Files
chatgpt-exporter/src/exporter/image.ts
T

350 lines
14 KiB
TypeScript

import { snapdom } from '@zumer/snapdom'
import i18n from '../i18n'
import { checkIfConversationStarted, getChatIdFromUrl } from '../page'
import { downloadFile, getFileNameWithFormat } from '../utils/download'
import { Effect } from '../utils/effect'
import { encodePng } from '../utils/png'
import { sleep } from '../utils/utils'
const MAX_SCREENSHOT_DIMENSION = 16_000
const MAX_TILE_PIXELS = 16_000_000
function scrollElementWithinRoot(scrollRoot: HTMLElement, target: HTMLElement, block: 'start' | 'center') {
const scrollRect = scrollRoot.getBoundingClientRect()
const targetRect = target.getBoundingClientRect()
const offset = targetRect.top - scrollRect.top
const alignment = block === 'center' ? (scrollRoot.clientHeight - targetRect.height) / 2 : 0
const requestedScrollTop = Math.max(0, Math.min(
scrollRoot.scrollHeight - scrollRoot.clientHeight,
scrollRoot.scrollTop + offset - alignment,
))
scrollRoot.scrollTop = requestedScrollTop
scrollRoot.dispatchEvent(new Event('scroll', { bubbles: true }))
}
function findCommonAncestor(elements: HTMLElement[]) {
let ancestor = elements[0]?.parentElement
while (ancestor && !elements.every(element => ancestor!.contains(element))) {
ancestor = ancestor.parentElement
}
return ancestor
}
export async function exportToPng(fileNameFormat: string) {
if (!checkIfConversationStarted()) {
alert(i18n.t('Please start a conversation first'))
return false
}
const effect = new Effect()
const conversationTurns = Array.from(document.querySelectorAll<HTMLElement>('#thread [data-testid^="conversation-turn-"]'))
const thread = findCommonAncestor(conversationTurns)
if (!thread || thread.children.length === 0 || thread.scrollHeight < 50) {
alert(i18n.t('Failed to export to PNG. Failed to find the element node.'))
return false
}
const isDarkMode = document.documentElement.classList.contains('dark')
const threadEl = thread as HTMLElement
const turnContainers = Array.from(threadEl.querySelectorAll<HTMLElement>('[data-turn-id-container][data-is-intersecting]'))
.filter(element => !!element.querySelector('[data-testid^="conversation-turn-"]') || element.offsetHeight > 0 || !!element.style.getPropertyValue('--last-known-height'))
const turnContainerIds = turnContainers
.map(element => element.dataset.turnIdContainer)
.filter((id): id is string => !!id && id !== 'client-created-root')
effect.add(() => {
threadEl.setAttribute('data-chatgpt-exporter-screenshot-root', '')
const style = document.createElement('style')
style.textContent = `
[data-chatgpt-exporter-screenshot-root],
#thread [data-testid^="conversation-turn-"] {
color: ${isDarkMode ? '#ececec' : '#0d0d0d'};
background-color: ${isDarkMode ? '#212121' : '#fff'};
}
/* https://github.com/niklasvh/html2canvas/issues/2775#issuecomment-1204988157 */
img {
display: initial !important;
}
pre {
margin-top: 8px !important;
}
pre > div > div > span {
margin-top: -12px;
padding-bottom: 2px;
}
#page-header,
#thread-bottom-container,
/* date separators such as "Yesterday 10:08 AM" */
[data-chatgpt-exporter-screenshot-root] [role="separator"],
/* any other elements that are not conversation turns */
[data-chatgpt-exporter-screenshot-root] > :not([data-turn-id-container]):not([data-testid^="conversation-turn-"]):not(:has([data-testid^="conversation-turn-"])),
/* hide back to top button */
button.absolute,
/* question button */
.group.absolute > button {
display: none;
}
/* Preserve the action row's spacing while hiding its toolbar. */
[data-testid^="conversation-turn-"] [role="group"]:has([data-testid="copy-turn-action-button"]),
/* code block buttons */
#thread pre button {
visibility: hidden;
}
/* Later user turns currently have much larger top padding than the first one. */
[data-testid^="conversation-turn-"][data-turn="user"] > h4 + div {
padding-top: 0 !important;
}
`
threadEl.appendChild(style)
return () => {
style.remove()
threadEl.removeAttribute('data-chatgpt-exporter-screenshot-root')
}
})
const scrollRoot = threadEl.closest<HTMLElement>('[data-scroll-root]')
if (scrollRoot) {
effect.add(() => {
const scrollTop = scrollRoot.scrollTop
const scrollLeft = scrollRoot.scrollLeft
const overflowAnchor = scrollRoot.style.overflowAnchor
scrollRoot.style.overflowAnchor = 'none'
return () => {
scrollRoot.style.overflowAnchor = overflowAnchor
scrollRoot.scrollTop = scrollTop
scrollRoot.scrollLeft = scrollLeft
}
})
}
effect.run()
// ChatGPT virtualizes off-screen turns. Visit every placeholder in DOM
// order and preserve a clone while each one is mounted. React cannot keep
// distant turns mounted at the same time, so the clones become our stable
// capture source.
