Tantabus
Filters (Default) SettingsRegisterLogin
Rules Update - For a short summary on changes, check Here
Welcome! Please check this thread

Image Analysis and Feedback Thread

Creative Corner » Image Analysis and Feedback Thread Search Posts
Started by AIPonyAnon
Posted
20 replies
Subscribe
AIPonyAnon

I’m creating this thread in the hope that we can get people to give each other useful feedback on the images that they create. This should help people to develop their own critical eye.
I was going to use ‘critique’ in the title, but that implies a level of interpretation. The intention here is to give reasonably objective feedback on errors or elements that detract from an image, and well-done things that improve the image.
If the creator has some stated goals or issues with the image, then feedback on how they might better reach these goals or solve these issues can also be given. Creators specifically asking for opinions on certain things is also fine.
In your analysis and feedback you should include, where possible, ideas and techniques for fixing the issues that you point out.
As an example, here’s one my recent gens:
Looking for basic errors we can see the following:
  • Cadance does not appear to have any sign of wings.
  • The bottom right of the inset section on the door behind Cadance is missing.
  • Her hooves (hoof shoes?) have slightly different shapes and embellishments.
  • Cadance’s tail is exceptionally long.
All of these issue can be fixed by starting with a different base gen, or by overpainting/inpainting/detailing.
A positive of the image is that the armor, clothing, and regalia all seem very coherent.
One issue I have is that I feel like the image is just sort of boring. Causes and solutions for this might include:
Problem Solutions
1girl standing Use more varied pose descriptions. Add more characters.
Basic, head-on camera angle Use a more interesting camera angle.
Basic lighting Use more interesting lighting descriptions (not well tagged, requires experimentation).
Background that is simultaneously dull yet distracting Use a different gen. Inpaint out distracting features. Describe the background differently.
No objects of interest. A weapon, shield, goblet, etc. might add some interest.
As I’m using a Noob based model, useful tags for things such as pose and camera-angle can be found in the e6 and danbooru tag wikis.
Some things I would like feedback from others on are:
  • Is the color grading okay? (difficult because I and many people do not have color accurate monitors)
  • What do people think of the face? I feel like it has a somewhat porcine influence to it and I’m not sure if I like it.
  • Is white and red a bad color combination with Cadance’s fur?
Please try to include feedback for others when you post a request for feedback. This should help to keep a good balance between giving and requesting feedback.
Posted Report
Edited
Star Streaker

I’m not terribly familiar with the Noob models, mostly stick to the PonyXL variants myself. I’m also by no means an artist so take what I say with a grain of salt.
Overall the image is pretty good, like you said if a little stale and bland. My impressions below.
Is white and red a bad color combination with Cadance’s fur? Is the color grading okay? (difficult because I and many people do not have color accurate monitors)
  • I’m fine with the red and white Cadance is wearing but what I think gets to me is Cadance is a little oversaturated. She somewhat harshly sticks out with the rest of the image (which is more painterly and muted) and I think that saturation is causing some artifacts particularly around the gold plate she’s wearing.
  • Another thing I noticed was that the doorway seems itself split in half with the line being covered up by Cadance herself. So one half of the door appears to be one style while the other half (Cadance’s left) has a mirror image but slightly different. On that same side, there’s a sort of knob/handle type thing just jutting out of the wall. What’s it for? The poster on the left (viewer’s pov) may make sense.
What do people think of the face? I feel like it has a somewhat porcine influence to it and I’m not sure if I like it.
  • I’d agree and I think it’s largely due to the nostrils/first person POV. From what I’ve noticed with my admittedly limited artistic talent is that pony faces when facing “forward” tend to point up slightly. Cadance’s muzzle is definitely pointing down but the face could be fixed with an inpaint or adetailer model.
Posted Report
Zerowinger

3-3/4" Army Man Fan
So this one is a bit of a problem that made me hem and haw a bit about uploading:
>>33453t (merged)
Night Glider has some of her hair sticking out of her helmet where it really shouldn’t be. I guess it could be argued there’s a slit for her mane in the back of her helmet that it’s sticking out from and so being blown along with her scarf, but that’s not a very realistic answer.
Google Imagen (or at least as used on NightCafe) unfortunately doesn’t allow for selective edits to select that errant hair for editing out. Would using a program like Gimp or Photoshop be a solution?
Hair sticking out of helmets also seems to be a frequent experience with some of these generations of mine, so would adjustment of negative prompts possibly help?
AIPonyAnon

@Zerowinger
You might be able to find some trick with your prompts that removes or reduces the issue, but that seems unlikely. This seems to be a fundamental misunderstanding of the model (or it had a lot of images with this mistake in its training data).
Can you do image-to-image with your current setup? If so, then doing a very basic edit of the image with any old image editor (even mspaint works fine) to just remove the hair and block the background colors in and then doing one or two passes with low-denoise image-to-image should fix it.
If you can’t image-to-image, then Photoshop is probably your best bet. I haven’t used the recent versions of Photoshop, but the tools in it should be able to handle removing things like that very quickly and easily. It’s doable in GIMP, but it’ll be a lot more work because its selection and fixing tools are not as advanced.
You could also try https://huggingface.co/spaces/Sanster/iopaint-lama which uses a model specifically trained on removing things from images, though you’ll likely want to do an image-to-image pass after using it.
Posted Report
AIPonyAnon

