--base1_v3
Actually, there are some minor but visible fixes, mostly stylistic. The old problem with certain expressions still exists, but maybe it will be fixed in the next version. Also skin texture not enough detailed (i think i'll made just a lora with nicegirls for new anima, this should be fast fix)
--base1_v2
To be completely honest, I trained this model on this specific caption format while I was pretty drunk, so it is what it is lol. I'm not even entirely sure how it works under the hood, and I didn't have enough time to thoroughly test everything out. Also, please don't come at me calling me retarded or saying "it doesn't work like that" - just chill, I'm just experimenting here. Hopefully, you guys can help me with that and test.
Anyway, here is a mini-guide on how it's supposed to work (written by Gemini):
To achieve maximum realism, analog grit, or authentic casual smartphone aesthetics, you should follow the exact prompt structure the model was trained on.
📐 The Prompt Formula:
[Prefix Tags] + [Natural Language Description] + [Suffix Tags & Score]
Prefix Tags (The Setup): Set the camera type, lighting, safety, and shot style at the very beginning.
Core Description (The Scene): Describe the subject, clothing, pose, and background using natural English sentences (avoid messy tag-soup).
Suffix Tags (The Quality & Era): Close your prompt with the simulated year of the photo and the quality score.
📋 KEYWORDS TO COPY-PASTE
1. Camera & Tech Prefixes:
@smartphone_photo— Casual, modern mobile look with subtle computational processing.@compact_digital_photo— Early 2000s "point-and-shoot" digicam vibe.@film_photo— Authentic analog look with rich organic textures and grain.@vhs_screencap— Retro video tape style with scanlines.@dslr_photo— Clean, professional camera rendering.
2. Lighting & Style Prefixes:
@available_light— Soft, natural indoor/outdoor daylight.@direct_flash— Harsh, flat flash (perfect for late-night party vibes or digicam looks).@candid_photo— Caught-on-camera, unposed, natural moments.@posed_photo— Deliberate posing.@mirror_reflection— Perfect for mirror selfies.@underexposed/@overexposed— For dramatic low-light or high-contrast shots.
3. Safety Blocks:
@sfwor@nsfw(choose depending on your target generation)
4. Era & Quality Suffixes (Put at the very end!):
Years:
@1995,@2000,@2005,@2010,@2020,@2025Scores:
score_5,score_6,score_7,score_8,score_9(can be used in negative)
💡 Tip: Use score_8 or score_9 for high definition and clean details. Use score_6 or score_7 combined with @smartphone_photo or @compact_digital_photo if you want a grittier, intentionally imperfect lo-fi look!
📸 EXAMPLE PROMPTS
Modern Smartphone Selfie:
@smartphone_photo @sfw @amateur_photo @candid_photo @available_light A close-up portrait selfie of a 20-year-old woman with neon green hair and heavy eyeliner. She is looking at the camera with a neutral expression. The background is a blurry minimalist bedroom. @2025 score_8
Retro Digicam Flash (Vibe from 2005):
@compact_digital_photo @sfw @amateur_photo @posed_photo @direct_flash An overexposed snapshot of a young woman posing in a cluttered room at night. Harsh flash lighting, red-eye effect, visible digital noise, and washed-out colors. @2005 score_7
Analog Film Portrait (Cosplay):
@film_photo @sfw @amateur_photo @candid_photo @available_light A medium shot of a young woman cosplaying Princess Zelda, sitting inside aP.S.: still WIP, i plan extend dataset and train more
--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 (59)
Looks like there are several @ tags added for the base 1.0 version, awaiting for a more detailed instruction on those, nice job btw! Hope this becomes the new pony on the booru tag based realism model front.
Hey. I still dunno if they work good, but seems like half of them at least a lil bit are working
woohoo!
You're doing great, keep it up!
Amazing work! I'm doing a ZiT refine and the results are amazing
Yo, can u please share zit refine? Cause i tried but result is so so
@Danrisi Sent in DM
@prompt_bit_sorcerer bro can you share with me too?