const turnSnapshots = new Map<string, HTMLElement>()
if (scrollRoot && turnContainerIds.length > 0) {
for (const turnContainerId of turnContainerIds) {
for (let pass = 0; pass < 10; pass++) {
const container = Array.from(threadEl.querySelectorAll<HTMLElement>('[data-turn-id-container][data-is-intersecting]'))
.find(element => element.dataset.turnIdContainer === turnContainerId)
if (!container) break
const renderedTurn = container.querySelector<HTMLElement>('[data-testid^="conversation-turn-"]')
if (renderedTurn) {
turnSnapshots.set(turnContainerId, container.cloneNode(true) as HTMLElement)
break
}
scrollElementWithinRoot(scrollRoot, container, 'center')
await sleep(250)
}
if (!turnSnapshots.has(turnContainerId)) {
const placeholder = Array.from(threadEl.querySelectorAll<HTMLElement>('[data-turn-id-container][data-is-intersecting]'))
.find(element => element.dataset.turnIdContainer === turnContainerId)
if (placeholder) turnSnapshots.set(turnContainerId, placeholder.cloneNode(true) as HTMLElement)
}
}
}
else if (scrollRoot && conversationTurns[0]) {
scrollElementWithinRoot(scrollRoot, conversationTurns[0], 'start')
await sleep(250)
}
await sleep(500)
let screenshotEl = threadEl
if (turnSnapshots.size > 0) {
const staticThread = threadEl.cloneNode(false) as HTMLElement
staticThread.setAttribute('data-chatgpt-exporter-screenshot-root', '')
staticThread.style.position = 'absolute'
staticThread.style.left = '-100000px'
staticThread.style.top = '0'
staticThread.style.width = `${threadEl.offsetWidth}px`
staticThread.style.height = 'auto'
staticThread.style.minHeight = '0'
staticThread.style.maxHeight = 'none'
staticThread.style.overflow = 'visible'
staticThread.style.pointerEvents = 'none'
for (const turnContainerId of turnContainerIds) {
const snapshot = turnSnapshots.get(turnContainerId)
if (snapshot) staticThread.appendChild(snapshot)
}
effect.add(() => {
document.body.appendChild(staticThread)
return () => staticThread.remove()
})
effect.run()
screenshotEl = staticThread
await sleep(100)
}
effect.add(() => {
const minHeight = screenshotEl.style.minHeight
screenshotEl.style.minHeight = `${screenshotEl.scrollHeight}px`
return () => {
screenshotEl.style.minHeight = minHeight
}
})
effect.run()
await sleep(0)
const backgroundColor = isDarkMode ? '#212121' : '#fff'
const width = Math.max(screenshotEl.offsetWidth, screenshotEl.scrollWidth)
const height = Math.max(screenshotEl.offsetHeight, screenshotEl.scrollHeight)
let capture: Awaited<ReturnType<typeof snapdom>> | null = null
try {
capture = await snapdom(screenshotEl, {
embedFonts: true,
backgroundColor,
})
}
catch (error) {
console.error('Failed to capture screenshot DOM', error)
}
const sourceWidth = capture?.meta.vbW || width
const sourceHeight = capture?.meta.vbH || height
// Keep export quality consistent across operating systems, display
// densities and browser zoom levels. Only reduce it when a tile would
// exceed the browser's maximum canvas width.
const requestedScale = Math.min(2, MAX_SCREENSHOT_DIMENSION / sourceWidth)
const desiredWidth = Math.max(1, Math.floor(sourceWidth * requestedScale))
// Use one exact scale derived from SnapDOM's viewBox. Mixing live DOM
// dimensions with capture dimensions can stretch and blur every tile.
