Adobe正式发布了他们新一代的AI绘图大模型:Adobe Firefly 3。

细节更强、语义理解更强、控制性更强等等。
还发了新一版本的PS AI。
不过这些不是重点。
Adobe Firefly 3的发布,结合前段时间发布的SD3,让我有了再一次搞一个AI绘图大模型竞技场,评测一下的想法。
在现在这个节点,过了近半年的时候,来再看一下现在进化过的巨头们,已经达到了什么样的水平。
四家分别为:
Midjourney V6、Adobe Firefly 3、Stable Diffusion 3、Dalle 3。

至于评测方式,我依然会从细节质量、审美(构图色彩等)、语义理解这三个维度来评测,剔除掉了风格多样化这个指标(没法测)。
细节质量、审美、语义理解每个类别14个case,总和42个Case(42这个数字的代表意义懂的都懂哈哈哈哈)
同时每个Prompt我会在AI绘图模型中roll3次出12张图,取效果最具有代表性的那个图,尽量减少偏见。同时为了保证公平,基本不会搞特别复杂的prompt。
同时,为了有最后整体可视化的评分让大家看着更直观,所以我会进行打分。在每个案例中,第一名为4分,第二为3分,第三为2分,最后一名为1分,最后计算平均分。
虽然每个case数量都不是很多,但是这也差不多了,而且是我个人的极限了。为了避免文章太长阅读体验极差,我就每个类别只放8个Case来做展示。
OK,让我们开始吧。
一. 细节质量
主要测试AI绘图对于细节的表现能力,比如人物面部皮肤的质感、比如织物纹理的细节、场景细微元素的细节等等,这个是对模型精度和输出质量一个非常重要的考量。
1.Prompt:
Selfie of charming kpop girl, outdoors, evening time, brunette, casual giggle, 2 bun tied hairstyle

Midjourney > SD3 > Adobe > Dalle
-
2.Prompt:
Portrait of a 2000s blonde woman posing on a sports car, white wired headphones, expressionless, 2000s hairstyle, 2000s fashion, sun rays, light teal and amber,Cinestill 50D

Midjourney > SD3 > Adobe > Dalle
-
3.Prompt:
Photo of smiling Labrador wearing sunglasses and straw hat sitting on the beach bench with glass of cocktail, beach scene, realistic

Midjourney > SD3 > Adobe > Dalle
-
4.Prompt:
a sports car drifting in a middle of partitions in a festival of vape and there is people around the car vaping, cinematic mood

SD3 > Adobe > Midjourney > Dalle
-
5.Prompt:
Realistic illustrations,The drumstick hits the frame and the drum bounces up water droplets

Midjourney > Adobe > Dalle > SD3
-
6.Prompt:
a house design inside of the perfect beach house, rustic malibu in style, the beach and surf included in the photos, Photography

Midjourney > Adobe > SD3 > Dalle
-
7.Prompt:
beautiful blonde model made out of porcelain, long hair, wearing sci-fi light mecha armor, in the style of balanced symmetry, white and blue LED lights on armor

Midjourney > SD3 > Adobe > Dalle
-
8.Prompt:
Delicious hamburger, floating in the air, food professional photography, studio lighting, studio background

Midjourney > Adobe > SD3 > Dalle
-
剩下case略。
在细节质量部分,Midjourney基本以绝对的优势压倒性胜利。


Midjourney > SD3 > Dalle > Adobe
-

Midjourney > Dalle > SD3 > Adobe
-

Midjourney > Adobe > SD3 > Dalle
-
4.Prompt:
woman smiling and having a cup of 7-eleven coffee outside a 7-eleven convenience store in the morning in the style of 90's anime, 1990s anime texture and colors, thick line work

Midjourney > Dalle > SD3 > Adobe
-
5.Prompt:
fantasy greatsword made from crimson metal, oil painting

Midjourney > SD3 > Dalle > Adobe
-
6.Prompt:
a dark ocean with great Sturm, Captive Souls Pirate's Redemption, ship emerging out of the fog, Giant octopus reaching out of the waters to pull down the ship

Midjourney > Dalle > SD3 > Adobe
-
7.Prompt:
warhammer 40K, Islamic space marine, white armor, black and gold trim, matte paintin

Midjourney > SD3 > Adobe > Dalle
-
8.Prompt:
oil painting of an angel with wings spread above the forest, light beam from its eyes illuminates path in bright green and blue colors

Midjourney > Adobe > SD3 > Dalle
-
剩下case略。
在审美部分,Midjourney依然以绝对的优势压倒性胜利,而以设计起家的Adobe,反而拉了最大的跨。

三. 语义理解
主要测试AI绘图对于复杂语义的理解能力,能否将文本内容都能清晰的表达出来并保证生成图片的质量。
Portrait photograph of an anthropomorphic tortoise seated on a New York City subway train

