RDBT [Anima]
This page contains the "original LoRA files" for advanced users.
See the main page for more info. (ckpt version, usage, faq, int8 ckpt...)
See this page for update log and version info.
Sharing merges using this model is not allowed. This "restriction" won't affect anyone. It's only aimed at those who steal others' models to sell.
Description
FAQ
Comments (15)
v0.6d is cfg distilled directly from the "anima preview". No finetuning. You can use it on any checkpoint based on "preview".
E.g. Cover images are using Animeyume v0.1.
"Use CFG scale = 1. Prefer euler a and euler." Well, if you're at CFG1 with a distilled LoRA you'd better use LCM, that's what is was made for. 🙌
No, LCM sampling was made for the LCM paper released 2023. Modern consistency methods are completely different. Some may work with LCM sampling but in general it's not the intended use.
Works insanely fast and allows me to even drop steps and still have a good image, reducing my gen time from 25 seconds down to 9.5 on an RTX 3090.
Unforunately it does completely screw over artstyles and some just outright don't work because of this LoRA.
But still a really great LoRA for when you want to experiment around with ideas.
Also it does kinda seem like dpmpp_2m_sde_gpu and uni_pc performs better than euler / euler a. But I haven't tested all that much so can't say for sure.
With NegPip you can use negatives even if you have set cfg to 1. https://github.com/pamparamm/ComfyUI-ppm In addition, EasyCache can be used to halve the sampling time.
how do you use easycache ,at least with the 15 steps euler-ancestral-cfg sampler I am using it doesn't seem to trigger at all.. does it need 30 + steps ?
@xpnrt i put easycache after model/load lora, try another sampler: euler a cfg++ It seems to take a long time. If i remind correctly when using easycache i get those sampling time (30 steps):
20 sec for euler a cfg++
10 sec for euler cfg++
5 sec for euler
Also i set easycache start_percent at 0 and end_percent at 100, reuse_threshold i keep at 20, increase it to get a speed boost but at the cost of quality
@Porzo ok that seems to work , what about the negpip implementation, if do this for example : "masterpiece, best quality, score_9, score_8, very aesthetic,hair bun, brown eyes, 1girl, (worst quality,:-1.3) (low quality,:-1.3) (score_1,:-1.3) (score_2,:-1.3) (score_3,:-1.3) (blurry,:-1.3) (sepia,:-1.3) (extra hands,:-1.3) (out of frame,:-1.3) (bad anatomy,:-1.3) (jpeg artifacts,:-1.3) (signature,:-1.3) (extra fingers,:-1.3) (missing fingers,:-1.3) (fewer digits:-1.3), (ye-pop,:-1.3) (deviantart,:-1.3)" this is on the positive prompt , negative empty or something on it doesn't seem to have any effect, the resulting image is just noise. this is with this lora connected, cfg set to 1, and the "CLIP NegPip" connected after model and clip so the rest of the loras , prompts are connected to it just as the flux example workflow....
@xpnrt i have posted a workflow https://civitai.com/articles/26248/fast-wdtagger-image-generation-workflow
So, not for lower steps too? Still need high steps?
Version 0.2d was heavily biasing towards 3D, photographic and "realistic-y" gens,
going to try out the new version and see what it makes.
You need to specify style in prompt. Prefer artist names, or "anime style", etc.
Thanks for the updated lora. What's the best method for training loras in Anima? Do you use the anima kohya branch? Is it important to use the danbooru tags for the text files? Thanks
I can't tell which tool is better or easier. I use my own script.
_ Is it important to use the danbooru tags for the text files?
Different people have different feeling and answer. My answer is: for small dataset < 100, use natural language + meaningful trigger words. Because nowadays text encoders are llm, they have their own thoughts.
kohya sd-script has anima already merged in main branch.
This worked well for a simple character lora on 6 images: Adamw8bit, 2e-4 LR, 16 Batch, 800 steps divided by batch = 50 total steps, cosine, 20 regularization images, Booru tags at top + Natural language at bottom generated by Kimi K2.5 (Used a script here), use keep_tokens_separator= "|||" on the captions for shuffled tags and to not shuffle natural language, caption_tag_dropout_rate = 0.05
You can train for 2K steps and reduce LR to 1e-4 if you want (default for sdxl), but I'm satisfied by the lora quality already despite being only 800 steps for Anima.