为什么需要真 alpha

一张好图应该能放在任何背景上。带 alpha 通道的 PNG 可以同时用于活动页、网站、演示文稿与商品目录，而不必每次手动抠图，也不会留下白框。

透明素材的四个典型用途：

• 季节性电商活动：产品图只生成一次，反复铺在不同节日的门店背景上。
• 企业演示：透明图表直接叠在指定的 PowerPoint 主题与渐变背景上，不破坏模板。
• 设计模板：图标、贴纸、装饰元素交给用户，放任意模板都不需要遮罩。
• 按需印花：一份设计稿套到不同颜色的空白服装与商品图上。

准备工作

pip install openai pillow

设置环境变量 OPENAI_API_KEY。本文使用 gpt-image-2.5-flare，请求 PNG 输出并设置 background="transparent"；要比较更重质量的模型，把模型名换成 gpt-image-2.5-sunburst。两者都支持透明背景。第一次比较时保持提示词、尺寸、quality="high" 不变，先看透明度、边缘细节与文字，再调质量或延迟。

import base64
import os
from pathlib import Path

from openai import OpenAI
from PIL import Image

if not os.getenv("OPENAI_API_KEY"):
    raise RuntimeError("Set OPENAI_API_KEY before running this notebook.")

IMAGE_MODEL = "gpt-image-2.5-flare"

client = OpenAI()
asset_dir = Path("images/transparent-image-assets")
asset_dir.mkdir(parents=True, exist_ok=True)

一条容易踩的规则：提示词优先于 background 参数

如果提示词里描述了背景、场景、颜色或「放在某物上」，模型可能照着描述生成背景，而不是给你透明底。

所以提示词必须同时做到两件事：

1. 只描述孤立的主体；
2. 明确要求真正的透明 alpha，并排除 backdrop、矩形、底座、投影、标签文字与水印。

一个可复用的「品牌 + 透明」约束段落：

One object from STILLROOM, an understated luxury botanical home-fragrance atelier:
artisanal warm amber and honey-tinted hand-blown glass, subtle brushed brass hardware,
natural ivory cotton cord, quiet Japanese-Scandinavian restraint,
warm side-lit editorial studio photography, photorealistic micro-detail.
Full object completely visible and generously padded. Preserve every natural transparency,
refraction, translucent layer and fine material edge.
Output an isolated object on actual fully transparent alpha; no backdrop, no rectangle,
no plinth, no cast shadow, no readable writing, no label text, no watermark.

把这段和每个产品的具体描述拼起来，就能让同系列的四件产品在视觉上保持一致。

生成透明 PNG

products = {
    "amber-parfum": "A sculptural clear honey-amber faceted glass eau-de-parfum bottle, ...",
    "botanical-oil": "A slim cylindrical deep honey-amber transparent hand-blown glass bottle, ...",
    "reed-diffuser": "A rounded low smoked-amber translucent glass diffuser vessel, ...",
    "garden-candle": "A handmade botanical candle in a thick translucent smoked-amber tumbler, ...",
}

for filename, product_prompt in products.items():
    result = client.images.generate(
        model=IMAGE_MODEL,
        prompt=f"{product_prompt} {brand_prompt}",
        background="transparent",
        size="1024x1536",
        quality="high",
        output_format="png",
    )
    image_bytes = base64.b64decode(result.data[0].b64_json)
    (asset_dir / f"{filename}.png").write_bytes(image_bytes)

注意 output_format="png" 与 background="transparent" 必须同时出现；只设背景参数、却让提示词描述场景，结果会退回不透明。

验证 alpha：不要相信肉眼

肉眼看到的「白底」可能只是白色像素。用 Pillow 检查通道与透明像素比例：

for filename in products:
    with Image.open(asset_dir / f"{filename}.png") as image:
        if image.mode != "RGBA":
            raise ValueError(f"{filename}.png does not have an alpha channel.")

        transparent_pixels = image.getchannel("A").histogram()[0]
        total_pixels = image.width * image.height
        transparent_percentage = 100 * transparent_pixels / total_pixels

        print(
            f"{filename}.png: {image.width}x{image.height}, "
            f"{transparent_percentage:.2f}% fully transparent"
        )

三个检查点：

• mode 必须是 RGBA。若是 RGB，说明根本没有 alpha 通道。
• 透明像素占比应当明显大于 0。为 0 意味着整幅图不透明。
• 占比也不能接近 100%——那通常意味着主体没生成出来。

预期的输出形态类似：

amber-parfum.png: 1024x1536, 78.41% fully transparent
botanical-oil.png: 1024x1536, 81.02% fully transparent