@TaiLong just any workflow that loads an image into the z-img sampling will do what they're saying. You could go straight from one decoder to the other encoder as simple as copying your anima workflow into your z-img workflow.
goat
thanx, i'll continue my work under this model
This actually works insanely well. I pretty much deleted most of my Illustrious/Pony models because of your model.
Gigantic downgrade from v1.0. Literally gore generator. Unlucky.
Let's hope in future versions he will fix it, having the same results
I have improved anatomy, but flatter colors -- but I'm using mostly natural language prompts. I find when I remove the @ tags it brings some depth back, but not entirely. I do fear the "score" and @ tags just decrease overall usability right now, either due to lack of data or training. They feel a little like debugging tags or something, meaning they could be incredibly powerful eventually.
One thing I haven't done, is go back to trying other sampler/schedulers. Most examples right now are much simpler than the 1.0, 1.2 examples. Might be a contributing factor. If you're just getting anatomy horror in general, I could only suggest checking the settings and using natural language prompts... but can understand if that feels like it defeats the purpose of the base model.
Another thing to consider, I think most examples must be upscaling their outputs, even using Zimg or Klein to improve skin and some anatomy. Cause yeah in some cases my outputs turn more into playdough than skin. But in most it actually gets the scene and emotion in an impressive way. I stand by my original, it's like using Chroma merged with Zimg, it needs help on the polish but can really understand what you're asking for.
Most likely this is a problem of the anima itself and it cannot be solved without strong training.
wow Anima has become too good now.
anyone know how to train a lora (realistic not anime) ? I was using AI tool kit to train other models ,but cannot find any Anima in the training setting.
I used this one https://github.com/gazingstars123/Anima-Standalone-Trainer and for loras and for checkpoint
@Danrisi any runpod option? my GPU may not be enough for training.
@er850 runpod? I'm using vast. But if u ask about gpu, I use 5090 in vast and default pytorch (vast) template
@Danrisi thx
Try this notebook citronlegacy/citron-colab-anima-lora-trainer
@Sailor_Luna Thx bro. This one works, but you need some manual changes with the model path from hf repo. The lora works suprisingly well.
@er850 so you have successfully trained a realistic Lora? Tell me parameters please, my attempts were far from what I achieved with Illustrious
@Sailor_Luna just the default , but 10 repeat as I remember correctly. and maybe I am using refiner with the lora so I am good with the output .However, with just the raw image generated from first pass, it's 70% there. btw i am using synthetic reaslitic photos( mostly crafted with different models but quite real). so i did ~35 photos with ~ 3500 steps. . and you have to change the hugging face repo to point to the latest Anima base model V1 in the notebook(ask gemini /claude to do that). hope that provide some information
could you release as Lora too?
I can only make Lenovo for base1. But a Lora trained on 2500 images will look like crap, it won't remember anything from the dataset
@Danrisi you can extract lora from this checkpoint. that looks same good then as checkpoint.
Difficult to get good darker lighting now, without weird hues and tints, hope that will improve. Otherwise, it's very good. I mean, I rarely comment on anything. Even if I started out with a negative that's the only thing I can think of, otherwise superb.
Thanks for the feedback. I collect all the "negative" but useful feedback because I know there's a lot that needs fixing, and your feedback helps with that. I don't just "parse the internet and get 50000 images for a dataset." I actually manually collect every image I think will improve the model, its style, and concept. Plus, there's tons of manual work with captions. Of course, I do a one-step generation with an AI model, but then I manually edit each caption. So, this dataset took me about two weeks of hard work. This time, I want to make a dataset of the same size but with images that can fix problems and train based on an already fine-tuned checkpoint, not from scratch.
With this finetune we now have the over-the-top exaggeration of illustrated imagery with a consistent layer of photorealism on top. Then If you pass all this through a properly trained photographic model like ZImage/Klein/Ernie you will easily achieve the best of the two styles. We'll soon be free from all the high-rated garbage-tier Loras and copycat AI-slopped finetunes that infest this place
Pretty good.
