--preview3
I'm also sharing a new experimental version of UltraReal FineTune Anima, this time trained on Anima_preview3.
This version was made because several people asked for a preview3-based release. In some cases it can produce better results than the preview1 version, especially depending on the prompt, but in other cases preview1 may still look better or behave more consistently.
So I don't really consider this a strict upgrade — it's more like an alternative version. Try both and use whichever one works better for your workflow and your prompts.
Model Features:
Based on Anima_preview3
Trained in the same way as preview1 release
Still highly prompt-sensitive
Some improvements in certain styles and generations
Some possible regressions compared to the preview1 version
Still Experimental / WIP
Special thanks to the Reddit donor who supported the project — your donation was one of the reasons I decided to retrain this for preview3 as well.
P.S.: in my flow i use custom sampler and scheduler, u can take it here https://github.com/WASasquatch/RES4SHO
--preview1
Hey everyone. I'm sharing my new experimental full finetune of the Anima_Preview1.
For this version, I collected a completely new dataset from scratch - it's entirely different from the one I used for my Flux.1 finetune.
Model Features:
🎛️ Highly Prompt-Sensitive: The stylistic range is quite diverse, but the final output relies heavily on your specific prompting.
📸 Analog & Digital Aesthetics: It can produce a wide variety of looks, from distinct analog grain to "high-quality" digital photos (well, as high-quality as it gets for a 1MP retro resolution).
⚡ Optimized: I've included Q8 and Q6_K_M quants for easier inference.
Honestly, I really love the image quality you can squeeze out of such a small model. However, this is still very much a WIP (Work in Progress).
I would love to hear your feedback and see your generations.
Also, NSFW capabilities weren't harmed
Description
FAQ
Comments (35)
Very good! I was expecting it losing character and booru tag knowledge like all those Illustrious realism merges but its really good!
Is this actually Anima Preview 1 based?
Yes. After some tests I found out that preview3 works slightly worse with realism, especially faces becomes distorted
This is really cool, probably one of the better if not the best anime-to-realism model. I tried it only briefly. For general subjects (non characters) it works very well (something like "photograph of a girl with fox ears and fox tail....etc". The backgrounds are great. The problem is for known characters/subjects. I tried both Ganyu from Genshin Impact and Ellen Joe from ZZZ and the faces are still very much 3D style/uncanny. It was a similar experience when i tried your loras: the NiceGirls excelled in faces/subjects but only at portraits, and the other one excelled in backgrounds but not subjects. I use Gemma4 for prompting and usually specify "a candid photograph of...". l'll have to test more.
I was testing realistic Anima setups when you suddenly released this fine tune! It's surprisingly competent for realism considering Anima is intended purely for 2D illustrations. I've gotten some pretty nice results pairing this with hinablue's R Tweaker Lora and sorta rdbt (both rolled back to their preview 1 versions).
The beta57 scheduler seems to help push realism and details but it kept ruining the faces for me so I stuck with sgm_uniform. I also preferred the lossy results from speed optimizations like EasyCache and Spectrum since they felt less polished and help break the shiny AI look.
Thanks for your work!
Edit: version 1.2 works a lot better
_____
For me, the images come out like SD 1.5. I'm using Forge Neo. Any ideas why that could be? I tried replicating some of the preview images' settings but they were came out totally different and with low quality. I used Q8
I have the same problem. All the images come out distorted.
Check parameters: 50 steps with ER SDE/Beta; CFG: 5, Shift: 6.
@mphobbit Thanks. is there a way to check for all the details in civitai? When I clicked on the images, I couldn't see Beta and Shift 6. I had Beta too when I tried but shift was at 3 or 4 I believe
I'll try it out soon, I'll give an update if it worked
Update: Images look 100x better but still not nearly as good as the preview images sadly
Update 2: If I try anything other than the preview prompts, it sometimes ignores 90% of what I typed, even if its something simple. On occasion it works but it never ever gets any character right for some reason.I have never had such issues before but perhaps just because Anime is new and all that. When I use the nornal Anima P3 it works pretty good tho..
I really want this to work because in the preview this model looks really damn good
@Lopodelas no way on Civit, only locally.
UPD: Civit generator recognizes the "Shift" parameter as "CFG scale" (instead value of CFG in Neo) .
@mphobbit any lightning lora or similar 4 step for this yet ?
@amazingbeauty Try: https://civitai.red/models/2560840?modelVersionId=2877687
@mphobbit thank you. anima is better or zit ?
@amazingbeauty they can't be compared directly. It depends what do you expect from model. ZIT and ZIB are much powerful (8B parameters + 4B LLM encoder vs. 3B parameters + 0.6B LLM encoder) . Anima is closer to Pony/Illustrious and its behavior seems to be better with tagging rather than NLP.
ZIT/ZIB look like a good and powerful general SDXL replacement. ZIT has a good photorealism out of box (but sometimes "flat" on my taste, but LORAs and finetunes solve this problem) + I see wider choice of finetunes for ZIT rather than for ZIB.
ZIB is more generalist model, better details (e.g. fur), but requiers >50 steps + NLP prompting to uncover its power. With <50 steps or turbo loras you'll recieve problems on a level of early SDXL (missing or extra limbs, various mutations etc).
@mphobbit seems as i expect, zit is more powerful but anima if tuned for specific thing it can do well maybe as zit level or maybe better! keep tuning it is the only test! , thanks for working on these finetuning..