const desiredScale = desiredWidth / sourceWidth
const desiredHeight = Math.max(1, Math.floor(sourceHeight * desiredScale))
const takeTiledScreenshot = async (): Promise<Blob | null> => {
if (!capture) {
console.warn('[ChatGPT Exporter:screenshot] tiled capture unavailable')
return null
}
if (typeof CompressionStream === 'undefined') {
console.warn('[ChatGPT Exporter:screenshot] CompressionStream unavailable; using downscaled fallback')
return null
}
const tileHeight = Math.max(1, Math.min(
MAX_SCREENSHOT_DIMENSION,
Math.floor(MAX_TILE_PIXELS / desiredWidth),
))
try {
return await encodePng(desiredWidth, async (appendRows) => {
for (let targetY = 0; targetY < desiredHeight; targetY += tileHeight) {
const targetTileHeight = Math.min(tileHeight, desiredHeight - targetY)
const sourceY = sourceHeight * targetY / desiredHeight
const sourceBottom = sourceHeight * (targetY + targetTileHeight) / desiredHeight
const canvas = await capture!.toCanvas({
crop: {
x: 0,
y: sourceY,
width: sourceWidth,
height: sourceBottom - sourceY,
},
scale: desiredScale,
dpr: 1,
backgroundColor,
})
if (canvas.width !== desiredWidth) {
throw new Error(`Unexpected screenshot tile width: ${canvas.width}`)
}
const context = canvas.getContext('2d', { willReadFrequently: true })
if (!context) throw new Error('Failed to read screenshot tile')
const imageData = context.getImageData(0, 0, canvas.width, canvas.height)
await appendRows(imageData)
canvas.width = 1
canvas.height = 1
}
})
}
catch (error) {
console.error('Failed to encode tiled screenshot', error)
return null
}
}
const passLimit = 10
const takeDownscaledScreenshot = async (additionalScale = 1, currentPass = 1): Promise<Blob | null> => {
if (!capture) return null
const scale = Math.min(
requestedScale,
MAX_SCREENSHOT_DIMENSION / sourceWidth,
MAX_SCREENSHOT_DIMENSION / sourceHeight,
) * additionalScale
const targetWidth = Math.max(1, Math.floor(sourceWidth * scale))
const targetHeight = Math.max(1, Math.floor(sourceHeight * scale))
let canvas: HTMLCanvasElement | null = null
try {
canvas = await capture.toCanvas({
scale,
dpr: 1,
backgroundColor,
})
const context = canvas.getContext('2d')
if (context) context.imageSmoothingEnabled = false
const blob = await new Promise<Blob | null>(resolve => canvas!.toBlob(resolve, 'image/png', 1))
if (blob) return blob
}
catch (error) {
console.error('Failed to take screenshot', error)
}
// eslint-disable-next-line no-console
console.log(`ChatGPT Exporter:takeScreenshot with height=${height} width=${width} targetHeight=${targetHeight} targetWidth=${targetWidth}`)
/**
* A browser may return an empty canvas instead of throwing when a
* raster exceeds its canvas or SVG decode limit.
* See https://developer.mozilla.org/en-US/docs/Web/HTML/Element/canvas#maximum_canvas_size
* Chromium will not throw, we can only get an empty canvas
* Firefox will throw "DOMException: CanvasRenderingContext2D.scale: Canvas exceeds max size."
*/
if (currentPass > passLimit) return null
// 1.4 ^ 5 ~= 5.37, should be enough for most cases
return takeDownscaledScreenshot(additionalScale / 1.4, currentPass + 1)
}
const shouldTile = desiredHeight > MAX_SCREENSHOT_DIMENSION
|| desiredWidth * desiredHeight > MAX_SCREENSHOT_DIMENSION * MAX_SCREENSHOT_DIMENSION
let png = shouldTile ? await takeTiledScreenshot() : await takeDownscaledScreenshot()
if (!png && shouldTile) {
console.warn('[ChatGPT Exporter:screenshot] tiled export failed; using downscaled fallback')
png = await takeDownscaledScreenshot()
}
effect.dispose()
if (!png) {
alert('Failed to export to PNG. This might be caused by the size of the conversation. Please try to export a smaller conversation.')
return false
}
const chatId = getChatIdFromUrl() || undefined
const fileName = getFileNameWithFormat(fileNameFormat, 'png', { chatId })
downloadFile(fileName, 'image/png', png)
return true
}