My take on an XL Pony, intended to veer more towards realism by default but without losing any knowledge of specific characters or concepts from the base model. Does not contain any third-party data from non-Pony XL checkpoints or Loras. Constructed solely from a series of bucketed 1280px Florence-2 Large / WD-VIT-3 captioned Loras that I trained myself directly on Base Pony, and iteratively merged at various strengths. All-in-all, version 1.0 has ~6500 images worth of new data.
Recommended basic positive prompt:
score_9, score_8_up, score_7_up, optional_rating_whatever, optional_source_whatever, your prompt here
Recommended basic negative prompt:
score_3_up, score_4_up, score_5_up, sketch, greyscale, monochrome, (simple background:1.3)
Recommended sampler / CFG / steps settings:
Euler Ancestral with Normal scheduling at 7.0 CFG and 25 - 35 steps is always a good starting spot. DPM++ SDE GPU and DPM++ 3M SDE GPU can also give nice results (at lower CFG, around 4.0 - 5.0, Exponential scheduling for 3M, Normal scheduling for uh, normal) if you're really aiming for realism.
If you use an up-to-date installation of ComfyUI, the "Clip Set Last Layer" node is NOT necessary with XL models like this one. Just do not use it.
Description
Initial release. Overall look of the model is basically where I want it (as is the re-introduction of concepts that Pony had forgotten about, like most types of cars lol), but there's still work to be done for future releases certainly.
VAE is baked in (SharpSpectrumXL, specifically. Meaning your images will look somewhat duller than intended if you load the regular SDXL VAE with this model for some reason. So don't do that.)
FAQ
Comments (11)
For the showcase images, note that the Tifa and car ones are actually Euler Normal via the "Restart" KSampler comfy node. CivitAI sadly doesn't have support for recognizing that so I thought I'd point it out here.
Do you happen to have an image with your restart sampler configuration embedded? I've tried it multiple times, but found it to be nothing but confusing and random, while sometimes taking an absurd amount of steps. Like 30 turning to 279 and fun like that :D
Would love to figure out a functioning configuration to understand its purpose
@redpinkretro The Tifa one I mentioned was actually stuck in analyzing hell lol, it should be visible now, I didn't notice before. You should be able to grab the comfy workflow directly from it. Or from the car one.
actually wait nevermind Civit just tosses the metadata when it doesn't recognize ONE node I guess, uh, here's a catbox for the Tifa pic:
@diffusionfanatic1173 Yeah, I actually built a workflow specifically for manually adding all the metadata of used LoRA and such as my main workflow ALWAYS has weird nodes and group nodes etc. that wouldn't be transferable :D
Thank you for the workflow! So you don't even need to mess around with the cryptic sampler/scheduler segment stuff to get better results?
@redpinkretro no, I find the a1111 parameter to do the trick.
@diffusionfanatic1173 I tried it with a1111 settings right next to a regular ksampler, same settings same everything, the results of the restart sampler looked a little different, but not better, rather slightly less contrasted and detailed. Do you use specific sampler/scheduler combinations to get an improvement out of it?
@redpinkretro Restart doesn't really work with anything other than Euler Normal, and Heun I think. Like just Euler, not the Ancestral version you'd more likely use with a regular KSampler. It's probably pretty prompt-dependent also, I don't exactly use Restart like that frequently anyways.
@diffusionfanatic1173 Thank you for clarifying. Still I can't get an advantage from using it. It takes about 50% more steps (20-29), 50% more time and doesn't even look better necessarily :D
I'm giving up on it for now. Thank you for your help!
One other thing: as with all my models, any images you see me post in the gallery will always be direct generations with no detailing or upscaling, with as much metadata included for recreation as is possible.
Training up V2.0 now. Tries to introduce some stuff that neither Pony nor any regular XL model were ever very good at, e.g. people fighting
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.