Character LoRA (v1) for the Anima model family — dual-purpose SFW + NSFW. Trained on 212 curated solo images with danbooru-tag captions (10 epochs, 3310 steps).
Trigger: hakurei reimu
Identity tags (usually appear on their own): hair tubes, long hair, brown hair, frilled hair tubes, red eyes, ascot, brown eyes
Canon outfit (add these to reproduce it): bow, hair bow, red bow, frills, detached sleeves, frilled bow, skirt, red skirt, ribbon trim, ribbon-trimmed sleeves
Recommended weight: 0.7 to 0.9 (0.8 default).
Trained on anima-base-v1.0 (Anima family).
Description
FAQ
Comments (12)
A quick PSA for those in the know so you don't waste your time: this model is a textbook mess - 212 images (shown ~30 times each), rank 32 across ALL attention and MLP layers (for a character, mind you!), constant 1e-4 LR, no warmup, no dropout, no norm constraints. You don't even need to check the actual tensor norms - it's immediately obvious this thing modifies the ENTIRE model way beyond character binding. It's aggressively memorizing data instead of preserving the Anima base. Hard pass.
People like my stuff and I offer them out for free.
Your welcome to make your own and release it as well.
Beyond that genuinely open to the idea that a lower rank or some min-SNR/dropout would be tidier and I may test it but "obviously rewrites the whole model, hard pass" doesn't survive contact with the norms.
Happy to share the per-layer breakdown.
@fluxis My actual point is that rank 32 across all 280 attention/MLP modules + constant 1e-4 for 3310 steps + roughly 30 presentations per image is a very broad and aggressive setup for a character LoRA. I don't know if you've tested it or not, but with "moderately aggressive" style training settings, there's a very high chance that the style will be stable long before 1000 steps and will NOT kill prompt adherence. Raw LoRA norms don’t prove that the base is preserved. Without comparing them to the corresponding base-weight norms and actual layer activations, they don’t say much about behavioural drift. Same-seed comparisons on unrelated characters, alternate styles, complex prompts and earlier checkpoints would be far more useful than a per-layer norm dump.
If those tests show that the LoRA keeps the base style and prompt adherence, fair enough, criticism withdrawn. But "people like it and it’s free" doesn’t demonstrate that. I think it's no secret to anyone by now that local users happily consume garbage models and garbage content, while the producers of all this get various medals like "Top X Lora Creator" and so on. isnt that sad?
@FFTDenjoyer As said I am open to looking into it .. I have tested and will continue to try what you asked but you also don't need to be a troll/jerk.
If you have constructive criticism share it if people are willing to listen that's great but there's no need to be so rude or aggressive.
@fluxis My words may very well have been aggressive, but that doesn't mean they weren't constructive. I outlined all the main problems twice (above)
@fluxis And honestly, I didn't mean to offend you with my "aggressive criticism" or anything like that - it just frustrates me to see open source stagnating, as I mentioned above. If you want to get better at training characters or styles, you can DM me - I'll give you simple advice that works in the vast majority of cases. gl!
@FFTDenjoyer I ran the tests you asked for again with this model .. LoRA on/off for unrelated character same seed gets me the same result.
No impact to base model.
Whether the extra steps (1000 vs 3200) is actually helpful I haven't found much evidence of but its not harming prompt adherence or the base model through my tests on same-seed comparisons.
I can post them if you would like.
@fluxis I recommend you try, first and foremost, generations with two characters performing different actions. Since simple scenarios can indeed be implemented without any issues on an overfitted model - that's a big plus for Anima
@FFTDenjoyer I just ran this test as well with the same clean exact match with lora on/off on the base model with the same seed.
I will still lower the epochs because it seems I can save myself some training time for a similar result but I have found no evidence of your claims through these tests in terms of it being harmful.
@fluxis Have you tried your char LoRA with any style LoRA? Or just with artist tags?
What baffles me even more so is the choice to pick the single most represented character in the whole model the knowledge of Hakurei Reimu in anima model is absolutely top tier, she sits supreme over ALL the other tags, even including objects - the only thing that is actually above her is "1girl".
She is the single most well-known character in the entire anime-related universe.
@yorgash800 I'm afraid you'll never get an answer to the question "why do people train well-known characters for the model" :) It's a classic