Gallery

Every asset on this page was generated by strands-sana itself on an NVIDIA Thor dev kit (sm_110, CUDA 13).

---

Hero — Sana 1.5 vs Sana-Sprint

Same prompt, same seed, vastly different cost.

Sana-1.5 1.6B
20 steps · 6.4s

Sana-Sprint 1.6B
2 steps · 0.6s

sana_generate(prompt=PROMPT, model="sana-1.5-1.6b-1024", steps=20, seed=42)
sana_sprint_generate(prompt=PROMPT, model="sana-sprint-1.6b-1024", steps=2, seed=42)

---

PAG — Perturbed Attention Guidance

pag_scale=2.0 boosts coherence at ~2× compute cost.

sana_generate(prompt=PROMPT, model="sana-1.5-1.6b-1024",
              pag_scale=2.0, steps=15, seed=42)

---

Img2Img — restyle anything

Source image (left) restyled with two different prompts.

Source (Sana-1.5)

"anime cel-shaded style"
strength=0.5

"oil painting, masterpiece"
strength=0.6

sana_img2img(prompt="anime cel-shaded style",
             image_path="hero.png",
             strength=0.5, steps=2)

---

Schedulers — same prompt, 4 schedulers

Sana is a flow-matching model, so flow-match-euler (default) is highest fidelity. DPM-Solver also works well; non-flow schedulers degrade quality but generate quickly.

flow-match-euler
(default · best fidelity)

dpm-solver
(13× faster)

euler
(lossy on flow models)

ddim
(lossy on flow models)

sana_set_scheduler(name="dpm-solver", model="sana-0.6b-1024")
sana_generate(prompt=PROMPT, ...)

---

Inference Scaling — 4 candidates, CLIP picks best

Generate N variants in parallel, score with CLIP, return the highest-aligned image.

seed 200
CLIP 0.431

seed 201
CLIP 0.453

seed 202 — winner
CLIP 0.462

seed 203
CLIP 0.459

sana_inference_scale(prompt=PROMPT, n_samples=4,
                     score_fn="clip", seed_start=200)

---

Image-to-Video

Promote a still into a 49-frame clip.

Source PNG

49 frames · 24 fps

sana_image_to_video(image_path="hero.png",
                    prompt="zoom out, neon flickers",
                    frames=49, steps=10)

→ Full MP4

---

SANA-Video — 3 aesthetics

Ocean waves crashing on cliffs

Watercolor fox running through autumn

Neon plasma swirling

for prompt in [...]:
    sana_video_generate(prompt=prompt, model="sana-video-2b-480",
                        frames=49, steps=10)

---

Performance on Thor

Tool	Model	Output	Time
sana_generate	sana-0.6b-512	512×512 PNG	45.8s
sana_generate PAG	sana-1.5-1.6b-1024	1024×1024	62.6s
sana_generate (Sana-1.5)	sana-1.5-1.6b-1024	1024×1024	163s
sana_sprint_generate	sana-sprint-0.6b-1024	1024×1024	48.9s
sana_img2img	sana-sprint-i2i-1.6b-1024	1024×1024	9.3s
sana_video_generate (33 frames)	sana-video-2b-480	480p MP4	75.6s
sana_video_generate (97 frames)	sana-video-2b-480	480p MP4	102.6s
sana_image_to_video	sana-video-i2v-480	480p MP4	74.8s
sana_inference_scale (4×CLIP)	sana-0.6b-1024	4 PNGs + winner	42.2s

Times include first-load model loading (~22s for 0.6B, ~52s for 2B video).