Updated: Jul 2, 2026
base model
Stardate 2026.07.01; Captain's Log: I'm finding some fidelity issues with background fuzzing and some foreground burning when using HRF at 1.5x (from 1024 x 1536) as this model seems to start tapping out somewhere around 2 Megapixels. After some testing, the sweet spot for HRF on a 2 x 3 image is 1.375x (2.0625 Mpxl = 1408 x 2112). The priests of the Temple of Syrinx have assumed control.
The High-Res Boost LoRa in Suggested Resources can also help maintain image sharpness when running HRF, but I recommend only using it during HRF runs as it may impact text-2-image results.
I went back and reviewed my settings for the model and after a lot of subjective comparison, I'm split between two recommendations:
dpm++_3m_sde_gpu + sgm_uniform @ 3.0 CFG & 2.718282 Flow Shift
er_sde + sgm_uniform @ 6.0 CFG & 2.718282 Flow Shift
On average, I got the best results scheduling with sgm_uniform with simple being a good fallback scheduler. Using dpm++_3m_sde without the gpu option produces a significantly different result so if sde_gpu is not available to you as a sampler, fall back to er_sde. The standard flow shift value is 3.0 and, maybe it's just seems kismet, but I enjoyed the results I got with Euler's number. I'm also a nerd who loves irrational numbers.
I got bored last night and decided to see if I could merge an Anima model for shits and giggles. I just vibed it out using some models that I really like, and it kind of hit my sweet spot for clarity, prompt adherence, lora compatibility, and overall versatility.
Thanks to the creators that I mooched this merge from:
I generate most everything at 1024 x 1536 using 3-5 CFG at 24 steps. I don't personally use turbo because it takes away from the model's versatility, but it works just fine all the same. All the showcase images are as-generated without inpainting or upscaling. Though I did iterate a bunch to get the some of the results I wanted.
For sampler and scheduler, I recommend using whatever looks good to you. I prefer using deis_2m_ode and bong_tangent, but er_sde or euler with whatever should work fine as well.
As far as positive and negative prompt quality tokens, I again recommend using whatever looks good to you. I'm happy with the results without them and have learned that less is more with Anima. So I just use basic modifiers like 'realistic', 'flat shading', 'toon style', etc to shape the output.
I consistently hi-res fix my stuff at 1.5x 1.375x upscaling using RealESRGAN_x2plus and lanczos downscaling even though it doesn't always need it. I just enjoy spending my time on things.
Anyway, that's all there is to it. If I feel cute, I might update it down the road. Enjoy!