i/Aether
1
0
Fork 0
mirror of https://github.com/fawney19/Aether.git synced 2026-09-10 13:10:21 +08:00
Code Issues Packages Projects Releases Wiki Activity

调整 ChatGPT Web 生图 token 估算口径

Browse Source
This commit is contained in:
Codex
2026-05-24 18:50:14 +08:00
parent 3bf7469d30
commit 74c8bfc59f
1 changed files with 252 additions and 25 deletions
Download Patch File Download Diff File Expand all files Collapse all files
apps/aether-gateway/src/execution_runtime/chatgpt_web_image.rs
+252 -25
View File
@@ -46,6 +46,11 @@ const CHATGPT_WEB_SEC_CH_UA: &str =
const CHATGPT_WEB_BROWSER_PROFILE: &str = "chrome143";
const CHATGPT_WEB_QUOTA_REFRESH_TIMEOUT_MS: u64 = 30_000;
const CHATGPT_WEB_QUOTA_REFRESH_PROXY_TIMEOUT_MS: u64 = 60_000;
const GPT_IMAGE2_TOKEN_MIN_PIXELS: u64 = 655_360;
const GPT_IMAGE2_TOKEN_MAX_PIXELS: u64 = 8_294_400;
const GPT_IMAGE2_TOKEN_MAX_EDGE: u64 = 3_840;
const GPT_IMAGE2_TOKEN_MAX_ASPECT_RATIO: u64 = 3;
const GPT_IMAGE2_PARTIAL_IMAGE_OUTPUT_TOKENS: u64 = 100;
pub(crate) struct ChatGptWebImageStream {
pub(crate) frame_stream: BoxStream<'static, Result<Bytes, IoError>>,
@@ -320,6 +325,8 @@ struct ChatGptWebImageRequest {
size: String,
ratio: String,
output_format: String,
quality: Option<String>,
partial_images: u64,
images: Vec<String>,
}
@@ -349,6 +356,8 @@ impl ChatGptWebImageRequest {
size: text("size").unwrap_or_else(|| "1024x1024".to_string()),
ratio: text("ratio").unwrap_or_else(|| "1:1".to_string()),
output_format: text("output_format").unwrap_or_else(|| "png".to_string()),
quality: text("quality"),
partial_images: json_u64(body.get("partial_images")).unwrap_or(0),
images,
})
}
@@ -2115,7 +2124,7 @@ fn chatgpt_web_image_usage(
report_context: Option<&Value>,
) -> (Value, Value) {
let input_tokens = chatgpt_web_image_input_tokens(request, report_context);
let estimated_output_tokens = chatgpt_web_image_output_tokens(image);
let estimated_output_tokens = chatgpt_web_image_output_tokens(request, image, report_context);
let usage = json!({
"input_tokens": input_tokens,
"output_tokens": estimated_output_tokens,
@@ -2130,8 +2139,8 @@ fn chatgpt_web_image_usage(
},
"output_tokens": estimated_output_tokens,
"output_tokens_details": {
"image_tokens": 0,
"text_tokens": estimated_output_tokens
"image_tokens": estimated_output_tokens,
"text_tokens": 0
},
"total_tokens": input_tokens.saturating_add(estimated_output_tokens),
}
@@ -2147,8 +2156,86 @@ fn chatgpt_web_image_input_tokens(
estimate_text_tokens(prompt.as_str())
}
fn chatgpt_web_image_output_tokens(image: &DownloadedImage) -> u64 {
estimate_text_tokens(image.b64_json.as_str())
fn chatgpt_web_image_output_tokens(
request: &ChatGptWebImageRequest,
image: &DownloadedImage,
report_context: Option<&Value>,
) -> u64 {
let quality = chatgpt_web_image_quality(request, report_context);
let size = chatgpt_web_image_size(request, image, report_context);
let partial_images = chatgpt_web_image_partial_images(request, report_context);
let base_tokens = size
.map(|(width, height)| gpt_image2_output_tokens(width, height, quality.as_str()))
.unwrap_or_else(|| gpt_image2_output_tokens(1024, 1024, quality.as_str()));
base_tokens
.saturating_add(partial_images.saturating_mul(GPT_IMAGE2_PARTIAL_IMAGE_OUTPUT_TOKENS))
}
fn chatgpt_web_image_quality(
request: &ChatGptWebImageRequest,
report_context: Option<&Value>,
) -> String {
let candidate = [
chatgpt_web_report_context_image_request_text(report_context, "quality"),
chatgpt_web_report_context_original_request_text(report_context, "quality"),
