ControlNet

from strands_sana import sana_controlnet_generate

sana_controlnet_generate(
    prompt="a futuristic robot",
    control_image="hed_edges.png",
    model="<a Sana-ControlNet checkpoint>",
    controlnet_conditioning_scale=0.8,
    steps=20,
)

Diffusers ControlNet weights are 'Coming soon' upstream

As of NVlabs/Sana model_zoo.md, the diffusers-format ControlNet checkpoints are listed as "Coming soon". Native checkpoints exist (e.g. Efficient-Large-Model/Sana_1600M_1024px_BF16_ControlNet_HED) but require upstream's app/sana_controlnet_pipeline.py loader.

The tool detects this and returns a helpful error pointing to the workaround.

When upstream ships diffusers ControlNet

Just point at the new repo:

sana_controlnet_generate(
    model="Efficient-Large-Model/Sana_1600M_1024px_BF16_ControlNet_HED_diffusers",
    prompt="...", control_image="...",
)

Control image preprocessing

ControlNet expects pre-processed control maps (HED edges, Canny, depth, normal). Use upstream's HED preprocessor or any ControlNet-compatible preprocessor:

import cv2
img = cv2.Canny(cv2.imread("source.png"), 100, 200)
cv2.imwrite("canny.png", img)

controlnet_conditioning_scale

value	effect
0.0	ControlNet disabled
0.5	gentle guidance
1.0	default — strong guidance
1.5+	over-saturated, may break