n-Arno/Anima-P3-Turbo-AIO-Q4_K

Update

Two new model has been uploaded:

• Anima-V1-Turbo-AIO-Q4_K.gguf
• nArnima-Turbo-SdCpp_Q4_K.gguf

The Anima V1 Turbo is an AIO model based on Anima Base V1 with three LoRA merged:

• An adapter of Qwen3-06b to partially abliterate refusal
• A custom Turbo LoRA: https://civitai.red/models/2619830/turbo-for-anima-less-steps
• A small portion of a LoRA adapter to increase Preview 3 LoRA compatibility

Here is an example result (with the same command as below):

nArnima is the quantization of this model: https://civitai.red/models/2626223/narnima-turbo-all-in-one-anima-v1-checkpoint

Both have been quantized using this command:

sd-cli -M convert -m model-BF16.safetensors -o model-Q4_K.gguf --tensor-type-rules "^model\.diffusion_model\.blocks\.[0-9]*\.(cross_attn|self_attn|mlp).*=q4_K,^model\.diffusion_model\.blocks\.[0-9]*\.adaln_modulation_.*=q6_K,^text_encoders\.llm.*=q8_0"

The choice of tensor data type is experimental: https://civitai.red/articles/30215/quantizing-anima-some-tests

---

Previously

This is a "All In One" version of Anima-Preview3 for the GPU Poor:

• Made specifically for Stable-diffusion.cpp, it will not work with ComfyUI/Forge (the text encoder prefix is different)
• It was manually concatenated (VAE+LLM+DiT) using torch and merged with the Turbo LoRA
• It was then quantized using sd.cpp: sd-cli -M convert -m Anima-P3-Turbo-AIO.safetensors --type q4_K -o ANIMA/Anima-P3-Turbo-AIO-Q4_K.gguf

Parameters:

• CFG 1
• sampler: er_sde
• scheduler: smoothstep (~beta)
• steps: 16

Example usage:

sd-cli -m Anima-P3-Turbo-AIO-Q4_K.gguf --vae-tiling --fa --offload-to-cpu --steps 16 --cfg-scale 1 -W 832 -H 1216 --sampling-method er_sde --scheduler smoothstep --cache-mode spectrum --seed 666 -p "score_9, masterpiece, best quality, newest, 1girl, solo, cowboy shot, facing viewer, holding a board with 'AIO ANIMA TURBO' written on it"

And as a server (with a LoRA folder provided):

sd-server -m Anima-P3-Turbo-AIO-Q4_K.gguf --vae-tiling --offload-to-cpu --fa -l 0.0.0.0 --lora-model-dir ./lora --steps 16 --cfg-scale 1 -W 832 -H 1216 --sampling-method er_sde --scheduler smoothstep --cache-mode spectrum --seed 666