Showcase images are base model (left) vs LoRA (right).
This LoRA does two things:
Allows generating at higher resolutions. 1536 works without any major issues, and even 2048 (4 MP) now works without completely falling apart.
Slight aesthetic increase toward higher-quality images.
It is trained on 10,000 high rated Danbooru images, with mixed-res training at [512, 1024, 1536] resolutions. The images are sampled to have a more even balance of SFW/NSFW and male/female (because otherwise, selecting on ratings only, you would get almost 100% NSFW female). The aesthetic boost is fairly weak, and is prompt dependent. It is not targeting any particular aesthetic, but simply "high rated Danbooru images", so there is still a lot of stylistic diversity, and hence the LoRA's effect is subtle.
Description
Trained on preview3.
FAQ
Comments (18)
Did a quick test with a Preview3 trained character Lora and stacking the highres lora seems to work fine. It helps A LOT with backgrounds from my check on x1536 base resolution.
Keeps style of character Lora fully intact.
looks like WAIslop ngl
Touchdown-R_MIX_31based.safetensors
If the same training step target is set,
will training
sequentially 10 epochs of [512] > [1024] > [1536]
be more beneficial for high-resolution details than 10 epochs of [512, 1024, 1536]?
I know the latter has already been implemented in the script, while the former requires manual operation in a certain sense.
Damn these loras are fire!
"The images are sampled to have a more even balance of SFW/NSFW and male/female (because otherwise, selecting on ratings only, you would get almost 100% NSFW female)"
Thank you for addressing this, you guys are truly gods ❤️🎉
Add anima support for WaveSpeed. With my highres fix config, it can reduce highres fix sampling time by up to 70%.
Add it just before highres fix and after base image sampling.
object_to_patch: diffusion_model
residual_diff_threshold: 0.200
start: 0
end: 1
max_consecutive_cache_hits: 2 residual_diff_threshold_end: 0.300 threshold_schedule: linear num_always_run_blocks: 4
enable_diagnostics: false
https://github.com/WalkingMeatAxolotl/Comfy-WaveSpeed
广告哥，草草
good
It definitely helps at 1024x1536 and above with complex prompts. At the same time it seems to reduce amount of details, which is really counter intuitive. Here is an example of simple prompt that preview3 handles generally well even without this lora at 1024x1536 https://civitai.com/images/127489918
At lower resolutions it is even more noticeable.
That "colorfix" LoRa is terrible, doesn't really fix anything and mostly the culprit that reduces details.
@NanahiraFan huh? I have comparisons on page, tf are you mumbling about? Show yours or gtfo
@Volnovik Your comparison isn't worth much when you're using a LoRA for a "semi-real" style when all it does is replicate the shitty SD1.5 default AI look lmao.
People aren't obligated to give comparisons to disagree with you btw. And if you're gonna do a comparison, at least do it properly, remove every other LoRA so you know it's not a LoRA fighting with another LoRA. You're comparing this LoRA, not colorfix or Semi-R style.
@upscaleanon537 that's not how loras work. But you can do it yourself. There will be 0 difference, but effect would be less noticeable, since preview3 sucks in details in general
@Volnovik It literally is how LoRAs work, you have multiple LoRAs fighting over what the output should be like and it makes no sense to add irrelevant LoRAs to whatever you're comparing. And when you use SD1.5 slop ass style, it speaks volumes about bad you are lmao.
If preview 3 sucks with details, then you're just bad. I've seen plenty of gens that have amazing details. But your gens looks like they're stuck in 2023
@upscaleanon537 none of the loras were trained on resolution above 1024 so you are simply wrong. Regarding style, anyone can just take a look at your shared images to get the idea that there is no discussion here. Even if you take into account subjectivity
where is the realism lora
good improvements for higher res generations, and helps a lot with some subjects that were not fully covered by the preview 3 checkpoint yet, but it does seem to be a bit too sfw. the idea of a balance is not just good but greatly appreciated but you might want to increase the number of imaged used. Also... is gemma really the best at tagging? Personally I found it to be a bit too "safety conscious" for that kind of work. Then again, I think it might just be up to the mproj used and luck. Sometimes you get good tags, sometimes the vision model just goes over detail as if it's not there.
Oh and... depending on the prompt, that "little" detail you bring up is a LOT more impressive that you make it sound in the LORA description.
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.