用例一：季节性电商活动

目标：同一组产品图，换四种节日背景，不重新抠图。

流程是：先定义四个产品提示词 → 生成透明 PNG → 用上面那段代码验证 → 把 PNG 与页面需求交给 Codex 生成单页门店 → 在春冬两套背景间切换，检查同一批 PNG 在新背景上的边缘。

交给 Codex 的指令要写清「不要生成替代图，也不要跑额外的去背步骤」：

Build a responsive single-page campaign website for STILLROOM using these existing transparent PNGs:
- images/transparent-image-assets/amber-parfum.png
- ...
Add Spring, Summer, Autumn, and Winter controls that change the background and copy.
Let visitors choose one product, two products, or the full collection.
Add a "Show transparency" checkerboard toggle.
Do not generate replacement images or run a separate background-removal step.

要盯住的难点边缘：透明玻璃与磨砂玻璃、香水瓶的薄纱丝带、藤条与羽毛状的蒲苇纤维、蜡烛上的细棉绳。传统去背工具容易把这些细节剪掉、加光晕，或把半透明材质压平。直接生成 alpha 通道，正是为了保住这些边缘。

用例二：企业演示图表

企业模板常规定背景色或渐变，AI 生成的图表如果自带白底，放上去就是一块白斑。透明 PNG 能让主题背景透出来。

quarterly_revenue = {"Q1": 42, "Q2": 51, "Q3": 63, "Q4": 78}
regional_pipeline = {"North America": 46, "Europe": 29, "APAC": 25}

图表提示词要明确「整块图表都透明」，而不只是轮廓外透明。深色主题还要指定白色或浅色标签、亮色系、不要卡片、边框、面板或填充的绘图区：

Create one polished enterprise bar chart on a genuinely transparent background
for a dark navy PowerPoint slide.
Use exactly four ascending bars with these values: Q1 = $42M, Q2 = $51M, Q3 = $63M, Q4 = $78M.
Add white quarter labels below each bar and white dollar-value labels above.
Keep the plot area, grid, and space between bars transparent.
Do not add a background, filled panel, frame, title, or card.

用横版 1536x1024 给标签留空间，quality="high" 帮助小字清晰。

一个必须承认的边界：模型收到的是准确数值，但生成的是位图。使用前必须逐项核对标签、数值与比例是否与源数据一致。对数字精度要求高的正式报表，应当用确定性工具渲染图表，图像生成只用于插画类元素。

用例三：设计模板素材

图标、贴纸、装饰元素的关键是「外侧全透明」：

Create one premium rounded-square app icon with an abstract folded paper bird
in cobalt blue, lavender, and peach. Keep everything outside the rounded icon tile transparent.

Create one illustrated strawberry character sticker with a close-fitting cream die-cut border.
Keep everything outside the sticker border transparent.

Create one delicate editorial botanical sprig.
Keep all space around and between the thin branches and leaves transparent.

细枝、叶片之间的空隙最容易被填成白色，检验时尤其要看这些区域。

用例四：印花商品

把设计稿与空白服装分开生成，才能一套稿子套多种商品：

merchandise-design: 户外主题圆形构图，山、松树、落日、河流，四周与图形间留空透明。
merchandise-blank-tshirt: 正面平铺的象牙色 T 恤，胸前留空，轮廓外全透明。
merchandise-blank-sweatshirt: 同样的平铺森林绿卫衣。

合成时用 Pillow 去掉透明留白，再按比例贴到胸前：

def visible_region(image):
    # 去掉全透明外边距，不改变可见像素
    bounds = image.getchannel("A").getbbox()
    if bounds is None:
        raise ValueError("Expected a transparent PNG with visible artwork.")
    return image.crop(bounds)

with Image.open(asset_dir / "merchandise-design.png") as image:
    print_design = visible_region(image.convert("RGBA"))

with Image.open(asset_dir / "merchandise-blank-tshirt.png") as image:
    garment = visible_region(image.convert("RGBA"))

artwork = print_design.copy()
artwork.thumbnail((round(garment.width * 0.32), round(garment.height * 0.32)),
                  Image.Resampling.LANCZOS)

chest_x = (garment.width - artwork.width) // 2
chest_y = round(garment.height * 0.28)
garment.alpha_composite(artwork, dest=(chest_x, chest_y))

alpha_composite 会按 alpha 混合，而不是直接覆盖，因此设计的边缘能自然融入布料。若用 paste 且不给遮罩，透明区域会变成黑色或白色。

评估生成素材

评估维度：透明度、边缘质量、品牌契合度、在后四种场景中的可用性。

可以组织人工评审，也用具备视觉能力的模型按固定评分表复核，重点标记：不透明背景、被剪掉的细节、服装上位置不对的印花、与源数据不符的图表标签。

理解检验

1. background="transparent" 已经设置了，为什么还可能在结果里看到白色背景？

参考答案：因为提示词优先于参数。如果提示词描述了背景、场景、颜色或「放在某物上」，模型可能照着描述生成背景。解决办法是让提示词只描述孤立主体，并明确写出「输出在真正的透明 alpha 上，不要 backdrop、矩形、底座、投影」。

