RC-AIO 2.9B / RealCosplay All In One Anima2.9B
All-in-One Multi-Style Base Model (Anima 2.9B Full Fine-tune)A full fine-tune of Anima 2.9B designed as a true multi-style (Realistic / 3D / Anime) All-in-One base model.It keeps the strong anime foundation of Anima by default, while high-quality real cosplay and scenery photos are used to improve anatomy, composition, and lighting. Realistic capability is significantly enhanced without destroying the original anime concept.Improvements over 2.0B
Realistic style fits noticeably better, and limbs/fingers are more stable and natural. This is currently my most satisfying version.Recommended SettingsPositive (Base)
masterpiece, best quality, highres,beauty facePositive for Realistic
photo \(medium\) 【0.5-1.5】Negative
worst quality, low quality, lowres, jpeg artifacts, mature female, fat, aegyo sal, puffy cheeks, ugly face, thick lips, Parameters
Sampler: ER SDE / Euler
Schedule: beta
CFG: 3~4 (I recommend 3.5)
Shift: 3
Hires.fix / Upscale Tip
If using secondary sampling or Hires.fix, keep the upscale multiplier at 1.8x or lower. Higher values can improve finger details but tend to introduce more noise.
Artist Style
@danbooru artist name
Specially Trained Tags
Background: blurry foreground, blurry background…
Art Style: photo \(medium\), figure \(medium\), realistic, 3D, impasto, draft, flat color, no lineart, @nyalia, @eknn…
Lighting: dark, chiaroscuro, night, backlighting, overexposure… (pls use "chiaroscuro" instead of "camera flash")
Composition: fisheye, partially underwater shot, dutch angle, pov crotch, foreshortening…
Age: … , man/woman, mature male/female, old man/woman, …
Others: aegyo sal, puffy cheeks, bloom, vignetting, chromatic aberration, etc.
Not Recommended Tag
solo (removed from training data, use solo focus instead)
Duplicate synonymous tags (e.g.
shirt, red shirt).These were filtered out during training to improve prompt adherence.
Description
14.6k dataset
FAQ
Comments (6)
Comfy isn't recognizing the checkpoint as containing a valid clip and vae. Even used your workflow.
Hey, Anima models (including the 2.9B one) aren’t regular all-in-one checkpoints — they don’t contain CLIP or VAE inside the main file. You need to load them separately:
RC-AIO 2.9b model → models/diffusion_models/
qwen_3_06b_base.safetensors → models/text_encoders/
qwen_image_vae.safetensors → models/vae/
Use Load Diffusion Model + Load CLIP + Load VAE instead of the normal Checkpoint Loader. Also make sure your ComfyUI is at least v0.33.1 (or newer), otherwise the 2.9B version may not load properly even with the right workflow.
@Sennke Yeah that (using diffusion_models) was what I first tried, but the output was garbled, either fully black or just grainy nothingness. In the example workflow there was a checkpoint loader, and "AIO" often refers to a unified model, so I thought I would try. I'll see if updating comfy helps at all, but I think it was already new enough.
@Sennke Actually yes, even though my Comfy was pretty new, updating Comfy fixed it. Thanks for the suggestion.
This is really really good, prompt coherence is great too. Thank you for this! The only thing I would say is for me anyway, when I tried making cunnilingus it often defaulted to facesitting, had to like really prompt against that and use negative. Though other than that, I look forward to the other versions
Thanks, I’ll strengthen it next version. This tag is underrepresented in my training data — not my kinks haha.