本实验用 5 句真实中文口语需求,分别走三条路线:① 工作流(LLM 改写提示词 + 扩散模型生图)② 原生 Gemini 图像(本环境缺 google 模块,全部失败)③ 原生 GPT-Image 2(直出)。
每个需求 × 每条路线的成 / 败。原生 Gemini 路线因运行环境缺少 google 模块而整体失败;工作流与 GPT-Image 2 全部成功出图。
| 需求(中文口语) | ① 工作流 | ② 原生 Gemini | ③ 原生 GPT-Image 2 |
|---|---|---|---|
| 帮我画一个周末加班的程序员,风格丧一点 | 成功 | 失败 ModuleNotFoundError: No module named 'google' | 成功 |
| 帮我画一盆放在窗台上的绿植,早晨的阳光刚好照进来 | 成功 | 失败 ModuleNotFoundError: No module named 'google' | 成功 |
| 帮我做一张新款降噪耳机的产品海报,主打“深夜独处也清净”这句文案,风格简约高级 | 成功 | 失败 ModuleNotFoundError: No module named 'google' | 成功 |
| 帮我画一个 AGI 实现以后程序员的工作场景 | 成功 | 失败 ModuleNotFoundError: No module named 'google' | 成功 |
| 帮我画一幅“未来城市的早晨”的画 | 成功 | 失败 ModuleNotFoundError: No module named 'google' | 成功 |
工作流路线 = 改写节点耗时(蓝,含 kimi-k3 推理) + 生图节点耗时(浅蓝,含提交/轮询/下载);原生 GPT-Image 2 路线 = 单次调用耗时(紫)。
工作流路线的“改写节点”调用 kimi-k3(带深度推理),token 开销显著高于 GPT-Image 2 的图像输出 token。这是工作流额外耗时的主要来源,也换来更可控的提示词。
两条成功路线产出的 PNG 大小。GPT-Image 2 直出的图片普遍略大;工作流路线经 wan2.2-t2i-flash 生成。
同一句中文需求,左为工作流路线、右为原生 GPT-Image 2 路线。点图可放大。注意:海报需求中“深夜独处也清净”文案由工作流路线在提示词里预留负空间、后期排版实现(经典扩散模型文字生成能力弱)。
工作流路线的核心增值步骤——把中文口语需求改写成 Stable Diffusion / FLUX 风格的可执行提示词(含负向词与画风/构图/光线/情绪增补)。
prompt: exhausted programmer sitting alone at desk, empty office at night, weekend overtime, messy hair, dark circles under eyes, wrinkled gray hoodie, hunched over keyboard, multiple monitors displaying glowing code, cold blue screen light on tired face, desk cluttered with empty coffee cups, energy drink cans, instant noodle cup, rain streaks on window, distant city lights bokeh, dim moody lighting, muted desaturated color palette, melancholic gloomy atmosphere, sense of loneliness and burnout, anime illustration style, cinematic composition, masterpiece, best quality, highly detailednegative_prompt: lowres, worst quality, bad anatomy, bad hands, extra fingers, missing limbs, deformed face, blurry, watermark, text, signature, logo, bright cheerful colors, oversaturated, smiling, happy expression, daylight, crowded officeprompt: masterpiece, best quality, highly detailed, photorealistic, a potted green plant, lush vibrant green leaves, terracotta clay pot, sitting on a wooden windowsill, morning sunlight streaming through the window, golden sunbeams, soft volumetric lighting, gentle light rays, dappled light and shadow on leaves, dewdrops on leaves, warm cozy atmosphere, tranquil morning mood, shallow depth of field, soft bokeh background, natural lighting, close-up composition, 8k resolutionnegative_prompt: lowres, bad anatomy, blurry, out of focus, watermark, text, signature, logo, jpeg artifacts, oversaturated, overexposed, underexposed, harsh shadows, wilted leaves, dead plant, cluttered background, cartoon, painting, illustrationprompt: masterpiece, best quality, highly detailed, premium noise-cancelling wireless over-ear headphones, matte black finish with subtle metallic accents, floating centered composition, deep midnight navy gradient background, soft moonlight rim lighting, faint blurred city night bokeh lights in distance, minimalist luxury product photography, commercial advertising poster style, clean negative space at top for copy text, dramatic soft shadows, calm serene late-night mood, cinematic lighting, 8k, sharp focus, studio shotnegative_prompt: lowres, bad anatomy, blurry, watermark, text, logo, signature, cluttered background, oversaturated, jpeg artifacts, extra objects, people, hands, cheap plastic look, harsh shadows, noisy, grainy, croppedprompt: masterpiece, best quality, highly detailed, a relaxed programmer leaning back in an ergonomic chair, sipping coffee, feet up on the desk, casually supervising holographic screens, translucent floating lines of code and glowing data streams in the air, humanoid AI robot assistant typing at a workstation, multiple holographic monitors displaying flowing code, futuristic minimalist office interior, floor-to-ceiling window with a futuristic city skyline at dusk, neon blue and cyan ambient lighting, soft screen glow, clean sci-fi aesthetic, cinematic lighting, wide shot, depth of field, concept art, digital painting, calm ironic atmosphere of full automationnegative_prompt: lowres, worst quality, low quality, bad anatomy, bad hands, extra fingers, missing fingers, extra limbs, deformed face, blurry, watermark, text, logo, signature, jpeg artifacts, oversaturated, cluttered compositionprompt: masterpiece, best quality, highly detailed, futuristic city at morning, sci-fi cityscape, towering skyscrapers, glass and chrome architecture, flying vehicles, holographic billboards, elevated sky bridges, sunrise, golden hour, soft warm sunlight, morning mist, light rays through buildings, atmospheric perspective, lens flare, blue and orange color palette, cinematic lighting, wide angle shot, epic composition, concept art, digital painting, ultra-detailed, 8knegative_prompt: lowres, worst quality, low quality, blurry, bad anatomy, watermark, text, signature, logo, jpeg artifacts, noise, oversaturated, deformed, cropped, out of frame仅列出两条成功路线的关键字段(原生 Gemini 路线整体失败,不列入)。
| 需求 | 路线 | 状态 | 耗时(s) | 模型 | 体积 |
|---|---|---|---|---|---|
| programmer-overtime | workflow | 成功 | 57587.4 | kimi-k3 + wan2.2-t2i-flash | 0.98 MB |
| programmer-overtime | native_gptimage | 成功 | 24.28 | gpt-image-2 | 1.34 MB |
| windowsill-plant | workflow | 成功 | 29197.4 | kimi-k3 + wan2.2-t2i-flash | 1.27 MB |
| windowsill-plant | native_gptimage | 成功 | 18.64 | gpt-image-2 | 1.47 MB |
| headphone-poster | workflow | 成功 | 33459.6 | kimi-k3 + wan2.2-t2i-flash | 0.86 MB |
| headphone-poster | native_gptimage | 成功 | 25.9 | gpt-image-2 | 1.12 MB |
| agi-programmer | workflow | 成功 | 42579.0 | kimi-k3 + wan2.2-t2i-flash | 1.09 MB |
| agi-programmer | native_gptimage | 成功 | 44.7 | gpt-image-2 | 1.35 MB |
| future-city-morning | workflow | 成功 | 30446.9 | kimi-k3 + wan2.2-t2i-flash | 1.27 MB |
| future-city-morning | native_gptimage | 成功 | 36.75 | gpt-image-2 | 1.78 MB |