Stuffing your belly into your jeans only works until you run out of belt loops and start overflowing them... although she's not going to admit she doesn't fit in her favorite getup just yet.

Recently discovered a fun new model called ChromaXL that's probably going to be a huge gamechanger; it's buggy as hell and can't upscale or inpaint well because of how overcooked it is, and it seems to fall into certain styles no matter what I do, BUT it knows every pokemon and species I throw at it and somehow infers what I want even though it can't understand natural language. Testing shows that ALL LoRAs work noticeably better with it than anything else, including the new symix immobility ones.

Honestly I hate this model a lot and it hates humanity, once the devs fix the "haze" that shows up in blank spaces and the weird twitchiness it'll be great, bonus points if they teach it natural language. My quality standards are out the window here since this is so unpredictable I have to hope I get the same image style across 2 pics.

For those curious, here's how to make the model actually work:

• Artist tags aren't super necessary, but the Spud version recognizes most of them and works well enough. Mango has less anime mixed in and recognizes furry styles better, but it tends to pixilate things more often. The regular v3.31 is unstable and will mess up inconsistently, not worth downloading.

• Since you don't need artist tags, using mediums like "matte painting" do the job, but you'll need to include "masterpiece, best quality, detailed image" garbage because for god knows what reason it actually works for once.

• Only e621 tags work in the prompt, and you can't use sentences or phrases.

• You have to use Euler A. I've seen DPM++ 2M work too, sometimes with Karras, but Euler is the only one that works reliably. If you have the latest version of Forge and can easily set your scheduler, set it to beta. Use 14-16 steps.

• Set your CFG at 1.75-2.5, or 2 if in Krita AI diffusion.

• Width or height need to be at least 1024, higher is better. 1228 is a good base, this model struggles a lot more for some reason.

• YOU MUST UPSCALE FOR FULL QUALITY, RAW GENS ALMOST ALWAYS LOOK LIKE SHIT.

• When upscaling, you'll need to set the hi-res steps to around 12 as the model is way too powerful and cooks itself if you run it for too many steps. Denoising of around 0.6 has worked for me. Inpainting works as usual but you might want to lower your step count to 8 and fiddle with the denoising and try upscaling by a small amount.

• I used the latest SymixPony v8 since v6 and v7 use the DoRA system, which won't load in most webuis right now (it was fine last month, don't update if the older ones work for you). Theoretically the other versions will work better in the future but for now we wait for bugfixes.

• It's an XL model, so you'll have to search for "emphasis" in the settings and set it to "no norm" in Forge.

Unedited original and workflow here.

Proompt:

solo, anthro, (obese), mature female, (aggron:0.9), blue eyes, slit pupils, black sclera, yellow shirt, (mom jeans, high waisted bottomwear:1.3), denim, standing
BREAK
(stocky), fat arms, big breasts, butt, huge thighs,
three-quarter portrait, side view, detailed background,
BREAK
[odisia, cettus : bellylliumu, erumeruta, radiowave, cosith : 0.4],
dark fantasy, unreal engine 5, raytracing, best quality

Negative prompt:

abs, navel, slim, belly overhang, (fat rolls), ribs, scales, white belly, bent over, crop top, bubble butt,
source filmmaker, flat color, patreon, watermark, worst quality

Steps: 14, Sampler: Euler a, CFG scale: 2.5, Seed: 1889583421, Size: 880x1024,
Model: chromaxlMix_v331Spud,
Denoising strength: 0.55, Hires upscale: 1.4, Hires steps: 12, Hires upscaler: Latent,
Clip skip: 2, Emphasis: No norm