Getting different (and much worse) results using the API and I can't figure out why.

[Piprian]

My results using the API consistently look worse than what I am getting using the same settings and prompt in the webui. (As far as I can tell)

For example this is what my results from the API always look like:

And this is what the same prompt and settings look like in the webui:

This is what my settings look like in the webui:

This is my settings when I'm using the API:
`"enable_hr": True,

"hr_scale": 2,

"hr_upscaler": "Latent",

"denoising_strength": 0.55,

"firstphase_width": 512,

"firstphase_height": 512,

"prompt": first_message,

"batch_size": 2,

"n_iter": 1,

"steps": 80,

"cfg_scale": 13,

"negative prompt": "bad-hands-5, badhandv4, easynegative",

"override_settings": {"CLIP_stop_at_last_layers": 2},

"sampler_index": "DDIM"`

It kinda looks like it's using a different model? I have previously successfully changed the model in the webui to get different results in the API so I assume it uses the webui setting anyway?
I added the model to the API call and it didn't help either.

[missionfloyd]

Your images lack metadata, so we can't see what model it's using. Here's how to set the model with the API.

option_payload = {
    "sd_model_checkpoint": "Anything-V3.0-pruned.ckpt [2700c435]",
}

response = requests.post(url="http://127.0.0.1:7860/sdapi/v1/options", json=option_payload)

[Piprian]

Thanks for the help!

That's weird. I didn't disable metadata. I'm guessing something about the way my bot converts the image strips it accidently?

This is how I tried to set the model:

I guess either that isn't right or it's not the model because it didn't fix the problem.

Sorry. I really don't know what I am doing...

Edit: I think it's probably not using the wrong model because it doesn't load a different model when I use the API right after using the webui.

Edit 2: I figured out how to print the "parameters" from the API response. It told me that it was using the correct model but I misspelled "negative_prompt" in my initial payload and it was generating without a negative prompt.
After fixing that, the results are now comparable to the ones I get from the webui! I am surprised how much of a difference there is from just removing the three embeddings I used as a negative.

Thank you for the help anyways!

[w-e-w]

tip, enable --enable-console-prompts
this should show the generation parameters in the console, useful for debugging

[Piprian]

Thank you!
I think the print parameters thing I found did roughly the same but this one might have more information.