--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 (43)
Does this great checkpoint work with the official Anima turbo Lora?
were can i get and how install sampler:hfx_stochastic_s2 used in your preview images for ultrarealFineTune_v12AnimaPreview3,
google did not help
https://github.com/WASasquatch/RES4SHO
ffs, i just realized that i didn't link this repo in description
@Danrisi Thanks
Please continue your work on this. It will turn out insane
Thanx <3
I'll do my best, cause i love anima. My main model rn
@Danrisi Same it can produce some insane images
Thank you! Love this model already, anima really does have a bright future ahead.
Thanks for the P3 version! I think the biggest boon here is being able to use newer LoRAs. I've been testing all the Anima LoRAs and have definitely passed by a few that are pretty decent for realism/semi-realism.
While realism isn't my focus, I definitely get the appeal of a model that understands both tags and natural language while also adapting well to styles. I'm sure some people will say Anima isn't meant for realism but, while that's true for the base model, even the creator of Anima has trained (non-public) realism LoRAs.
I clicked on that shit faster than I read the (non-public) portion lol
This feels like Chroma with improvements sometimes. I don't really use all the anime styles, I don't really know them, but simply porting over my F2K and Chroma prompts have worked well. Impressive honestly, is almost a 1-1 replacement for Chroma in most cases.
Edit: Okay after some more "hardcore" cases, it's not nearly as responsive. It misses some text, as well. I'd say it's closer to 0.8-1 replacement, that last 20% of hardcore nsfw understanding gets lost, but with it comes a better contextual understanding like Z-img brings. This sits in a weird place between Z-img and Chroma for me, maybe not directly replacing it, but more fun than the constant abominations or raw feeling in F2K.
1.2 works a lot better for me but still sometimes leans toward semi-3d/anime
Great job on this finetune and the loras, but I've got a couple of questions :
1) Are there recurring captioning words/phrases that would enhance the model's ability to steer away from anime/3d?
2) The negatives used are not booru tags nor natural language version of them. They remind SD 1.5 schizo prompting and they are also applied using weights which does absolutely nothing on clip-less models. Am I missing something?
3) What about the recommended prompting structure of the base model (masterpiece, best quality, score_7/ ye-pop -- worst quality, low quality, score_1, score_2, score_3, artist name)? Is there a reason why it is not used?
Hey, thanks for the feedback.
1) Absolutely. I rely heavily on photography-related terms to steer it away from 3D. Phrases like "analog film", "vintage digicam", "film grain", "amateur photo", and "phone camera" work wonders. Adding specific lighting descriptions (like "harsh flash" or "natural sunlight") also really helps push it toward realism. (using stuff like hasselblad 16k ultramsterpiece doesn't really helps)
2) You won't believe, but I just grabbed negative prompt from the most popular 1.5 model on civit. Surprisingly, it still seems to improve the results. Even if the weight syntax doesn't technically function the same way on this architecture, those specific negative tokens help suppress the plastic/3D look, especially in combination with my third point (at least i believe in it, maybe some natural language phrases helps even better)
3) The main reason I avoid the base model's recommended structure (score_8, masterpiece, etc.) is that those tags are deeply tied to its anime/3D training data. Triggering them acts like a magnet for that style — if you use them, it immediately pulls the generation away from realism and makes it look like anime or 3D again.
This has tons of potential. I'm using a lora extracted at rank 512 and performs as well as the previews on this page when using 'amateur photo' in the prompt and 'cartoon, anime' in the negative prompt.
I've seen on reddit you mentioned you trained this in pretty much just 1 day - that's impressive. Makes me wonder how good this could be if you trained at a lower LR for longer.
Also what do you think of using '@photo' (or something similar) as triggerword for 'style'? Since the style of the output is so heavily influenced by '@artist' tags - you could take advantage of that to make the model learn realism as a new style per-say. Again, also, maybe adding some anime images to the datasets and ensuring their captions contain 'anime' tag - could perhaps also help when using 'anime' in the negative prompt?
Really wish lodestones would try and finetune this model in a way that makes sense and does not break the model's characters knowledge because one big use case for this is to generate known anime characters in a realistic content akin to cosplay.
Anyways, you are awesome and so is this finetune. Thank you.
could you please share the lore
@Ceylon_Ai Sure: https://drive.google.com/file/d/1rXZD2rYymJyd5U75MsKkFTuMgL4rSJWS/view
It was done with the lora extract node from KJ nodes in comfyui
I think that longer training times might cause the model to forget some of its original knowledge. Finding a balance between these two factors would be very difficult.
