feat: add Anima support

[rmatif]

For #1245

./build/bin/sd-cli --diffusion-model models/anima-preview.safetensors --llm models/qwen_3_06b_base.safetensors --vae models/qwen_image_vae.safetensors -p "a cute cat" --fa -v -H 1024 -W 1024 --cfg-scale 4

Download model: https://huggingface.co/circlestone-labs/Anima/tree/main/split_files

Output eg:

[Juste-Leo2]

I tested the configuration on my Windows computer and it worked ! Thank you very much for this PR :) . I just had a VRAM issue with VAE (probably a compatibility issue with Flash-attention, judging by the warnings I got during compilation). This limited my images to 512*512 pixels, given that I have little RAM (16GB).

Note: For compilation, I had to add a flag.

cmake .. -DSD_CUDA=ON -DCMAKE_CXX_FLAGS="/bigobj"
cmake --build . --config Release

Here is the command I ran:

./build/bin/sd-cli.exe --diffusion-model models/anima-preview.safetensors --llm models/qwen_3_06b_base.safetensors --vae models/qwen_image_vae.safetensors -p "a cute blue cat in a park" --fa -v -H 512 -W 512 --cfg-scale 4

Here is the result :)

[wbruna]

This limited my images to 512*512 pixels, given that I have little RAM (16GB).

Just turn on VAE tiling; this took around 6G with no Flash Attention on my RX 7600 XT (on ROCm; Vulkan works too):

./sd-cli --diffusion-model anima-preview.safetensors --llm Qwen3-0.6B-UD-Q4_K_XL.gguf -W 1024 -H 1024 --vae Qwen_Image-VAE.safetensors --vae-tiling --lora-model-dir . --steps 9 --cfg-scale 1 -p "a cute blue cat in a park <lora:Anima-Preview_Turbo_8step:1>" -o ./test_anima_cat.png

(also with https://huggingface.co/Einhorn/Anima-Preview_8_Step_Turbo_Lora/ because I'm impatient 🙂)

[Juste-Leo2]

Just turn on VAE tiling; this took around 6G with no Flash Attention on my RX 7600 XT (on ROCm; Vulkan works too):

Thanks a lot! I didn't know about that :)
I successfully reproduced it using the safetensors format for the encoder.

I have high hopes for this architecture for inference on the phone (via Termux, for example), given the small size of the model and the encoder.

.\build\bin\sd-cli.exe --diffusion-model models\anima-preview.safetensors --llm models\qwen_3_06b_base.safetensors -W 1024 -H 1024 --vae models\qwen_image_vae.safetensors --vae-tiling --lora-model-dir .\models --steps 9 --cfg-scale 1 -p "a cute blue cat in a park <lora:Anima-Preview_Turbo_8step:1>" -o .\test_anima_cat.png

[leejet]

Thank you for your contribution.