request.quality.clone(),
]
.into_iter()
.flatten()
.find(|value| !value.is_empty())
.unwrap_or_else(|| "medium".to_string());
normalize_gpt_image2_quality(candidate.as_str())
}
fn chatgpt_web_image_size(
request: &ChatGptWebImageRequest,
image: &DownloadedImage,
report_context: Option<&Value>,
) -> Option<(u64, u64)> {
if let Some(candidate) = downloaded_image_dimensions(image)
.filter(|(width, height)| gpt_image2_dimensions_are_plausible(*width, *height))
{
return Some(candidate);
}
let candidates = [
chatgpt_web_report_context_image_request_text(report_context, "size")
.and_then(|value| parse_gpt_image2_size(value.as_str())),
chatgpt_web_report_context_original_request_text(report_context, "size")
.and_then(|value| parse_gpt_image2_size(value.as_str())),
parse_gpt_image2_size(request.size.as_str()),
];
for candidate in candidates.into_iter().flatten() {
if gpt_image2_dimensions_are_valid(candidate.0, candidate.1) {
return Some(candidate);
}
}
let ratio = chatgpt_web_image_ratio(request, report_context);
Some(chatgpt_web_fallback_size_for_ratio(ratio.as_str()))
}
fn chatgpt_web_image_partial_images(
request: &ChatGptWebImageRequest,
report_context: Option<&Value>,
) -> u64 {
chatgpt_web_report_context_image_request_u64(report_context, "partial_images")
.or_else(|| {
chatgpt_web_report_context_original_request_u64(report_context, "partial_images")
})
.unwrap_or(request.partial_images)
}
fn chatgpt_web_image_ratio(
request: &ChatGptWebImageRequest,
report_context: Option<&Value>,
) -> String {
chatgpt_web_report_context_image_request_text(report_context, "ratio")
.or_else(|| chatgpt_web_report_context_original_request_text(report_context, "ratio"))
.or_else(|| {
chatgpt_web_report_context_original_request_text(report_context, "aspect_ratio")
})
.unwrap_or_else(|| request.ratio.clone())
}
fn chatgpt_web_report_context_image_request_text(
@@ -2164,6 +2251,16 @@ fn chatgpt_web_report_context_image_request_text(
.map(ToOwned::to_owned)
}
fn chatgpt_web_report_context_image_request_u64(
report_context: Option<&Value>,
key: &str,
) -> Option<u64> {
report_context
.and_then(|value| value.get("image_request"))
.and_then(|value| value.get(key))
.and_then(|value| json_u64(Some(value)))
}
fn chatgpt_web_report_context_original_request_text(
report_context: Option<&Value>,
key: &str,
@@ -2175,6 +2272,16 @@ fn chatgpt_web_report_context_original_request_text(
})
}
fn chatgpt_web_report_context_original_request_u64(
report_context: Option<&Value>,
key: &str,
) -> Option<u64> {
let original = report_context?.get("original_request_body")?;
json_u64(original.get(key)).or_else(|| {
chatgpt_web_original_image_tool_value(original, key).and_then(|value| json_u64(Some(value)))
})
}
fn chatgpt_web_original_image_tool_value<'a>(original: &'a Value, key: &str) -> Option<&'a Value> {
original
.get("tools")
@@ -2206,6 +2313,107 @@ fn value_text(value: Option<&Value>) -> Option<String> {
.map(ToOwned::to_owned)
}
fn downloaded_image_dimensions(image: &DownloadedImage) -> Option<(u64, u64)> {
match (image.width, image.height) {
(Some(width), Some(height)) if width > 0 && height > 0 => {
Some((width as u64, height as u64))
}
_ => None,
}
}
fn gpt_image2_dimensions_are_plausible(width: u64, height: u64) -> bool {
let pixels = width.saturating_mul(height);
if !(GPT_IMAGE2_TOKEN_MIN_PIXELS..=GPT_IMAGE2_TOKEN_MAX_PIXELS).contains(&pixels) {
return false;
}
let max_edge = width.max(height);
let min_edge = width.min(height);
if max_edge > GPT_IMAGE2_TOKEN_MAX_EDGE {