@mphobbit Do you know if ZIB/ZIT has been trained on anime/game characters? That's my main focus, which is why I really wanted this Anima finetune to work
@Lopodelas Here are couple ideas:
https://civitai.red/models/2483351/z-anime
https://civitai.red/models/2520590/z-imageturbo-anime-v2-bf16
@mphobbit My bad, I actually meant for realistic images
@Lopodelas just look through the site setting filters to "Checkpoints / Z-Image Turbo" and "Checkpoints / Z-Image Base" and choose what do you like most.
Start from:
Sophos-ZIT by sophosympatheia
https://civitai.red/models/2506618/sophos-zit?modelVersionId=2885923
GonzaLomo-ZPop by GBRX (by the way, I can recommend you his SDXL checkpoints)
https://civitai.red/models/2192562/gonzalomo-zpop?modelVersionId=2598704
z-image-turbo_djm_realistic by DjM_
https://civitai.red/models/2308742/z-image-turbodjmrealistic?modelVersionId=2770039
Z-Image_clear_photoreal by easygoing0114
https://civitai.red/models/2369132/z-imageclearphotoreal?modelVersionId=2664542
unStable Revolution ZIB by Peli86 - I like its outputs
https://civitai.red/models/2402485/unstable-revolution-zib?modelVersionId=2701478
PonziReality (ZiB) by ponzi:
https://civitai.red/models/2382954/ponzireality-zib?modelVersionId=2691241
Warning: if you have plans to use LoRAs the best choice are vanilla ZiT/ZiB. Sometimes finetunes conflict with loras and give mess.
This is a great star for photorealism, love the details profesional and amateurish, I noticed that most of the time the skin looks very soft, poreless, some hand and face distortion when the subject is on half body in a certain distance , also,I noticed that has a hard time with light for example crack of light,dimly lighting etc, but overall an awesome start
Feels like a great start. I still seem to get better results with anima preview 3 and your LORAs though, at least as far as clarity and variability. But your finetine can pull off better adherence in many cases. Also, enabling Flow Shift to 1 can produce some nice results.
Thanx. I think you are right. Maybe I should to train once more, but with slightly another dataset
for my tests, the tags for each image were copy and modified from online sources, including pure tags and natural language plus tags. covers NSFW and normal images.
Each test set consists of four images: 1.default workflow, 2.default workflow + NGA, 3.dual CLIP ( using CLIP skip = 2), 4.dual CLIP + NGA.
personally feel the default workflow is already sufficient.
-------------
I think this is the best model for turning animated characters into realistic human forms with NSFW capability, far superior to Klein and ZIT.
Those two models require a lot of LoRAs to generate the desired anime characters, whereas this model has a much richer built-in knowledge of characters.
-------------
for normal images, the characters are actually very faithful to the original.
for more complex poses and more extreme/adventurous scenes (such as modifying the character's body proportions into something more extreme), the clothing becomes less accurate. (small issue, can all be resolved with LoRAs.)
this is amazing, i just spent last 5 hours playing with it, great prompt adherance, amazing amateur style, please go on for the love of god!
please write how to open and use this neural network. I have installed the fp16 SafeTensor file
Half precision, best balance • 3.89 GB what should I do next?
@df4348624153 you could use the raw huggingface diffusers library, or read an online tutorial on ComfyUI
https://github.com/Comfy-Org/ComfyUI#installing
From https://huggingface.co/circlestone-labs/Anima:
"The model is natively supported in ComfyUI. The above image contains a workflow; you can open it in ComfyUI or drag-and-drop to get the workflow. The model files go in their respective folders inside your model directory:
anima-preview3-base.safetensors goes in ComfyUI/models/diffusion_models
qwen_3_06b_base.safetensors goes in ComfyUI/models/text_encoders
qwen_image_vae.safetensors goes in ComfyUI/models/vae (this is the Qwen-Image VAE, you might already have it)"
Incredible model, I hope the final anima will become an even better base for your lora/tune. Is there a list of tags that can be used? Because your examples contain very complex tags, the existence of which is impossible to guess.
Thanks, I hope so too <3
About the tags - yeah, it’s kind of tricky. I’m still experimenting with them while generating images. Some of them come from my own head, and for others I ask Gemini or ChatGPT for ideas
You need the Booru tag autocomplete extension
@_1_ I'm talking about style tags. If you look at the author's examples, he uses natural language to set a realistic style. So I was interested in what style tags its model supports. I know about anima prompting and danboru.
Do you have a workflow that does not involve Fast Groups Bypasser (rgthree). I cannot get this node to install.
Just... delete it. It's just there to help bypass the feet and face groups
Just be aware that RES4SHO samplers/schedulers you use in some examples are not shown by ComfyUI if missing, so it's better to put a link somewhere.
the way samplers are implemented feels frustrating sometimes, always the sinking feeling when "No Missing Nodes"
This is amazing and, if done properly, could make me redundant in the best way. A lot of what I do is use anime to real conversion workflows to get a good sample of what an anime character would look like as a real person, then train character loras based on that. I hope this will make that unnecessary.
That said, while I have high hope for this project in the future, at the moment it seems to struggle in two areas. The first is output quality. The second is character consistency.
As an example, I have been able to get a few good pics of Weiss Schnee from RWBY out of this model, but each Weiss looks different than the last. If I was attempting to animate a video out of this, or even illustrate a fanfic, I wouldn't be able to use the different generations in the same project. This makes it, for my work case, unusable at the moment.
I hope that by the final iteration of this, the flaws (generation quality and character consistency) will be ironed out.