ComfyUI-ClarkAirSana

ComfyUI-ClarkAirSana

Run Clark Air Sana 1.6B — ternary (~1.58-bit) weights on real GemLite INT2 CUDA kernels — natively in ComfyUI under the standard KSampler. Self-contained: the Sana model, the Gemma text encoder, and the DC-AE VAE are all included, so there is no other custom node to install.

Quick start

1. In ComfyUI-Manager, install Clark Air Sana and restart ComfyUI. (That's the only node pack you need. The Python deps install automatically.)
2. Download the model → put it in ComfyUI/models/clark_air_sana/: clark_air_sana_gemlite_comfy.safetensors (495 MB).
3. ⬇ Download the workflow: example_workflow.json → drag it onto the ComfyUI canvas → press Queue.

The Gemma text encoder and DC-AE VAE download themselves on the first run.

demo

Requirements

• NVIDIA CUDA GPU. GemLite INT2 runs on Triton kernels, so Linux or WSL2 is the smooth path. On native Windows you need triton-windows (the deps file installs it automatically, but Triton on native Windows is community-supported and may not work on every setup — WSL2 is recommended).
• Python deps (installed for you from requirements.txt): gemlite, bitsandbytes, triton, transformers, diffusers, accelerate, timm, einops, safetensors, numpy. If your ComfyUI is the portable build, the Manager handles this; to do it by hand use the bundled Python: .\python_embeded\python.exe -s -m pip install -r ComfyUI\custom_nodes\ComfyUI-ClarkAirSana\requirements.txt.

Nodes

All provided by this pack (category ClarkAir/Sana):

Node	Output	Role
Clark Air Sana Loader	MODEL	the GemLite INT2 ternary transformer
Clark Air Gemma Loader / Encode	CONDITIONING	Gemma-2 text encoder (4-bit by default)
Clark Air DC-AE VAE Loader / Decode	IMAGE	DC-AE latents → pixels
Clark Air Sana Empty Latent	LATENT	32-channel Sana latent

Workflow

example_workflow.json (drag onto the canvas) wires:

Clark Air Gemma Loader ─┬─ Gemma Encode (positive) ─┐
                        └─ Gemma Encode (negative) ─┤
Clark Air Sana Loader ───────── MODEL ──────────────┼─ KSampler ─ Clark Air VAE Decode ─ SaveImage
Clark Air DC-AE VAE Loader ──── VAE ─────────────────┘     │
Clark Air Sana Empty Latent ─── LATENT ───────────────────┘

KSampler: euler / normal, 20 steps, cfg 4.5, 512×512 (verified end-to-end through ComfyUI's KSampler). example_workflow_api.json is the same graph in API format.

Footprint (~3.2 GB)

Component	Download
transformer (this repo)	495 MB (ternary, GemLite INT2 + FP8 islands)
Gemma text encoder	~2.1 GB (unsloth/gemma-2-2b-it-bnb-4bit, 4-bit; switch to Efficient-Large-Model/gemma-2-2b-it for fp16)
DC-AE VAE	1.2 GB (loaded bf16)

Notes

• Keep the transformer GPU-resident. GemLite holds packed codes as buffers that ComfyUI's lowvram weight-streaming can't move; the loader pins the ~0.5 GB trunk on the GPU. Avoid --lowvram for it (Gemma/VAE are fine).
• Works offline. Once the Gemma encoder and DC-AE VAE have been fetched once, the nodes load them straight from the local cache with no network check, so a queue never stalls on huggingface.co when you are offline.
• The Sana model code under sana/ is vendored from ComfyUI_ExtraModels (Apache-2.0), which adapts NVlabs/Sana — see NOTICE. This pack bundles it so it runs with no external node dependency.