@Star Streaker
She somewhat harshly sticks out with the rest of the image (which is more painterly and muted)
I don’t think Cadance is oversaturated in the technical sense of the term, but I think I get what you mean. The brown of the background doesn’t highlight her, rather it makes her look out of place.
I hadn’t noticed that she does look slightly less painterly than the background when viewed from a distance, thanks for pointing that out. It’s been a struggle with all of the SDXL based models to get good, consistent painterly styles.
Another thing I noticed was that the doorway seems itself split in half with the line being covered up by Cadance herself.
Aaaah, consistent background details are a bane. Trying to get anything to maintain symmetry when split by a character is such a pain.
I’d agree and I think it’s largely due to the nostrils/first person POV. From what I’ve noticed with my admittedly limited artistic talent is that pony faces when facing “forward” tend to point up slightly. Cadance’s muzzle is definitely pointing down but the face could be fixed with an inpaint or adetailer model.
Ah, thanks for pointing that out, the bridge of the muzzle is definitely pointing down too much. I knew something was up, but couldn’t quite figure out what it was. I do detail the face and eyes, but that comes with its own problems. I think this one in particularly is probably a result of cropping too close when detailing and using meesh in the artist mix, as he tends to draw more sloped muzzles.
Posted Report
Star Streaker

@AIPonyAnon
I’m kinda curious, what kind of AI gen workflow do you use? I largely use SD Forge but have been intending to get into Comfy.
The door itself is actually relatively easy to fix. I haven’t posted one image yet but there was a Canterlot street view image I generated with a terrible cobblestone path. So I used inpainting and img-img and specified the prompt to only focus on exclusively the road and it gave a much more coherent road while preserving the subjects. You could also do the same for inpainting Cadance’s wings though you’d likely want to use soft inpainting for those.
Posted Report
Zerowinger

3-3/4" Army Man Fan
@AIPonyAnon
Looking for basic errors we can see the following:
  • Cadance does not appear to have any sign of wings.
I know I use a very different model from yourself and everyone else, but one thing I’ve noticed that gets the Imagen generator to put wings on characters that should have them is to include prompts like “spread wings” or “wings folded in,” or some other prompt or detail that focuses on wings. Perhaps something similar may work with this checkpoint?
Posted Report
AIPonyAnon

@Zerowinger
Yeah, that usually works quite well. With both Pony and Noob just adding wings will usually fix it. spread wings and folded wings can help with positioning, but a lot of the time the model will just do what it wants with them.
Posted Report
Zerowinger

3-3/4" Army Man Fan
@AIPonyAnon
Well, it is a very unusual image of Dash. Her chest fur and mane kind of look like feathers, plus having that extra splotch of color on her chest might throw people off. Plus the color for the surrounding scenery is kind of strange and gives it a bit of a surreal look, I don’t know if that’s what you were going for or not.
Posted Report
AIPonyAnon

@Zerowinger
Yeah, I was going for overly colorful and surreal. The rainbow chest fluff was a side product of bumping the colorfulness up, but I thought it looked cool.
I’m not so much worried about it being ignored because it’s unusual, but rather that there’s some issue with it that I’m not perceiving. It not being to the taste of the users here is fine, if that is the case.
Posted Report
Star Streaker

@AIPonyAnon
I’d say no, there’s not anything terribly “wrong” with it. Rainbow’s forelegs look a little strange but overall I don’t see anything that’d “break” the image. The surreal style is pretty good and consistent throughout.
Posted Report
Background Pony #FE39
@AIPonyAnon
A lot of people into AI are very, uh, “biased towards horny”.
Could be a good idea to keep this thread populated with SFW stuff you’re proud of (and maybe promote the thread around, on images or elsewhere). Plus, can subscribe to threads.
Posted Report
AIPonyAnon

Something felt compositionally off in this one:
I realized it after upscaling and couldn’t get any outpainting to work well. I had hoped to add another section on the right-hand side in the hopes that it would balance the image a little better. Would be interested to hear if anyone has thoughts on if/why this has compositional issues, and ways you might fix it.
Posted Report
Zerowinger

3-3/4" Army Man Fan
@AIPonyAnon
Aside from the fact that Shining looks like he’s leaning to one side, I don’t see any issues with the image. I like the armor design you got on him and how it has his cutie mark as an emblem.
Posted Report
AIPonyAnon

@Zerowinger
I hadn’t considered the lean being a little weird. He does look like he’s walking a little too leisurely to be leaning that much to turn.
The Noob based models seem much better at weapons and armor than other SDXL/1.5 models that I’ve used. The models in general also like to use cutiemarks in places where you might find an emblem.
Posted Report
Edit Preview
Syntax quick reference: **bold** *italic* ||hide text|| `code` __underline__ ~~strike~~ ^sup^ ~sub~

Detailed syntax guide