Akira Mizutani from visual novel <a target="_blank" rel="ugc" href="https://growth-academy.itch.io/growth-academy">Growth Academy.</a> <a target="_blank" rel="ugc" href="https://tensor.art/models/715555622097018533?source_id=nj-zo13ml0C0ovcubnf99BIl">Tensor.art</a> | <a target="_blank" rel="ugc" href="https://yodayo.com/models/9c0d6cd3-8538-4d20-8767-9fcd1b54d36d/">Yodayo</a> (<a target="_blank" rel="ugc" href="https://yodayo.com/models/2251cdba-2044-4f55-98ca-f3827f0abe6e/">yodayo PDXL</a>) | <a target="_blank" rel="ugc" href="https://pixai.art/model/1731579570114072224?utm_source=copy_web">Pixai.art</a> (without XL v0.97 forever) | <a target="_blank" rel="ugc" href="https://www.seaart.ai/models/detail/f9a9a8e950fd7199f42be5d49352c791">Seaart.ai</a> Please read the description before choosing a version.v0.99 FLUXFirst version on flux.Training this lora was on the civit ai site, so I'll be glad to get buzz to improve the quality of the lora and be able to experiment.For original style, you can try the tags: cartoon style, anime screencapor just increase weightRecommended weight: 0.9-1.2 (1)CFG scale: 3.5-5 (4)Minimum recommended steps: 20Trained on FLUX DEV, natural language. Detailed settings are specified in the version. v0.97 PDXL Pony XL release, for attempts at the original style score_7 is recommended, try this prompt for it: score_7, sharp, anime screencap, flat colors, flat shadowsFor a completely original style, you can try this prompt: anime screencap, flat colors, flat shadows, sharp (without score at all) Otherwise many of the minuses of the sd 1.5 versions were fixed simply by switching to a different model. Didn't notice any particular problems, except theoretically slightly less flexibility. Minimum recommended steps: 28 Recommended weight: 0.7-1.1 (0.85-1) The image is recognized almost always, and I haven't noticed any problems when increasing the weight, so you can use 0.9-1 by default Trained on Pony diffusion v6 checkpoint.Short pros and cons of each version. v0.9 org The first version of lora, however, has its advantages. + • A most original style, this applies to the face and anatomy as well. - • Theoretically less flexibility. • Overcooked. v0.90 org Improved version v0.9 with an attempt to retain the original style. For the original char style, I recommend this version. + • Greater flexibility compared to v0.9. • Dataset has better quality data. - • The original style is quite a bit worse, it's different. • Theoretically less flexibility than v0.96 • Slightly overcooked. v0.96 + • Best flexibility. • Enlarged and improved dataset. • The least amount of overcooking. - • The original style has been lost and changed. Further details For all ver: Trigger word: akiraga Recommend words: 1girl,solo,muscular,ponytail,brown hair,brown eyes for original clothing red shirt, pleated blue skirt Recommended weight: 0.3-0.5 Minimum recommended steps: 25 Trained on aamAnyloraAnimeMix checkpoint. Does not work well with checkpoint nyanMix Recommend checkpoint in arts. Recommended to use adetailer, hires fix, all the basic outfits are present, but you have to pick up the prompt on your own. Possible problems: There are 2 lines from the pelvis to the torso, often turning into panty straps. Without specifying specific poses, it will be very similar. The back muscles can be a little strange. The clarity of the fingers may be lost. v0.9/v0.90 Weight for original style: 0.55-0.7 Recommended checkpoints for the org char style: aamAnyloraAnimeMixAnime_v1, coffeensfw_v10 Possible problems: Due to overcooked the finger problems are most pronounced, the flexibility of poses without a promt is limited, the original character has double eyebrows, but sometimes they can triple up. Org rec steps: 30+ v0.96 The most stable and less overcooking sd1.5 version, all problems described above except torso lines are reduced for it. I recommend it if the character image is enough. The style is more 2.5d compared to previous versions. Because of this, the increased weight will not result in a sustainable original style. Would love to see your generations! All versions are actually pretty stable.

