A single Lora to allow prompting of mini giantess size people that appear 6-30+ feet tall. The goal is to teach SD to tweak the scale of the background and provide full body poses.
V3 LyCoris:
Try 0.3-0.85 Strength, DPM++ 2M Karrras
What's New:
· Re-trained as a LyCoris, it's WAY BETTER. Less chaotic, more flexible, preserves model style.
· Expanded training data. Expanded on v2 training data from 278, 768x768 images to 896 images up to 768x1360. Added "macrogts" training images. Also see version notes.
Re-tagged trigger words to better prompt approximate scaling using mg-sg guide:
tallgts: Taller than average
minigiantess: Giantesses 7' to 12' tall
mesogts: Giantesses 12'to 50' tall
macrogts: Giantesses 50' to 500' tall
Also re-tagged training data by subject:
solo: one subject
couple: two subjects
group: one minigts subject and a group of people
V2 Lora:
Try 0.2-0.85 Strength, lower is better for realism.
What's New:
· Re-trained as a Lora.
· Upscaled training images to 768x768
V1.2 - Older: Hypernetwork
A hyper-network to allow prompting of mini giantess size people that appear 6-30+ feet tall. The goal is to teach SD to tweak the scale of the background and provide full body poses.
See version list to pick anime or realism styles. Unfortunately LoRa testing trained poses but not the sense of scale. Instead, v1.2 attempts to provide a more stable, adaptable hyper-network. Try ~0.7 strength for realism, or ~0.3 for Grapefruit.
Although no prompt keyword is needed, try "minigiantess" and specifying a pose: standing, sitting, kneeling, crouching, full body portrait, long shot, head touching ceiling, laying, etc. Best results are to turn off hyper-network for final image cleanup.
v1.2 - Realism
Trained on Delibratev1.1 in SD 2.1. Same training data, but works significantly better in SD 2.1 and still functions in SD 1.5.6.
v1.2 - Grapefruit
Trained on Grapefruit V4 in SD 2.1 to provide anime illustration styles. See version notes.
v1.1 and Older
Doesn't work well on SD 2.1. Tested on SD 1.5.6 on Automatic1111.
Try using 0.6 strength.
Other process tips: 30 steps/DDIM to generate poses, then move to img2img and reduce hyper-network strength ~0.3.
Trained on Delibratev1.1 in SD 1.5.6. Hypernetwork potentially works to a lesser extent on other models.
Training notes:
· v1 - 100,000 steps, 1 epoch, wide net. ~100 images focused on a single subject indoors with BLIP tagging.
· v1.1 - 100,000 steps, 1 epoch, wide deep and bottle-necked, ~140 images, some with two subjects and outdoors backgrounds.
· v1.2 Grapefruit - 20,000 steps, 4 epochs, wide deep and bottle-necked with scaled LR. Same training images as v1.1.
· v1.2 Realism - 20,000 steps, 4 epochs, wide deep and bottle-necked with scaled LR. Same training images as v1.1.
Description
~140 images, then mirrored. Pruned the original image set to higher quality images, added new images with two subjects and outdoors backgrounds. Also manually hand tagged all images after BLIP tagging. In-painting a second subject may still be easier.
FAQ
Comments (13)
how do i use this? i am very interested but cant seem to figure it out
In Automatic1111 web-ui, under settings, select the Hypernetwork and adjust it's strength. It's then applied to the model you're using. Or are you having issues prompting?
@periledascent676 think it just struggles with the model im using (anything v4.5) as i get a ton of very weird images, it seems like its trying i guess but i might be doing it wrong as im used to easier stuff like lora xd. I hope you can refine it as much as possible as its hard to make good minigts indoors stuff rn
@Mi739 Yeah, makes sense. I'll try to train one on that model and see how it goes.
@periledascent676 would be nice to be able to generate women outgrowing rooms in that style, cant really get that rn
@themangamer123 I trained a version on Anythingv4.5 but it didn't seem to work so well. But, v1.2 - Grapefruit was trained on Grapefruit model (I just picked to most downloaded Anime model), and prompts well. Try it out!
@periledascent676 way better than before even on anything v4.5, not bad, im actually getting some results
@periledascent676 is 1.2 the final version or is this still gonna get improved?
@themangamer123 Probably, yes. I think both better training (upwards of 72-100+ GPU hours) and improving the training data quality could help prompting and reduce weird poses. I'm kind of also looking at a version specifically for better size difference between couples.
I'm seeing strange results with weights anything larger then 0.4 on the lora version which doesn't have a big impact before it starts making really abstract bodies. Your pictures are great so I'm keen, I could post results but I would have to make a review. Using deliberate and the same prompts as you.
Thanks for the feedback, is this on the v1.1? It definitely generates some mutations, I'm working on refining the training, I'm still pretty new at it. SD already kind of struggles with full body portraits and large proportions.
I'm having pretty good results in v1.1 @0.6strength using txt2img DPM++ 2M Karras/150+ steps with ~50-60% poses being pretty normal/usable. Feel free to review and share images, can always edit later as you play around with it.
@jerrycurly I've updated the model, try v1.2. It's a lot more stable, and if you're using SD 2.1 you probably weren't getting great results.