IceRealistic
A 2.5D hybrid model that renders hyper-detailed "doll-like" characters — sitting at the boundary between photorealism and anime. Fine-tuned on 50,000 images over 60 epochs on an Anima2B base.
50k Training images
60 Epochs
Anima2B Base model
IceRealistic is a 2.5D hybrid checkpoint trained to produce what I call "dolls" — characters that sit right at the edge between photorealism and anime. Think hyper-detailed skin, sculpted anatomy, glossy fabrics, and expressive anime faces grounded in a realistic render. Not quite a photo, not quite a drawing — something in between.
Fine-tuned on 50,000 images over 60 epochs on top of Anima2B. This model excels at character portraits, full-body shots, fashion and costume renders, and stylized scenes with strong lighting.
---
What it's good at:
- 2.5D character renders with a "doll" quality
- Detailed clothing, fabrics, and accessories
- Expressive faces with realistic anatomy
- Strong lighting and indoor/outdoor scenes
What it's not ideal for:
- Pure photorealism without anime influence
- Landscapes or scenes without characters
---
Description
FAQ
Comments (7)
Is there any hope for Qwen 3.5 4B Text Encoder for Anima 2B base?
Honestly, there is hope, but the results are not amazing. On my benchmarks qwen 3.5 4b wins about 15% of the time. If there is demand I can upload it but again don't expect amazing results
@LyloGummy I have also been testing Anima at the architectural level. The current bottleneck seems to be its dual-encoder setup. From what I can tell, the T5 encoder carries most of the stable text-conditioning signal, while the Qwen LLM path appears to have a much smaller influence.
Because of that, upgrading Qwen to a larger LLM probably will not create drastic improvements by itself. Even if the larger LLM understands prompts better, the diffusion model still has to be trained to properly use that embedding space.
My guess is that making a larger LLM path actually matter would require retraining the adapter/conditioning layers at minimum, and possibly retraining the base model with less dependence on T5.
its got a unique artstyle and looks objectively very good. but worse at prompt adherence
Hi @goeypants400! Appreciate the feedback! Do you have any examples of this? I will work on fixing the issues on the next version
@LyloGummy dont really got picture example. i used proper settings. the finetuning made it better looks-wise, but lost prompt adherence such as adding random stuff like magic and bright lights or wrong proportions. eyes and hair is place it happens a lot, hair floating or not shaded. in one of your own examples, the beach pink hair one. the thigh fades into the background, many times stuff has faded. these are not just rare occurrences, its probably just the style is not fully done finetuning
在相同的提示词下服装细节会变得更多，但是色准在有多种颜色的时候似乎有色相偏移