return false;
}
if max_edge > min_edge.saturating_mul(GPT_IMAGE2_TOKEN_MAX_ASPECT_RATIO) {
return false;
}
true
}
fn gpt_image2_dimensions_are_valid(width: u64, height: u64) -> bool {
width % 16 == 0 && height % 16 == 0 && gpt_image2_dimensions_are_plausible(width, height)
}
fn normalize_gpt_image2_quality(value: &str) -> String {
match value.trim().to_ascii_lowercase().as_str() {
"low" => "low".to_string(),
"medium" | "standard" | "auto" => "medium".to_string(),
"high" | "hd" => "high".to_string(),
_ => "medium".to_string(),
}
}
fn parse_gpt_image2_size(size: &str) -> Option<(u64, u64)> {
let normalized = size.trim().to_ascii_lowercase().replace('×', "x");
let (width, height) = normalized.split_once('x')?;
let width = width
.trim()
.parse::<u64>()
.ok()
.filter(|value| *value > 0)?;
let height = height
.trim()
.parse::<u64>()
.ok()
.filter(|value| *value > 0)?;
Some((width, height))
}
fn chatgpt_web_fallback_size_for_ratio(ratio: &str) -> (u64, u64) {
match ratio.trim() {
"3:2" => (1216, 832),
"2:3" => (832, 1216),
"4:3" => (1152, 864),
"3:4" => (864, 1152),
"5:4" => (1120, 896),
"4:5" => (896, 1120),
"16:9" => (1344, 768),
"9:16" => (768, 1344),
"21:9" => (1536, 640),
_ => (1024, 1024),
}
}
// Estimate GPT Image 2 image-token output using the same dimensions and quality
// drivers as OpenAI's public cost calculator. This intentionally ignores the
// base64 response length, which is only a transport encoding.
fn gpt_image2_output_tokens(width: u64, height: u64, quality: &str) -> u64 {
let quality_scale = match quality.trim().to_ascii_lowercase().as_str() {
"low" => 16u64,
"high" => 96u64,
_ => 48u64,
};
let long = width.max(height);
let short = width.min(height);
let short_scale = round_div_u64(quality_scale.saturating_mul(short), long);
let (long_scale, short_scale) = if width >= height {
(quality_scale, short_scale)
} else {
(short_scale, quality_scale)
};
let latent_pixels = u128::from(long_scale).saturating_mul(u128::from(short_scale));
let image_pixels = u128::from(width).saturating_mul(u128::from(height));
let numerator =
latent_pixels.saturating_mul(u128::from(2_000_000u64).saturating_add(image_pixels));
let tokens = (numerator.saturating_add(4_000_000u128 - 1)) / 4_000_000u128;
u64::try_from(tokens).unwrap_or(u64::MAX)
}
fn round_div_u64(numerator: u64, denominator: u64) -> u64 {
if denominator == 0 {
return 0;
}
numerator.saturating_add(denominator / 2) / denominator
}
fn estimate_text_tokens(text: &str) -> u64 {
let chars = text.chars().count() as u64;
if chars == 0 {
@@ -2215,6 +2423,23 @@ fn estimate_text_tokens(text: &str) -> u64 {
}
}
fn json_u64(value: Option<&Value>) -> Option<u64> {
value.and_then(|value| {
value
.as_u64()
.or_else(|| {
value
.as_i64()
.and_then(|number| (number >= 0).then_some(number as u64))
})
.or_else(|| {
value
.as_str()
.and_then(|number| number.trim().parse::<u64>().ok())
})
})
}
fn build_failed_sse(request: &ChatGptWebImageRequest, failure: &Value) -> String {
let failed = if failure.get("type").and_then(Value::as_str) == Some("response.failed") {
failure.clone()
@@ -2889,9 +3114,17 @@ mod tests {
.expect("completed response should be present")
}
#[test]
fn gpt_image2_output_token_estimator_matches_pricing_calculator_examples() {
assert_eq!(gpt_image2_output_tokens(1024, 1024, "low"), 196);
assert_eq!(gpt_image2_output_tokens(1024, 1024, "medium"), 1756);
assert_eq!(gpt_image2_output_tokens(1536, 1024, "medium"), 1372);
assert_eq!(gpt_image2_output_tokens(1024, 1536, "medium"), 1372);