2. 只检查「看起来是白底」为什么不够？应该怎么验证？

参考答案：白色可能是不透明的白像素。必须检查文件本身：用 Pillow 打开后确认 mode == "RGBA"，再用 alpha 通道的直方图统计完全透明像素的比例。占比为 0 说明整幅不透明；接近 100% 说明主体没生成出来。两个极端都要报错或重生成。

3. 为什么透明素材要分别生成图表与商品，而不是在成品图上抠图？

参考答案：因为传统去背工具在处理半透明材质、细纤维、薄纱、发丝时容易剪掉细节、产生光晕或把透明压平。直接在生成阶段获得 alpha 通道，能保留这些边缘；而且素材是独立的，可以反复复用到不同背景、不同服装与不同版式上。

4. 把设计稿贴到 T 恤上时，为什么用 alpha_composite 而不是 paste？

参考答案：alpha_composite 会按 alpha 通道混合，设计图的半透明边缘能与布料自然过渡；paste 不带遮罩时直接把矩形区域覆盖过去，透明部分会被填成黑色或白色，边缘也会出现硬切。

5. 生成企业图表时，为什么提示词要求「绘图区、网格、柱间空隙都透明」，而不只是轮廓外透明？

参考答案：因为图表要叠在深色渐变背景上，如果绘图区或网格自带浅色底，就会形成一块可见的矩形面板，破坏模板。要求内部也透明，才能让背景从柱子之间、图例周围透出来。

6. 图表素材在正式使用前必须做哪一项与视觉无关的检查？

参考答案：核对数值。模型虽然收到了准确数字，但输出是位图，标签、数值、比例都可能出错。必须逐项对照源数据；对精度要求高的正式报表，应改用确定性工具渲染，图像生成只用于插画。

与前后内容的关系

这篇把单张图的生成推进到「可复用素材」：难点从写提示词转到验证与合成。它依赖同系列「图像提示与编辑」一篇里的提示词原则与参数约束；反过来，任何需要批量出图的流程（活动页、演示、商品目录）都应当先通过这里的 alpha 校验，再进入排版环节。

Why real alpha matters

A good image should work on any background. A PNG with a real alpha channel can serve campaign pages, websites, presentations, and product catalogs at once, with no manual cutout and no white box.

Four typical uses:

• Seasonal e-commerce campaigns: generate product images once, reuse them across holiday storefronts.
• Enterprise presentations: place transparent charts directly on a required PowerPoint theme and gradient background.
• Design templates: icons, stickers, and decorative elements that drop onto any layout without masking.
• Print-on-demand merchandise: apply one print design to different blank garments and product mockups.

Setup

pip install openai pillow

Set OPENAI_API_KEY. The examples use gpt-image-2.5-flare, request PNG output, and set background="transparent". To compare a model optimized for demanding quality, switch to gpt-image-2.5-sunburst. Both support transparent backgrounds. For the first comparison, keep prompts, sizes, and quality="high" unchanged; inspect transparency, fine edges, and text before tuning quality or latency.

import base64
import os
from pathlib import Path

from openai import OpenAI
from PIL import Image

if not os.getenv("OPENAI_API_KEY"):
    raise RuntimeError("Set OPENAI_API_KEY before running this notebook.")

IMAGE_MODEL = "gpt-image-2.5-flare"

client = OpenAI()
asset_dir = Path("images/transparent-image-assets")
asset_dir.mkdir(parents=True, exist_ok=True)

The rule that bites: prompts override the background parameter

If the prompt describes a backdrop, scene, color, or an object "on" something, the model may generate that background instead of returning transparency.

So the prompt must do two things at once:

1. Describe only an isolated subject.
2. Explicitly demand real transparent alpha, excluding backdrop, rectangle, plinth, cast shadow, label text, and watermark.

A reusable brand-plus-transparency constraint:

One object from STILLROOM, an understated luxury botanical home-fragrance atelier:
artisanal warm amber and honey-tinted hand-blown glass, subtle brushed brass hardware,
natural ivory cotton cord, quiet Japanese-Scandinavian restraint,
warm side-lit editorial studio photography, photorealistic micro-detail.
Full object completely visible and generously padded. Preserve every natural transparency,
refraction, translucent layer and fine material edge.
Output an isolated object on actual fully transparent alpha; no backdrop, no rectangle,
no plinth, no cast shadow, no readable writing, no label text, no watermark.

