runs/BX-20261014-B-0003-official-M0034-002/)
|
||||
|
||||
E:I’m on the Enterprise Plan, so far I love the results of Argon compared to Pro 3.1 which was garbage compared to Sonnet 5.5
|
||||
我用的是企业套餐,到目前为止我很喜欢 Argon 的效果,相比 Pro 3.1,Pro 3.1 比 Sonnet 5.5 差很多
|
||||
|
||||
When I switched to Argon, I immediately noticed how well it can handle large codebases which may even require “third” codebases that are like a puzzle. Gemini 3.1 Pro had issues here with consistency across my codebases but Argon seemed to understand perfectly what I want.
|
||||
当我切换到 Argon 时,立刻注意到它能很好地处理大型代码库,甚至可能需要像拼图一样的“第三”代码库。Gemini 3.1 Pro 在我的代码库中存在不一致的问题,但 Argon 似乎完全理解我的需求。
|
||||
|
||||
D:For me it is the absolute GOAT in every day chat, HUGE parameters for sure becuse it is rich and fun to chat with in different languages, finally not feels like im interacting with a robot. For coding or scripts i am still using gpt6 and opus 5.5 still but their chat is just soulless.
|
||||
对我来说,它是日常聊天中的绝对最强,参数非常大,因为它丰富且有趣,用不同语言聊天,终于不再感觉像是在和机器人互动。编程或脚本方面我还在用 GPT6 和 Opus 5.5,但他们的聊天功能完全没有灵魂。
|
||||
|
||||
F:I'm a Pro tier subscriber and I've just asked it a couple regular text prompts in the Gemini app and it's burned through 52% of my 5-hour quota in seven different queries. Not even complex questions, mainly things where I'm asking basic questions like when will Argon come to Pro tier subscribers in the anti-gravity app, what model are you, and what's the information on the hallucination rate between 3.8 Flash and Argon. It burned through 52% of my quota just like that.
|
||||
我是专业版订阅者,刚刚在 Gemini 应用里问了几个普通文本提示,结果在七个不同的查询中,它已经用完了我 5 小时配额的 52%。甚至不是复杂的问题,主要是我问一些基础问题,比如 Argon 什么时候会在反重力应用的 Pro 级订阅用户中推出,你用的是什么型号,以及 3.8 Flash 和 Argon 之间的幻觉率信息。它就这样消耗了我 52%的配额。
|
||||
|
||||
I haven't even attempted more complex prompts where I normally use Gemini for legal research and legal drafting and document analysis, and I'm worried that I'll barely be able to start a basic legal task on Argon before running through my 5-hour limit.
|
||||
我甚至还没尝试过那些我通常用 Gemini 做法律研究、法律起草和文件分析的更复杂提示,我担心自己在 Argon 上几乎无法开始一个基本的法律任务,就要用完 5 小时的限制。
|
||||
|
||||
I also noticed that 3.1 Pro is burning through almost 7% per prompt now too, and I used to be able to use 3.1 Pro on extended thinking for long multi-hour sessions without burning through my 5-hour limit. Something is seriously different in how they're calculating quota usage when selecting the Pro model or the Argon model. 3.8 Flash I can use non-stop basically.
|
||||
我还注意到 3.1 Pro 现在每个提示几乎消耗了 7%的资源,而我以前能用 3.1 Pro 进行长时间的长时间思考而不用超过 5 小时的限制。他们在选择 Pro 型号或 Argon 型号时,计算配额使用的方式确实有很大不同。3.8 闪光灯我基本上可以不停使用。
|
||||