📋 Sana Test

🧑 You

Create a uv project in ~/sana for this thing: https://huggingface.co/clark-labs/clark-air-sana-1.6b-1.58bit and run 10 test images that would show interesting output from a typical model. Keep the images in the same directory.

▶ Tool: read_file [path='~/Personal/skills/skill_index.md']

▶ Tool: fetch_url [url='https://huggingface.co/clark-labs/clark-air-sana-1.6b-]…

▶ Tool: run_bash [command='nvidia-smi 2>/dev/null && echo "CUDA AVAILABLE" ||]…

▶ Tool: directory_tree [path='~/sana']

▶ Tool: run_bash [command='mkdir -p ~/sana && cd ~/sana && uv init --no-readm]…

▶ Tool: run_bash [command='cd ~/sana && uv add diffusers transformers acceler]…

▶ Tool: write_file [content='#!/usr/bin/env -S uv run """Generate 10 test image]…

▶ Tool: run_bash [command='cd ~/sana && uv run python generate.py 2>&1', time]…

▶ Tool: replace_in_file [new_str=' pipe = SanaPipeline.from_pretrained( "]…

▶ Tool: run_bash [command='cd ~/sana && uv run python generate.py 2>&1', time]…

▶ Tool: run_bash [command='ls -lh ~/sana/sana_*.png && echo "---" && cd ~/san]…

▶ Tool: directory_tree [path='~/sana']

🤖 Assistant

Done! Here's what's in ~/sana/:

Project: Clark Air Sana 1.6B (1.58-bit ternary)

Setup: uv project with diffusers, transformers, accelerate, and torch — all dependencies self-contained in .venv.

Script: generate.py — run any time with uv run python generate.py

10 test images covering diverse styles and subjects:

Performance: ~8–9s per 512×512 image on your RTX 3070 (8GB), using enable_model_cpu_offload() to fit VRAM. The packed ternary model is only 374 MB vs 3.21 GB FP16 — an 8.6× compression.