@dawn66666666 Yes, that's not even debatable - it would definitely happen. Which is why for a real finetune it should be done with many anime images as well - just a small ratio of anime to real content. Ideally the same dataset it was used for anime would be used for this purpose but its unlikely the anima team would share it plus the model is only 2B and was originally realism-focused so going back to realism after an anime finetune would be difficult and it would never achieve quality akin to what Z-Image-Turbo can gives us. But hey, it would still be better than this.
Another option would be to still train like I mentioned but extend the parameters to 3B so we can now have 'more room' to teach the additional realistic data.
@Danrisi include nsfw real life images and this could easily be the best finetune before the full release of Anima
Already, but not too much. Next update will include a lot of this in dataset
@Danrisi And feets, lots of female feet and very feminine shemale, not your thing but I beg you, lets get rid of that ugly soles (see my post to know what im talking about). Do you want help with pics? I have everything from the natural look to the softest and most well-groomed look.
I just cant caption them I can use gemma4 26b it captions very well but it would take a long time and I would not dare using a smaller vision model than 26B to make the dataset bad with crap captions
It looks like the 'base model' is missing from the category. Regardless, the photorealism of the Anima model is impressive. Do you have any plans to strengthen Asian image datasets or develop a dedicated model for them later on?
+
Goal
Hope to see this improved on, faces can be a bit warped from afar
Yeah, I know. I hope to fix it in the next release, so people stop throwing shit on Reddit and telling me about wasting compute power
@Danrisi Ah, typical, 4chan feels like the worst in that regard some of them are so heated about Anima in general even though it's exactly what them and many others have been begging for albeit with minor problems. Don't listen to them lol, sorry if I came off like one of those
Судя по фотографиям мы разные аниме смотрим)))
?
"P.S.: in my flow i use custom sampler and scheduler, u can take it here https://github.com/WASasquatch/RES4SHO"
I got this and started using HFX_STOCHASTIC_S2 like your examples and I like it, slow but detailed.
You also mention scheduler - what do you use?
atan_detail (from the same node pack)
Anima Base 1.0 is out! Do you plan on trying fine-tuning it?
Yeah, ofc. I'm already sorting and captioning images for bigger dataset. Previous was 2k, new one will be 5k. I think everything will be ready in 4 days approximately
@Danrisi Exciting! Can I ask how you caption your images?
@Danrisi But.... I want to jerk off sooner than 4 days! :( Thanks man, you do good work, I've been following your loras for a long time now... Like a decade in AI years.
dear creater why i got this ？could u help me？
"prompt_id": "27c68cef-123c-4856-91c0-4ca504bce2ad",
"node_id": "NegativeCLIP_Base",
"node_type": "CLIPTextEncode",
"executed": [],
"exception_message": "ERROR: clip input is invalid: None\n\nIf the clip is from a checkpoint loader node your checkpoint does not contain a valid clip or text encoder model.\n",
"exception_type": "RuntimeError",
"traceback": [
" File \u0022D:\\AI\\Packages\\ComfyUI\\execution.py\u0022, line 535, in execute\n output_data, output_ui, has_subgraph, has_pending_tasks = await get_output_data(prompt_id, unique_id, obj, input_data_all, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb, v3_data=v3_data)\n ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^\n",
" File \u0022D:\\AI\\Packages\\ComfyUI\\execution.py\u0022, line 335, in get_output_data\n return_values = await asyncmap_node_over_list(prompt_id, unique_id, obj, input_data_all, obj.FUNCTION, allow_interrupt=True, execution_block_cb=execution_block_cb, pre_execute_cb=pre_execute_cb, v3_data=v3_data)\n ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^\n",
" File \u0022D:\\AI\\Packages\\ComfyUI\\execution.py\u0022, line 309, in asyncmap_node_over_list\n await process_inputs(input_dict, i)\n",
" File \u0022D:\\AI\\Packages\\ComfyUI\\execution.py\u0022, line 297, in process_inputs\n result = f(**inputs)\n ^^^^^^^^^^^\n",
" File \u0022D:\\AI\\Packages\\ComfyUI\\nodes.py\u0022, line 78, in encode\n raise RuntimeError(\u0022ERROR: clip input is invalid: None\\n\\nIf the clip is from a checkpoint loader node your checkpoint does not contain a valid clip or text encoder model.\u0022)\n"
],
"current_inputs": {
"clip": [
I'm hoping that you take this as far as it'll go while keeping Anima knowledge in place. Perhaps with an addition of using one of the reality models as refiner? Maybe a K9B semi-real to real?
I really like your work. Have you ever thought about doing a Chroma1-HD fine-tune Checkpoint as the realism is great and it's pretty much uncensored out the box. They do also have Chroma Flash DMD2 LORAs to. I think rank 256 and 128 are great.