Dalle > Midjourney > SD3 > Adobe
-
A businessman on a throne. The AI agents gathered behind him like royal guards. Photo Real

Dalle > Midjourney > SD3 > Adobe
-
A cup of coffee sitting on a table in front of a window, outside the window is a futuristic city; a futuristic monorail can be seen close by, many lush plants around, shot from ground floor, clouds above

Dalle > Adobe > SD3 > Midjourney
-
A hyper-realistic image of an anthropomorphic corn cob working as a cashier at a convenience store, depicted with a cheerful expression while laughing. The corn cob, dressed in the store's uniform, features a friendly face with eyes and a mouth on the husk, showing a big, joyful smile. The scene captures the corn cob scanning items at the cash register, wearing a typical convenience store uniform that includes a neat polo shirt and a name tag

Dalle > Midjourney > SD3 > Adobe
-
Editorial photography of astronaut cooking Christmas colorful chocolate honey cookies on spaceship, Christmas honey cookies floating around astronaut, no gravity, in spaceship, levitated

Dalle > Midjourney > SD3 > Adobe
-
a close up hyper realistic image of a medieval knight facing off against the grim reaper. Dramatic lighting

Dalle = Midjourney > Adobe > SD3
-
a very pretty young woman smilling flying over an aztec city with a dog, both the woman and the dog are flying, she is wearing an aztec outfit, the dog is wearing a colourful collar. they both seem to be having fun, ultra realistic

Dalle = Midjourney > Adobe > SD3
-
8.Prompt:
dungeons and dragons, high detailed, fantastic realism, female centaur with unicorn horn on head, hyper realistic

Midjourney > SD3 > Dalle> Adobe
-
剩下case略。
Dalle3和Midjourney基本上处于领先地位,Dalle还是领先一筹。Adobe继续垫底。

最后总结
在四个大模型三个维度评完了以后,我相信大家应该能对这几个大模型有大概的了解了。
但是为了更直观一些,我再来做个雷达图吧。

细节质量方面,MJ V6 > SD3 > Adobe Fiefly 3 > Dalle 3。
审美方面,MJ V6 > SD3 > Dalle 3 > Adobe Fiefly 3。
语义理解方面,Dalle 3 > MJ V6> SD3 > Adobe Fiefly 3。
MJ依然稳坐头把交椅,很多人跟我说,啥XX大模型在什么什么参数评测中已经超越了MJ啥啥的,我每次都点点头:哦。
而Adobe Fiefly 3的全面拉胯以至于我几度怀疑自己是不是选错了模型,直到我再三确认我选的确实就是Fiefly Image 3预览版。
就...拉胯的令人难以置信。
而SD3至少在我以API方式接入使用下,也没有很多自媒体或者其他人吹的那么神乎其神。
希望这个评测,能抛砖引玉吧,让大家对AI绘图综合有一些了解。
更建议的是,自己上手去试试。
又跑了十几个小时,虽然跟大家说的是只有42个Case,但是背后跑了不知道多少。希望能对大家有所帮助吧。
好文章,需要你的鼓励
微软近年来频繁出现技术故障和服务中断,从Windows更新删除用户文件到Azure云服务因配置错误而崩溃,质量控制问题愈发突出。2014年公司大幅裁减测试团队后,采用敏捷开发模式替代传统测试方法,但结果并不理想。虽然Windows生态系统庞大复杂,某些问题在所难免,但Azure作为微软核心云服务,反复因配置变更导致客户服务中断,已不仅仅是质量控制问题,更是对公司技术能力的质疑。
Meta研究团队发现仅仅改变AI示例间的分隔符号就能导致模型性能产生高达45%的巨大差异,甚至可以操纵AI排行榜排名。这个看似微不足道的格式选择问题普遍存在于所有主流AI模型中,包括最先进的GPT-4o,揭示了当前AI评测体系的根本性缺陷。研究提出通过明确说明分隔符类型等方法可以部分缓解这一问题。
当团队准备部署大语言模型时,面临开源与闭源的选择。专家讨论显示,美国在开源AI领域相对落后,而中国有更多开源模型。开源系统建立在信任基础上,需要开放数据、模型架构和参数。然而,即使是被称为"开源"的DeepSeek也并非完全开源。企业客户往往倾向于闭源系统,但开源权重模型仍能提供基础设施选择自由。AI主权成为国家安全考量,各国希望控制本地化AI发展命运。
香港中文大学研究团队开发出CALM训练框架和STORM模型,通过轻量化干预方式让40亿参数小模型在优化建模任务上达到6710亿参数大模型的性能。该方法保护模型原生推理能力,仅修改2.6%内容就实现显著提升,为AI优化建模应用大幅降低了技术门槛和成本。