Here's a thing almost nobody uses, because everybody thinks the only knobs are prompt + CFG + sampler. There's a control layer underneath the prompt — the image you start from — and it quietly decides your entire palette, mood, and brightness.
I'll show you the proof, then the exact recipe. Every image below is the same prompt, the same seed — the only thing that changes is the solid color I start the img2img from.
!same prompt, same seed, only the canvas changes
A near-black canvas gives a genuinely dark, moody scene. A near-white canvas gives that airy, high-key look. A warm-orange canvas gives a real golden hour. Same prompt. Same seed. The model kept full creative freedom — it just obeyed the color.
Why this matters: SDXL literally can't do dark
This isn't just a color grader. It fixes a real, widely-complained-about SDXL weakness: SDXL cannot make a genuinely dark or genuinely bright image from a text prompt. The training noise schedule biases every generation toward mid-grey average brightness. Prompt "pitch black night, almost no light" all you want — you get dim daylight. The canvas trick is the fix, because you're no longer asking for dark, you're starting from it.
The technique
Make a solid color image the size you want (any paint app, or a 1-node fill in ComfyUI).
Run img2img with your normal prompt, from that canvas.
Set denoise strength ≈ 0.80. This is the whole trick.
That's it. Works on any SDXL checkpoint, in any UI with an img2img tab.
The number that makes it work: 0.80
I swept the denoise strength so you don't have to. This is the part people get wrong:
!strength sweep: 0.70 too strong, 0.80 perfect, 0.88 ignored
0.70 — too strong. The canvas overpowers the model. The black canvas stays a near-black rectangle; there's barely an image.
0.80 — the sweet spot. The canvas anchors palette and brightness, and the model fills in a fully coherent, detailed scene. Look at that dark mountain lake — that is the dark, atmospheric image SDXL "can't" make.
0.88 — too weak. The model ignores the canvas and snaps back to its default daylight average. Barely different from a normal generation.
If your result looks like a solid color, go up. If it looks like a normal daytime image, go down. Live around 0.78–0.82.
Palette control, not just brightness
The same mechanism gives you a color-grade dial. Here's one prompt/seed across seven canvases:
!seven canvases, seven palettes
Warm/saturated canvases anchor hardest (that sunset is unmistakable). Cool and neutral tones grade more gently. Pick the canvas that matches the mood you're chasing and let the model do the rest.
Where I'd use it
Dark scenes SDXL refuses: night, dungeons, noir, horror, moody landscapes.
High-key / minimalist product and portrait looks that normally come out muddy.
Consistent color grade across a set — same canvas = same palette family, shot to shot.
Time of day on a landscape without touching the prompt.
No LoRA, no ControlNet, no extension. Just a colored rectangle and one slider.
All test images were generated on AlbedoBase XL (SDXL 1.0), Euler a, on a single RTX 4090 — same prompt and seed throughout, only the starting canvas and denoise strength changed. If you try it, post your dark scenes below; I'd love to see the ones SDXL always told you it couldn't make.