Akira Mizutani Growth Academy (GA) [FLUX, Pony XL, SD 1.5]

First version, stable i think.The training was on civit, the site decided to self-delete the data after publication, so the information below is based on recollection, as far as I know forge allows you to view information of lora.image count:170Steps: ~2000trainBatchSize: 4unetLR: 0.0004ClipSkip: 1NetworkDim: 4NetworkAlpha: 2resolution: 512lrScheduler: cosine_with_restartsminSnrGamma: 5noiseOffset: 0enableBucket: trueoptimizerType: AdamW8BitmaxTrainEpochs: 10

Test flux character AkiraGA_epoch_9.safetensors

v0.99 FLUX

v0.97 PDXL

v0.96

v0.90

Akira Mizutani from visual novel Growth Academy. <a target="_blank" rel="noopener noreferrer nofollow" href="https://civitai.com/models/378135">Civitai</a> | Yodayo <a target="_blank" rel="noopener noreferrer nofollow" href="https://yodayo.com/models/9c0d6cd3-8538-4d20-8767-9fcd1b54d36d/">https://yodayo.com/models/9c0d6cd3-8538-4d20-8767-9fcd1b54d36d/</a> | Pixai.art <a target="_blank" rel="noopener noreferrer nofollow" href="https://pixai.art/model/1731579570114072224?utm_source=copy_web">https://pixai.art/model/1731579570114072224?utm_source=copy_web</a> <a target="_blank" rel="noopener noreferrer nofollow" href="https://pixai.art/model/1731579570114072224?utm_source=copy_web">Please re</a>ad the description before choosing a version. Short pros and cons of each version. v0.9 org The first version of lora, however, has its advantages. + • A most original style, this applies to the face and anatomy as well. - • Theoretically less flexibility. • Overcooked. v0.90 org Improved version v0.9 with an attempt to retain the original style. For the original char style, I recommend this version. + • Greater flexibility compared to v0.9. • Dataset has better quality data. - • The original style is quite a bit worse, it's different. • Theoretically less flexibility than v0.96 • Slightly overcooked. v0.96 + • Best flexibility. • Enlarged and improved dataset. • The least amount of overcooking. - • The original style has been lost and changed. Further details For all ver: Trigger word: akiragaRecommend words: 1girl,solo,muscular,ponytail,brown hair,brown eyes for original clothing red shirt, pleated blue skirt Recommended weight: 0.3-0.5 Minimum recommended steps: 25 Trained on aamAnyloraAnimeMix checkpoint. Does not work well with checkpoint nyanMix Recommend checkpoint in arts. Recommended to use adetailer, hires fix, all the basic outfits are present, but you have to pick up the prompt on your own. Possible problems: There are 2 lines from the pelvis to the torso, often turning into panty straps. Without specifying specific poses, it will be very similar. The back muscles can be a little strange. The clarity of the fingers may be lost. v0.9/v0.90 Weight for original style: 0.55-0.7 Recommended checkpoints for the org char style: aamAnyloraAnimeMixAnime_v1, coffeensfw_v10 Possible problems: Due to overcooked the finger problems are most pronounced, the flexibility of poses without a promt is limited, the original character has double eyebrows, but sometimes they can triple up. Org rec steps: 30+ v0.96 The most stable and less overcooking version, all problems described above except torso lines are reduced for it. I recommend it if the character image is enough. The style is more 2.5d compared to previous versions. Because of this, the increased weight will not result in a sustainable original style. Would love to see your generations! It took me a lot longer than I had planned, overall all versions are actually pretty stable.

Akira Mizutani Growth Academy (GA)

v0.99 FLUX

First version on flux.

For original style, you can try the tags: cartoon style, anime screencap

or just increase weight

Recommended weight: 0.9-1.2 (1)

CFG scale: 3.5-5 (4)

Minimum recommended steps: 20

Trained on FLUX DEV, natural language. Detailed settings are specified in the version.
First version, stable i think.

The training was on civit, the site decided to self-delete the data after publication, so the information below is based on recollection, as far as I know forge allows you to view information of lora.

image count:170

Steps: ~2000

trainBatchSize: 4

unetLR: 0.0004

ClipSkip: 1

NetworkDim: 4

NetworkAlpha: 2

resolution: 512

lrScheduler: cosine_with_restarts

minSnrGamma: 5

noiseOffset: 0

enableBucket: true

optimizerType: AdamW8Bit

maxTrainEpochs: 10

Files with SHA256: 011b628890f81d34ce1545bcc2223eafef0f0f3ba69c7955e8c8e741fd19ab47