I suggest including more photos for next versions (specially close ups) that focuses on hair and genitals, those are the 2 areas where still looks very "anime".
Are you doing any kinda of special workflow, not really getting same results much more of cartoon semi-realistic look. Also, my crappy AMD GPU likes to be unstable when it does a lot of AI stuff and reboot my whole computer, so a workflow that is a little lighter would be nice
Try RDBT's lora with different amount of strength. Aura Flow node also helps with certain checkpoints I find
this might sound crazy, but use the anima turbo lora alongside the lenovo ultrareal lora. er-sde or dpmpp-2m-sde-gpu for sampler seems to be working
edit: normal euler is the best
same amd gpu and plastic skin ı did everything to fix it nothing works so far
I was having the same issue. Check out the example images they have. the workflow used for them use a different ksampler (Clownshark) and a sampling method that the default Ksampler doesn't have. there are also some other tags used that seem to help. I still get the plastic look sometimes, but much less than when i basically copied the example image workflwo
It seems that the written guide is incomplete since many tags you've used, such as @amateur_photo @wide_angle and others are not listed
it's a good question. i'll recheck it today
every danrisi example workflow: how many steps do you want?
yes
anima base is doing great with 30 steps to seal the general composition. using 2s or 2m sampling would be around 25 for the full 50. whats the point of going through virtually 100 steps? the difference it yields is minimal at best.
i gen while working, so i dont care about gen time, i just wanna make sure i use ultimate settings. ofc it gen good even with 30 steps
This model creates a wow effect. <@_@>
Thank you for your work! ❤
This is excellent work!!
Even when I specify doggy style in the prompt, the model often produces a pov of cowgirl position quite frequently.
so I think this model would be even better if that were improved.
>Even when I specify doggy style in the prompt, the model often produces a pov of cowgirl position quite frequently.
Upon closer inspection, it appears that the model wasn’t actually generating cowgirl-position images despite the doggy-style prompt; rather, the images showed the woman in a doggy-style position as viewed from directly above her head. However, the model frequently crops out the expression from the mouth up when rendering these images.
z-img on chroma-distilled acid. it makes images that hit parts of your brain differently. I want to make horror content but I also.. dont
Indeed - mindblowing model ANIMA. Bit slower than SDXL but with benefits of FLUX and Z
@PurePrompt having the qwen clip is so nice, even with the updates to chroma-radiance, all it's missing is the sentence comprehension. I do wish this supported higher resolutions, but we'll see if more training in the finetunes makes a difference. It's still an amazing chroma substitute as a first step into klein/ltx.
@makiaeveli definitely it will happen in near future. High resolution, inpaint ability, controlnet and so on. Btw - there is already a hires LoRA for AN!MA, that allow to get images with over 2048x2048. Not classic fix but still good enough for paint styles... mostly. https://civitai.com/models/2540444/anima-highresaesthetic-boost
Lenvo Ultrareal and NiceGirls Ultrareal LoRAs is for initial ANIMA model, they are already included in UltraReal Fine-Tuned ANIMA - am i right?
If you're asking if their datasets are already in finetune, then yes, and I have many more of my own datasets too in finetune (that long time ago should be separate loras)
@Danrisi i see. Thx for response.
This is amazing!
This one + Photanima v1 merged @ 0.5 with cosmo dmd2 at 0.85 and lenovo ultrareal at 0.8 12 steps 1.3 cfg er_sde beta 544x784 produces great results
Thanks for the tip!
This model has done wonders, can't wait to see more from it
This is a great checkpoint, thank you for this! I have noticed however an issue with one cunnilingus concept. Like it seems to always have 3 girls, especially if I try to make it pov, and while that's great and all it really doesn't listen to the prompt can also break easily when trying to make it only two girl in cunnilingus. It seems since the beginning of ai this concept really has been the most difficult to have consistency in any model.
this is a good lora for group sex image structure, which can be used in contronet inpaint