This damn gen. This image was an insane comedy of errors the whole process and I'll be happy if I never have to go through something like this ever again. I was so close to just abandoning the project several times because there were just problems out the ass the whole way through, but hey, now I'm finally done and I solved an issue I was having for a while in the process. Let me walk you through what happened.
Initially, I had the idea for this gen, but it was going to be solo Leafan talking to herself in a mirror. Right away I knew there would be 3 big hurdles with that, being getting everything consistent between Leafan and her reflection, deciding what the background should be, and arranging everything in a way that would look right with the massive text bubble. I tried a few rawgens to get the different angles I needed but it didn't take long for me to decide to change plans. Oh well, happens.
Next I thought "okay, how about instead of talking to herself, she's talking to another pal?". I thought it made a little more sense, and it would get the point across just the same. Sekhmet seemed like a good choice for the other pal, so I downloaded some loras and started rawgenning again. Eventually I had 2 rawgens I settled on, one for Leafan and one for Sekhmet.
However, there was an issue that I've been dealing with for a long time that was just especially bad with this artstyle. My img2img outputs were coming out very grainy and low quality, like when you save an image as jpg at medium quality. It looked compressed. In my earlier gens I would mostly dance around this by either smudging out the compression myself or just fiddling around with some settings to reduce the effect, but with the dark color of Leafan's body and the no lineart style of her it was extremely noticeable. The upscales were unsalvageable as is, so I knew I had no choice but to finally put belt to ass and find out the root cause of this problem and fix it.
In the meantime though, I still had a lot of work to do actually writing the dialogue and arranging the final image. So, just because I didn't feel like diagnosing a technical issue at the time, I loaded up a new gimp project and starting arranging the character and typesetting everything. This took me about an hour. I had it originally that Leafan's text bubble would completely surround Sehkmet to emphasize that she's not shutting up, but it really reduced the readability of her dialogue. So I restarted essentially from scratch. At this point I was pretty frustrated so I decided to switch gears again and just troubleshoot the tech problem instead. That way I'd have actual clean images to work with too, instead of low quality placeholders.
I spent another few hours playing around with settings. I switched models, adjusted samplers, denoise, resolution, CFG, VAE, pretty much anything with a sliderm, dropdown, or checkbox in forge. In the end I had a few leads but nothing that conclusively solved the issue. Then I made my biggest mistake. I asked grok for help.
Grok suggested it could be a VAE issue, which made sense to me since VAE operations are lossy. In the end, I tried reloading the UI with a flag that forces full precision during the gen process. UI loaded fine, so I reset all my img2img settings to their usual values and started an upscale to test how it worked. It didn't work at all. It bluescreened by laptop.
Now here's the thing. I never saved gimp. I still had the project open since I was working on it on the side, and I didn't realize I never saved it. So all the work I did to write the dialogue, draw the speech bubbles, and arrange everything was lost. I would be starting essentially from scratch. I was several hours deep into this at this point so I just ragequit and went to bed.
A few days later I did some more troubleshooting and eventually found the cause of the img2img issue. It was so stupidly simple. There's a setting in forge called "apply color correction to img2img outputs" that for some reason was enabled. I turned it off, instantly fixed. I won't lie, I was pretty pissed the solution ended up being so simple, but that was overshadowed by the fact that it was SOLVED. I finally didn't have to deal with this crap anymore. From there I had to still do all the work for the project, upscaling the rawgens, rewriting all the dialogue, drawing the textboxes, everything. But it's done. I can rest.
I'll take my 5 upvotes and leave all of this behind me now. And save my damn projects.

Step 1 - txt2img (Leafan)

l3afan_pal, furry, pal_(species), panda girl, black-white fur, white pupils, red eyes, leaf belt, big arms, big legs, <lora:Leafan_Palworld_IL:1>,
solo, three-quarter portrait, side view, rear view, short stack, huge thighs, nervous, sweat, sweaty butt, open mouth,
masterpiece, best quality, no lineart, pyroxtra, <lora:PyroxtraStyle_Vpred-V2.3:0.8>, kekitopu, (frumples:0.9), sssonic2, <lora:illustrious_flat_color_v2:0.5>,
simple background,
Negative prompt: worst quality, looking back,
Steps: 28, Sampler: Euler a, Schedule type: SGM Uniform, CFG scale: 5, Seed: 3691777559, Size: 832x1344, Model hash: 63098ee7ee, Model: StableMondAI-SDG, RNG: CPU, Style Selector Enabled: True, Style Selector Randomize: False, Style Selector Style: base, Lora hashes: "Leafan_Palworld_IL: 0d7f73283f3b, PyroxtraStyle_Vpred-V2.3: 9474a6c332d4, illustrious_flat_color_v2: 75ef486a9f3f", Rescale CFG: 0.7, Noise Schedule: Zero Terminal SNR, Version: neo, Module 1: sdxl_vae

Step 2 - txt2img (Sehkmet)
Sekhmet_p4l, Egyptian, pal_(species), anthro, feline, bob cut, gold collar, red eyes, black dress, navel cutout, <lora:Sekhmet_Palworld_IL:1>
solo, three-quarter portrait, side view, front view, open mouth, neutral expression, bored,
masterpiece, best quality, no lineart, pyroxtra, <lora:PyroxtraStyle_Vpred-V2.3:0.8>, kekitopu, (frumples:0.9), sssonic2, <lora:illustrious_flat_color_v2:0.5>,
simple background,
Negative prompt: worst quality, looking back,
Steps: 28, Sampler: Euler a, Schedule type: SGM Uniform, CFG scale: 5, Seed: 3653470466, Size: 832x1344, Model hash: 63098ee7ee, Model: StableMondAI-SDG, RNG: CPU, Style Selector Enabled: True, Style Selector Randomize: False, Style Selector Style: base, Lora hashes: "Sekhmet_Palworld_IL: ff3a92acfef2, PyroxtraStyle_Vpred-V2.3: 9474a6c332d4, illustrious_flat_color_v2: 75ef486a9f3f", Rescale CFG: 0.7, Noise Schedule: Zero Terminal SNR, Version: neo, Module 1: sdxl_vae

Step 4 - Manual edits (Sekhmet)
*Cleaned up artifacting in eyes, mouth

Step 5 - Manual edits
*Took a screenshot in palworld and loaded it into a new gimp project as the background

Step 6 - img2img (Leafan)
same prompt, 0.2 denoise

Step 7 - img2img (Sekhmet)
same prompt, 0.18 denoise

Step 8 - Manual edits
*Removed background on both characters and created transparent copies, then added them to gimp project
*Added text