Combining this with each product's own description keeps a collection visually consistent.

Generating transparent PNGs

products = {
    "amber-parfum": "A sculptural clear honey-amber faceted glass eau-de-parfum bottle, ...",
    "botanical-oil": "A slim cylindrical deep honey-amber transparent hand-blown glass bottle, ...",
    "reed-diffuser": "A rounded low smoked-amber translucent glass diffuser vessel, ...",
    "garden-candle": "A handmade botanical candle in a thick translucent smoked-amber tumbler, ...",
}

for filename, product_prompt in products.items():
    result = client.images.generate(
        model=IMAGE_MODEL,
        prompt=f"{product_prompt} {brand_prompt}",
        background="transparent",
        size="1024x1536",
        quality="high",
        output_format="png",
    )
    image_bytes = base64.b64decode(result.data[0].b64_json)
    (asset_dir / f"{filename}.png").write_bytes(image_bytes)

output_format="png" and background="transparent" must appear together. Setting the background flag while describing a scene in the prompt falls back to an opaque image.

Verify alpha: do not trust your eyes

What looks like a white background may just be white pixels. Check the channel and the transparent-pixel share with Pillow:

for filename in products:
    with Image.open(asset_dir / f"{filename}.png") as image:
        if image.mode != "RGBA":
            raise ValueError(f"{filename}.png does not have an alpha channel.")

        transparent_pixels = image.getchannel("A").histogram()[0]
        total_pixels = image.width * image.height
        transparent_percentage = 100 * transparent_pixels / total_pixels

        print(
            f"{filename}.png: {image.width}x{image.height}, "
            f"{transparent_percentage:.2f}% fully transparent"
        )

Three checks:

• mode must be RGBA; RGB means there is no alpha channel at all.
• The transparent share must be clearly above 0. Zero means the whole image is opaque.
• It must not be near 100% either — that usually means the subject failed to generate.

Expected output looks like:

amber-parfum.png: 1024x1536, 78.41% fully transparent
botanical-oil.png: 1024x1536, 81.02% fully transparent

Use case 1: seasonal e-commerce campaigns

Goal: one set of product images, four seasonal backgrounds, no re-cutout.

The sequence is: define four product prompts → generate transparent PNGs → validate with the code above → hand the PNGs and page requirements to Codex for a single-page storefront → switch between Spring and Winter and inspect the same PNGs on new backgrounds.

Tell Codex explicitly not to replace the images or add a background-removal step:

Build a responsive single-page campaign website for STILLROOM using these existing transparent PNGs:
- images/transparent-image-assets/amber-parfum.png
- ...
Add Spring, Summer, Autumn, and Winter controls that change the background and copy.
Let visitors choose one product, two products, or the full collection.
Add a "Show transparency" checkerboard toggle.
Do not generate replacement images or run a separate background-removal step.

Edges to inspect: clear and frosted glass, the perfume bottle's sheer ribbon, the diffuser's thin reeds and feathery pampas fibers, the candle's fine cotton cord. Conventional background-removal tools clip these details, add halos, or flatten translucent materials. Generating the alpha channel directly is what preserves them.

Use case 2: enterprise presentation charts

Enterprise templates often mandate a background color or gradient. An AI-generated chart with a white background leaves a visible white patch. A transparent PNG lets the theme show through.

quarterly_revenue = {"Q1": 42, "Q2": 51, "Q3": 63, "Q4": 78}
regional_pipeline = {"North America": 46, "Europe": 29, "APAC": 25}

The prompt must ask for transparency across the entire chart, not just outside its silhouette. For a dark theme, also specify white or pale labels, bright colors, and no card, frame, panel, or filled plot area:

Create one polished enterprise bar chart on a genuinely transparent background
for a dark navy PowerPoint slide.
Use exactly four ascending bars with these values: Q1 = $42M, Q2 = $51M, Q3 = $63M, Q4 = $78M.
Add white quarter labels below each bar and white dollar-value labels above.
Keep the plot area, grid, and space between bars transparent.
Do not add a background, filled panel, frame, title, or card.

Use a landscape 1536x1024 to leave room for labels, and quality="high" to help small text.

A boundary to accept: the model receives exact values, but the output is raster. Verify labels, values, and proportions against the source data before use. For production reporting where numerical precision is critical, render charts with a deterministic tool and reserve image generation for illustrative elements.

