rule of thirds composition works in Midjourney, Flux, and SDXL — but the syntax diverges sharply by architecture. This article covers 5 framing terms (rule of thirds, negative space, Dutch angle, leading lines, foreground/background layering) with copy-paste phrases, a cross-tool reference table (MJ V7/V8.1, Flux dev/schnell, SDXL, SD3), and the one discipline that separates consistent results from random: keep composition tokens in their own prompt slot, away from lighting.
rule of thirds composition in Midjourney, the hiker landed exactly at the bottom-right intersection, and the image looked cinematic. Then you pasted the same phrase into SDXL with a (rule of thirds:1.3) weight and got something passable. Then you dropped it into Flux with the weight syntax still attached — and the model ignored the weight entirely, yet the framing still worked. Then you tried it on SD3 and couldn't tell if it changed anything at all.rule of thirds, negative space, dutch angle, leading lines) map to the geometric priors a model learned from labeled training data. They act on the structure of the image. Lighting tokens act on its surface. According to Nightjar's analysis of prompt patterns for AI product photos, these two token families operate in independent attention spaces — adjusting one doesn't change the other. 1soft golden hour from a 45-degree angle is ambiguous: is the 45-degree the light direction or the camera angle? The model guesses. 1 Keep composition framing in its own phrase group, away from lighting descriptions.[Subject + action], [composition framing], [environment], [lighting as separate clause]lone hiker on a mountain peak, rule of thirds composition subject at bottom-right intersection, vast sky filling the left two-thirds, golden hour side lightrule of thirds composition is confirmed working on Midjourney V7/V8.1, Atlabs AI, getimg.ai, and Flux. 2 On SDXL you can also add an optional weight. On SD3, reliability is low (more on that in the table below).rule of thirds composition, subject positioned at the [upper/lower] [left/right] intersection--v 7): 3Canal path in French outskirts at dawn, lead-in lines curving to distant bell tower,
rule of thirds framing, warm lighting, pastel reflections --ar 16:9 --v 7
vast negative space, [sky / wall / fog] occupying [left/right/upper] two-thirds, minimalismminimal composition with large empty grey sky, lone figure in lower-left third, sharp focus on subject lands more consistently than negative space alone.dutch angle. Alternatives canted angle and tilted camera angle work across Midjourney, Flux, and Atlabs AI. 5 It produces a rotated horizon that reads as psychological unease — Atlabs categorizes it under diagonal lines, noting it creates "a sense of movement, speed, instability, or unrest." 2dutch angle, tilted horizon, off-kilter framingCyberpunk street scene with tilted camera angle, strong diagonal lines created by neon signs,
sense of chaos and speed, dutch angle
leading lines. The alternate lead-in lines works equally well. Atlabs describes these as "the most reliable way to control where the viewer looks" — actual lines in the environment pointing directly at the subject. 2leading lines pointing toward [subject], vanishing point at [location]a long wooden pier stretching into fog, strong leading lines drawing the eye toward a distant lighthouse — the prose tells the model where the lines go, which Flux handles better than the abstract label alone.blurred [foreground element] in foreground, [subject] in sharp focus in middle ground,
[background element] in background, three distinct layers of depth
| Tool | Syntax mode | Weight syntax | Composition framing behavior | Key restriction |
|---|---|---|---|---|
| MJ V7 | Natural language prose | :: weight removed in V7+ | All 5 terms reliable; pairs well with --raw for literal framing | