[BUG] - Does not read Invoke 3 SDXL data
Void2258 opened this issue · comments
Description
Nothing comes up for images generated in Invoke 3 using SDXL. All fields show blank. "No data selected or unsupported format".
Image file
Example image:
Yes I know the image is kinda bad; it was a test for my new card to see gen times.
Invoke metadata shown internally for comparison:
{
"app_version": "3.1.1",
"generation_mode": "sdxl_txt2img",
"positive_prompt": "supergirl flying through the sky, billowing cape, city background",
"negative_prompt": "Multiple, several, bad hands, extra arms",
"width": 1024,
"height": 1024,
"seed": 2507332253,
"rand_device": "cpu",
"cfg_scale": 7.5,
"steps": 25,
"scheduler": "ddpm",
"clip_skip": 0,
"model": {
"model_name": "dreamshaperXL10_alpha2Xl10",
"base_model": "sdxl",
"model_type": "main"
},
"controlnets": [],
"loras": [],
"positive_style_prompt": "Comic, drawn, animated, high quality, masterpiece, artgerm",
"negative_style_prompt": "Realistic, poor quality, warped, deformed"
}
Version
1.3.3Post1
OS
Windows
Tool
InvokeAI
To the @receyuki :
import re
def split_string(text):
patterns = [r"(?<=,)\s*\([^)]+\)\s*(?=,)", r"\w+\s*\([^)]+\)\s*\w+"]
matches = []
for pattern in patterns:
matches += re.findall(pattern, text)
if matches:
for match in matches:
text = text.replace(match, "").strip()
result = text.split(",")
result = [item.strip() for item in result if item]
result = matches + result
return result
image_dict = {
"invokeai_metadata": {
"app_version": "3.1.1",
"generation_mode": "sdxl_txt2img",
"positive_prompt": "supergirl flying through the sky, billowing cape, city background",
"negative_prompt": "Multiple, several, bad hands, extra arms",
"width": 1024,
"height": 1024,
"seed": 2507332253,
"rand_device": "cpu",
"cfg_scale": 7.5,
"steps": 25,
"scheduler": "ddpm",
"clip_skip": 0,
"model": {
"model_name": "dreamshaperXL10_alpha2Xl10",
"base_model": "sdxl",
"model_type": "main",
},
"controlnets": [],
"loras": [],
"positive_style_prompt": "Comic, drawn, animated, high quality, masterpiece, artgerm",
"negative_style_prompt": "Realistic, poor quality, warped, deformed",
}
}
def extract_image_prompt(image_data):
data = image_data.get("invokeai_metadata")
image_prompt = {
"positive": split_string(data.get("positive_prompt")),
"negative": split_string(data.get("negative_prompt")),
"parameter": {
"model": data.get("model", {}).get("model_name"),
"sampler": data.get("scheduler"),
"seed": data.get("seed"),
"cfg": data.get("cfg_scale"),
"steps": data.get("steps"),
"size": f'{data.get("width")}x{data.get("height")}',
},
}
return image_prompt
extract_image_prompt(image_dict)output:
{
"positive": ["supergirl flying through the sky", "billowing cape", "city background"],
"negative": ["Multiple", "several", "bad hands", "extra arms"],
"parameter": {
"model": "dreamshaperXL10_alpha2Xl10",
"sampler": "ddpm",
"seed": 2507332253,
"cfg": 7.5,
"steps": 25,
"size": "1024x1024"
}
}
@Void2258
I've just realized that Invoke3 has a completely different metadata structure, and it can't even properly import images generated by Invoke2 lol. I will be adding support for Invoke3 in the next release.
Supposedly the import issues for invoke 2 into invoke 3 have been addressed in 3.2
https://github.com/receyuki/stable-diffusion-prompt-reader/releases/tag/v1.3.4-beta.1
I just released 1.3.4-beta.1. Try it out.
Opens invoke 3 images, but almost everything is in a big blob in the bottom box (settingS).
However, now doesn't even respond to invoke 2 images being dragged in and just opens nothing if you try the browse function for them. Invoke 2 images open fine in older version.
About Invoke 2, please upload any images that cannot be read here. I have refactored the code for InvokeAI, so there may be new issues that were not present before.
For Invoke 3, as mentioned in the release notes, previous versions only extracted essential metadata. However, I believe some users may need additional data beyond the essentials, such as control nets or LoRAs. Therefore, I have decided to directly display the original settings. Simple mode will be a better choice if you only need essential metadata. Additionally, you can click on the text box in simple mode to copy a single parameter value.
It's all of my invoke 2 images. I have several hundred and while I haven't tried every single one, every one I have tried does not work. Here is just one item.
I guess the thing with invoke 3 is there is more "essential" stuff, like style prompts, than there was before. So the big block at the end carries 99% of things and the other boxes look empty.
I must say, InvokeAI has made a mess of ensuring metadata compatibility, and once again, this image contains a format I've never seen before. I will fix this issue in beta2.
As for the new display format, you're right, style_prompt and refiner are making the settings very long. I thought the simple mode would be sufficient for most users, which is why I made this change. Additionally, in version 1.4.x, I will add an option to modify the default display mode so that users can have the simple mode as their default.
You might want to go to the invoke discord and discuss things with them. They are very responsive to feedback and willing to help out with issues.
I originally planned to wait for comfyUI node to be ready, but it seems like it might take a while, so I released a beta2.
https://github.com/receyuki/stable-diffusion-prompt-reader/releases/tag/v1.3.4-beta.2