V4
This version is based on Turbo V4. The layers for VAE and TE are still using the old AIO format that is incompatible with current version of Forge Neo, i'll see if i can make an alternative version.
V3.1 Patch
This version is leveraging Turbo V2.9 and should be able to handle better LoRAs from my tests 😊
V3 update
Using Turbo V3, far less qwen_heretic, no more p3-to-v1 backward move. It should be pretty close to a vanilla anima turbo, but with my own take on this aio checkpoint :D
V2 update
Same idea as V1, sligth change in content:
model = load_file("anima_baseV10.safetensors")
prefix_replace(model, "net.", "model.diffusion_model.")
te = load_file("qwen_3_06b_base.safetensors")
prefix_replace(te, "model.", "cond_stage_model.qwen3_06b.transformer.model.")
for k in te:
model[k] = te[k]
merge_lora(model, "qwen-heretic-lora.safetensors", 0.4)
merge_lora(model, "Turbo-ANIMA-v1.5.safetensors", 0.9)
merge_lora(model, "anima-P3-to-V1.safetensors", -0.1)
vae = load_file("finalcut_vae.safetensors")
prefix_replace(vae, "", "first_stage_model.")
for k in vae:
model[k] = vae[k]
save_file(model, "ANIMA-V1-Turbo-AIO_V2.safetensors")If reading this is too hard, the summary is:
less change in TE (0.8 before, now 0.4)
Turbo V1.5 instead of V1
FP32 VAE instead of BF16
V1
I had some great reviews regarding nArnima AIO, but some people would have prefered than i didn't bake in my own style to allow more flexibility... so here we go!
This is ANIMA Base V1 with the same treatment as nArnima AIO without the style part of the LoRA (only TE/Turbo/Adapter LoRA).
Usage:
Resolution: Base V1 had a huge boost in quality with a final finetune at 1536x1536 resolution = Showcases picture are 960x1440 without any Hires.Fix or Upscale.
sampler/scheduler: ER_SDE + Beta (or Smoothstep for sd.cpp)
CFG: 1 since the Turbo LoRA is included (but 1.1 to 1.5 also works and allow back some negative prompting)
steps: 8-12 should be enough with the Turbo LoRA but i go for 16.
positive prompt: i like to start with masterpiece, best quality, score_9, newest, detailed
flow shift: keep the default at 3 if your tool of choice display the value
Works with Forge Neo, should work with ComfyUI.
Quantized version for sd.cpp coming up on HF soon is here: https://huggingface.co/n-Arno/Anima-P3-Turbo-AIO-Q4_K/blob/main/Anima-V1-Turbo-AIO-Q4_K.gguf😁
PS: don't use my nArnima V2 LoRA on top of this model as-is or the TE will get funky (not completely broken, but it may not play nice). Override this model TE with base Qwen3-0.6b if you intend to use this very specific LoRA on top (but in this case, why not use the full model?), otherwise, for any other LoRA, no problem should arise 😅
Legal
The CircleStone Model is licensed by CircleStone Labs LLC under the CircleStone Non-Commercial License. Copyright CircleStone Labs LLC. IN NO EVENT SHALL CIRCLESTONE LABS LLC BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH USE OF THIS MODEL.
Description
Posted on Civitai, upload found elsewhere are FORBIDDEN and done without the author permission.
FAQ
Comments (20)
Running smoothly on ComfyUI. I’m genuinely amazed and very satisfied with the results. Thank you so much!
Thanks for confirming this! 🙏
pretty neat- using the exact same workflow i've been using to investigate styles and prompting with anima, only changing the 3-part loader for a checkpoint loader, and accounting for sampling settings, my gen speed is nearly cut in half, which i expected. but also it's removed some of the sketchy-quality and 'shabbiness' from this particular artist style i'm investigating, and made it look significantly cleaner, while still keeping the style recognizable! granted so far this is only a data set of 1 (aka an anecdote), but i'm pretty impressed. appreciate ya!
e: compared to base anima 1.0
That's nice! This "improvement" can come from several parts => the turbo is using a part of an original DMD2 LoRA, the text encoder should "refuse" less (aka, some combinaison of tokens could trigger an incorrect conditioning due to "censorship") and the adapter i used could have restored some artists that may have suffered during the last fine-tune if the source data was not available in higher resolution 🥰
so this is basically Anima base 1.0 with turbo lora?
Not exactly:
- I have added at 0.8 a LoRA to the text encoder (to limit refusal)
- The turbo LoRA used is my own custom merge
- An adapter has been used at low strength (0.1) to improve results from Preview 3 LoRA
lora to limit refusal? don't think text encoder is censored. Anima has nsfw by default. i will try it out
@MrFlex It is NSFW capable yes, but qwen3 models have a guardrail by default.
First time trying Anima and I went for this checkpoint using ComfyUI. Pretty awesome so far but having slight trouble generating a more mature look.
That's something i didn't try actually, since Anima use Danbooru as a dataset, it probably suffer from the same problem as Illustrious at the beginning until some people trained LoRA to get the mature feeling.
Absolutely, still it's awesome, listens to whatever i can come up with and generations are very clean.
try adding tags like “aged up” to the prompt, and adding “aged down” to the negative words. I think there will be a slight change.
even when the subject is “woman” or “milf,” the model sometimes ends up looking like a little kid, doesn't it? 😂
@imaginak I didn't even think of "aged up", guess i've been spoiled by all the different mature checkpoints. Thanks!
Can you upload your custom turbo LoRA?
It is already available here: https://civitai.red/models/2619830/turbo-for-anima-less-steps
I did not upload the Qwen3 LoRA but it is available here: https://huggingface.co/n-Arno/ANIMA_extract/blob/main/qwen-heretic-lora.safetensors
Maintains great stability at high res [1536x] when used with character Loras.
Most of the Anima 1.0 Loras I trained tend to fall apart at higher res on the base model [maybe cuz of 1024 train res] but this checkpoint handles high res extremely well.
Would be nice to have the option for only the safetensor tho [without TE/VAE].
I could do without the VAE as it was not altered, but the TE has been changed a bit. You could get the same result using this Turbo LoRA and the Qwen abliterated TE (even if my TE in there is slightly different) with base Anima if you wish 😊
(Thanks for the tips 💙)
@n_Arno Great thanks for the info
This damn model is amazing, it completely surprised me. A round of applause for its creator. Quality, few steps, resolution doesn't matter, zero bugs or glitches, time efficiency. Perfection.