Use case 3: design-template assets

The key for icons, stickers, and decorative elements is "transparent outside":

Create one premium rounded-square app icon with an abstract folded paper bird
in cobalt blue, lavender, and peach. Keep everything outside the rounded icon tile transparent.

Create one illustrated strawberry character sticker with a close-fitting cream die-cut border.
Keep everything outside the sticker border transparent.

Create one delicate editorial botanical sprig.
Keep all space around and between the thin branches and leaves transparent.

Gaps between thin branches and leaves are the most likely regions to be filled white; inspect them specifically.

Use case 4: print-on-demand merchandise

Generating the design and the blank garments separately is what lets one artwork serve several products:

merchandise-design: an outdoor circular composition with mountains, pines, a setting sun, a river;
keep the surrounding area and the spaces between elements transparent.
merchandise-blank-tshirt: a flat, front-facing ivory crew-neck tee with a blank chest;
everything outside the garment silhouette transparent.
merchandise-blank-sweatshirt: the same, in forest green.

When compositing, trim the transparent padding and place the artwork proportionally on the chest:

def visible_region(image):
    # Remove fully transparent padding without changing visible pixels.
    bounds = image.getchannel("A").getbbox()
    if bounds is None:
        raise ValueError("Expected a transparent PNG with visible artwork.")
    return image.crop(bounds)

with Image.open(asset_dir / "merchandise-design.png") as image:
    print_design = visible_region(image.convert("RGBA"))

with Image.open(asset_dir / "merchandise-blank-tshirt.png") as image:
    garment = visible_region(image.convert("RGBA"))

artwork = print_design.copy()
artwork.thumbnail((round(garment.width * 0.32), round(garment.height * 0.32)),
                  Image.Resampling.LANCZOS)

chest_x = (garment.width - artwork.width) // 2
chest_y = round(garment.height * 0.28)
garment.alpha_composite(artwork, dest=(chest_x, chest_y))

alpha_composite blends by alpha rather than overwriting, so the artwork's edges sit naturally on fabric. Using paste without a mask turns transparent regions black or white.

Evaluating generated assets

Evaluate transparency, edge quality, brand fit, and usefulness across representative campaigns, slides, templates, and mockups.

Human review can be paired with a vision-capable model against a clear rubric, flagging opaque backgrounds, clipped details, mislocated prints on garments, and chart labels that disagree with the source data.

Comprehension check

1. background="transparent" is set, so why can the result still come back with a white background?

Reference answer: Because the prompt takes priority over the parameter. If the prompt describes a backdrop, scene, color, or an object "on" something, the model may generate that background. The fix is to describe only an isolated subject and to state explicitly that the output must be on real transparent alpha with no backdrop, rectangle, plinth, or cast shadow.

2. Why is "it looks like a white background" insufficient, and how should you verify?

Reference answer: White may be opaque white pixels. Check the file itself: open it with Pillow and confirm mode == "RGBA", then use the alpha channel histogram to measure the share of fully transparent pixels. Zero means the image is entirely opaque; near 100% means the subject failed to generate. Both extremes should raise an error or trigger regeneration.

3. Why generate charts and products with transparency instead of cutting them out of a finished image?

Reference answer: Conventional cutout tools clip fine detail on translucent materials, thin fibers, sheer fabric, and hair, and often add halos or flatten transparency. Generating the alpha channel directly preserves those edges, and the assets stay independent so they can be reused on different backgrounds, garments, and layouts.

4. When placing a design on a T-shirt, why use alpha_composite rather than paste?

Reference answer: alpha_composite blends according to the alpha channel, so the artwork's semi-transparent edges transition naturally into the fabric. paste without a mask overwrites a rectangular region, filling transparent areas with black or white and leaving hard edges.

5. Why does the chart prompt ask for the plot area, grid, and gaps between bars to be transparent, not just the outside?

Reference answer: The chart is placed on a dark gradient. If the plot area or grid carries a light background, it forms a visible rectangular panel that breaks the template. Requiring internal transparency lets the background show through between bars and around the legend.

6. What non-visual check is mandatory before using a chart asset?

Reference answer: Verify the numbers. The model received exact values, but the output is raster, so labels, values, and proportions can be wrong. Each item must be checked against the source data; for precision-critical reporting, render charts with deterministic tools and use image generation only for illustrations.

How this connects

This note moves from producing a single image to producing reusable assets, where the hard part shifts from prompting to validation and compositing. It relies on the prompting principles and parameter constraints in the image prompting chapter; conversely, any batch pipeline — campaign pages, decks, catalogs — should pass the alpha checks here before layout.