How to run Sana locally.

Which model.

Sana is NVIDIA’s efficient image model: a linear diffusion transformer with a deep-compression autoencoder that shrinks images 32 times per side, so large sizes stay cheap. The family has grown in three steps.

• Sana 1.0 (late 2024): 0.6B and 1.6B at 1024 px, plus 1.6B models trained for 2K and 4K.
• SANA 1.5 (March 2025): 1.6B and 4.8B at 1024 px.
• Sana Sprint (March 2025): 0.6B and 1.6B distilled for one to four steps.

Start with SANA 1.5 1.6B for quality or Sprint for speed. In ComfyUI, results are a little behind NVIDIA’s own code, which samples with a Flow-DPM solver ComfyUI doesn’t have, and text in images is weaker than SDXL’s.

Two node packs.

ComfyUI has no Sana loader of its own. Two packs fill the gap, and they read different files.

• ComfyUI_ExtraModels, NVIDIA’s fork loads the original .pth checkpoints and samples with ComfyUI’s own KSampler, so you get live previews and ComfyUI’s samplers. It runs on NVIDIA GPUs or the CPU, not on a Mac.
• ComfyUI-SANA wraps the diffusers pipeline and loads diffusers folders. It runs on Apple Silicon, NVIDIA and CPU, with its own generate node instead of a KSampler.

Files you need.

For ExtraModels

Pick a preset whose name starts with Efficient-Large-Model/ in the SanaCheckpointLoader and it downloads the checkpoint itself. The Gemma encoder and the VAE download the same way.

Plus the Gemma 2 2B encoder (5.2 GB) and the DC-AE 1.1 VAE (1.2 GB), shared by all of them.

For ComfyUI-SANA

Whole diffusers folders, each with its own encoder and VAE inside, placed in ComfyUI/models/diffusers/<name>. Download only the variant you need: a full snapshot of the 1.6B repo is about 22 GB of duplicates.

• Sprint Sana_Sprint_0.6B_1024px_diffusers ComfyUI/models/diffusers/ folder: transformer, text encoder, VAE
  7.7 GB Download
• SANA 1.5 SANA1.5_1.6B_1024px_diffusers ComfyUI/models/diffusers/ folder: transformer, text encoder, VAE
  9.7 GB Download

hf download Efficient-Large-Model/Sana_Sprint_0.6B_1024px_diffusers --local-dir models/diffusers/Sana_Sprint_0.6B_1024px_diffusers

What fits your computer.

At 1024 × 1024 unless noted. The 18 GB for 4K is NVIDIA’s figure; the rest comes from user reports and file sizes.

• 8 GBTight

  The 0.6B models. The 1.6B at 1024 px needs about 8.7 GB with VAE offloading, and an 8 GB RTX 4060 laptop ran out of memory.

• 12 GBFits

  Sana 1.6B, SANA 1.5 1.6B and Sprint at 1024 px.

• 16 GBFits

  All the 1.6B models with room to spare. The 4K model needs more.

• 24 GBFits

  The 4K workflow (18 GB) and SANA 1.5 4.8B, which runs on an RTX 3090.

• MacVia pack

  Only through ComfyUI-SANA with device set to mps. Its README says Sprint 0.6B runs comfortably on Apple Silicon; no timings given.

• 8 GB graphics cards
• 12 GB cards
• 24 GB cards
• Mac with Apple Silicon

Set it up.

With ExtraModels (NVIDIA GPUs)

1. 1

   Remove the old pack

   If custom_nodes has city96’s ComfyUI_ExtraModels, delete or move that folder first. Both use the same folder name.

2. 2

   Install NVIDIA’s fork

   Manager can’t find it by name, so clone it and install its requirements with ComfyUI’s own Python. ExtraModels’ VAE loader also needs diffusers.

   Terminal, in ComfyUI/custom_nodes Copy

   git clone https://github.com/lawrence-cj/ComfyUI_ExtraModels.git
   pip install -r ComfyUI_ExtraModels/requirements.txt diffusers

3. 3

   Load NVIDIA’s workflow

   Restart ComfyUI and open Sana_FlowEuler.json (or the 2K, 4K, SANA-1.5 or Sprint one) from NVlabs/Sana’s ComfyUI folder.

4. 4

   Pick presets

   In SanaCheckpointLoader, choose an Efficient-Large-Model/… preset. In GemmaLoader, use Efficient-Large-Model/gemma-2-2b-it: it’s the same model as Google’s, without the gated access.

5. 5

   Swap the latent node

   EmptySanaLatentImage fails on current ComfyUI. Use ComfyUI’s EmptyHunyuanImageLatent instead: same 32 channels, same 1/32 scale. Keep the size a multiple of 32 and matched to the model.

6. 6

   Set CFG and run

   CFG 2 for Sana 1.0, 4.5 for SANA 1.5. The first run downloads the checkpoint, Gemma and the VAE.

With ComfyUI-SANA (Mac, or anywhere)

git clone https://github.com/geoffitect/ComfyUI-SANA.git
pip install -r ComfyUI-SANA/requirements.txt

Download a diffusers folder as shown above, restart ComfyUI and wire SANA Model Loader into SANA Generate and a Save Image node. On a Mac, set the loader’s device to mps. The original .pth files don’t load here.

Settings that work.

These follow NVIDIA’s own ComfyUI workflows for ExtraModels.

Sana 1.0 and SANA 1.5

Steps

28

CFG

2SANA 1.5: 4.5

Sampler

euler

Scheduler

normal

Size

1024 × 10242K and 4K models: their own size

Size step

32

Negative

yes

Latent

EmptyHunyuanImageLatent

Sprint

Steps

2

CFG

14.5 in ScmModelSampling

Sampler

scm

Scheduler

sgm_uniform

Sprint’s guidance goes in through the ScmModelSampling node, while the KSampler stays at CFG 1. Negative prompts do nothing on Sprint. In ComfyUI-SANA, the README suggests about 20 steps at guidance 4.5 for regular models and 2 steps for Sprint. Community tests find CFG 2 to 7 useful depending on style, and longer, more detailed prompts help.

How fast.

NVIDIA’s figures from its own code, not ComfyUI.

When it goes wrong.

Grey, black or yellow images

city96’s ExtraModels. Replace it with lawrence-cj/ComfyUI_ExtraModels and pick the Efficient-Large-Model/… presets.

Access to model google/gemma-2-2b-it is restricted or 401 Client Error … Cannot access gated repo

The workflow points at Google’s gated repo. Use Efficient-Large-Model/gemma-2-2b-it in GemmaLoader.

The checkpoint you are trying to load has model type gemma2 but Transformers does not recognize this architecture

Update transformers in ComfyUI’s own Python (python_embeded in the portable build).

ExtraVAELoader: No module named 'diffusers'

Install diffusers into ComfyUI’s Python environment.

Value not in list: vae_type: 'dcae-f32c32-sana-1.1-diffusers'

The node pack is too old. Update NVIDIA’s fork.

size mismatch for pos_embed

The preset and resolution don’t match the checkpoint, or the pack is old. Use the 2K or 4K preset with its own size, and update the fork.

TypeError: 'int' object is not subscriptable in KSampler

Another city96 ExtraModels symptom. Switch to NVIDIA’s fork.

Input type (torch.cuda.HalfTensor) and weight type (torch.HalfTensor) should be the same

One loader is on the CPU and another on the GPU. Put them on the same device.

Questions.

Sources: NVlabs/Sana repository [1], NVIDIA’s ComfyUI workflows, Sana 1.6B model card, ComfyUI_ExtraModels fork, ComfyUI-SANA, settings and quality thread, memory thread, VAE type thread.