Unofficial custom-node for SANA: Efficient High-Resolution Image Synthesis with Linear Diffusion Transformer
- A init node with lots of bugs, do not try unless interested.
- ram will not released (tried but failed).
- 4gb vram cuda device will oom in text_encoder with non-4bit, but retry works.
- test on rtx cuda-device with win10+py311+torch 2.5.1+cu126.
- Batch_size not work, it's a loop which i don't think is a good idea, so i left it empty.
- ram & vram: 16+gb ram, init model needs lots of ram. 4gb vram at least.
- text_encoder: gemma-2-2b-it ~ 5gb vram, gemma-2-2b-it-bnb-4bit ~ 2.3gb vram.
- dit: ~ 3.5gb vram.
- vae: ~ 3.1gb vram for 4k.
- https://huggingface.co/Efficient-Large-Model/Sana_1600M_1024px/blob/main/checkpoints/Sana_1600M_1024px.pth
- https://hf-mirror.com/Efficient-Large-Model/Sana_1600M_1024px/blob/main/checkpoints/Sana_1600M_1024px.pth China mainland users.
vae: autodownload or manual download or git from below links into ComfyUI\models\vae, rename folder_name to models--mit-han-lab--dc-ae-f32c32-sana-1.0.
- https://huggingface.co/mit-han-lab/dc-ae-f32c32-in-1.0
- https://hf-mirror.com/mit-han-lab/dc-ae-f32c32-in-1.0 China mainland users.
text_encoder: autodownload or manual download or git from below links into ComfyUI\models\text_encoders, rename folder_name to models--unsloth--gemma-2-2b-it.
- https://huggingface.co/unsloth/gemma-2-2b-it
- https://hf-mirror.com/unsloth/gemma-2-2b-it China mainland users.
4bit text_encoder: autodownload or manual download or git from below links into ComfyUI\models\text_encoders, rename folder_name to models--unsloth--gemma-2-2b-it-bnb-4bit. It will remain 1.5gb in vram after text_encode, which can be freed by ComfyUI-Manager Free model and node cache.
- https://huggingface.co/unsloth/gemma-2-2b-it-bnb-4bit
- https://hf-mirror.com/unsloth/gemma-2-2b-it-bnb-4bit China mainland users.
- I only installed 2 modules, so i do not know what are really needed.