assert_eq!(gpt_image2_output_tokens(1024, 1024, "high"), 7024);
}
#[test]
fn chatgpt_web_success_sse_includes_estimated_image_usage() {
let output_text = "aGVsbG8=".repeat(128);
let request = ChatGptWebImageRequest {
model: "gpt-image-2".to_string(),
web_model: "gpt-5-5-thinking".to_string(),
@@ -2899,10 +3132,12 @@ mod tests {
size: "1024x1024".to_string(),
ratio: "1:1".to_string(),
output_format: "png".to_string(),
quality: Some("low".to_string()),
partial_images: 0,
images: Vec::new(),
};
let image = DownloadedImage {
b64_json: output_text.clone(),
b64_json: "aGVsbG8=".repeat(128),
mime: "image/png".to_string(),
width: Some(1024),
height: Some(1024),
@@ -2919,7 +3154,7 @@ mod tests {
);
let completed = completed_response_from_sse(body.as_str());
let input_tokens = estimate_text_tokens("draw a test image");
let output_tokens = estimate_text_tokens(output_text.as_str());
let output_tokens = 196;
assert_eq!(completed["usage"]["input_tokens"], json!(input_tokens));
assert_eq!(completed["usage"]["output_tokens"], json!(output_tokens));
@@ -2932,7 +3167,7 @@ mod tests {
json!(input_tokens)
);
assert_eq!(
completed["tool_usage"]["image_gen"]["output_tokens_details"]["text_tokens"],
completed["tool_usage"]["image_gen"]["output_tokens_details"]["image_tokens"],
json!(output_tokens)
);
assert_eq!(
@@ -2942,8 +3177,7 @@ mod tests {
}
#[test]
fn chatgpt_web_success_sse_counts_output_text_not_dimensions() {
let output_text = "iVBORw0KGgoAAAANSUhEUgAA".repeat(64);
fn chatgpt_web_success_sse_uses_image_dimensions_not_output_text() {
let request = ChatGptWebImageRequest {
model: "gpt-image-2".to_string(),
web_model: "gpt-5-5-thinking".to_string(),
@@ -2951,10 +3185,12 @@ mod tests {
size: "1024x1024".to_string(),
ratio: "1:1".to_string(),
output_format: "png".to_string(),
quality: Some("low".to_string()),
partial_images: 0,
images: Vec::new(),
};
let image = DownloadedImage {
b64_json: output_text.clone(),
b64_json: "iVBORw0KGgoAAAANSUhEUgAA".repeat(64),
mime: "image/png".to_string(),
width: Some(1402),
height: Some(1122),
@@ -2973,7 +3209,7 @@ mod tests {
assert_eq!(
completed["usage"]["output_tokens"],
json!(estimate_text_tokens(output_text.as_str()))
json!(gpt_image2_output_tokens(1402, 1122, "low"))
);
}
@@ -3428,16 +3664,12 @@ data: [DONE]
assert!(body.contains("\"height\":3"));
let expected_output_text =
base64::engine::general_purpose::STANDARD.encode(png_header_bytes(2, 3));
let expected_output_tokens = estimate_text_tokens(expected_output_text.as_str());
assert!(body.contains(&expected_output_text));
let completed = completed_response_from_sse(body.as_str());
assert_eq!(
completed["usage"]["output_tokens"],
json!(expected_output_tokens)
);
assert_eq!(completed["usage"]["output_tokens"], json!(1756));
assert_eq!(
completed["tool_usage"]["image_gen"]["output_tokens"],
json!(expected_output_tokens)
json!(1756)
);
handle.abort();
@@ -3503,11 +3735,6 @@ data: [DONE]
assert!(decoded_data.contains("\"width\":2"));
assert!(decoded_data.contains("\"height\":3"));
assert!(text.contains("\"type\":\"eof\""));
let expected_output_tokens = estimate_text_tokens(
base64::engine::general_purpose::STANDARD
.encode(png_header_bytes(2, 3))
.as_str(),
);
let eof_frame = text
.lines()
.filter_map(|line| serde_json::from_str::<Value>(line).ok())
@@ -3520,7 +3747,7 @@ data: [DONE]
.and_then(|summary| summary.get("standardized_usage"))
.and_then(|usage| usage.get("output_tokens"))
.and_then(Value::as_i64),
i64::try_from(expected_output_tokens).ok()
Some(1756)
);
assert_eq!(
eof_frame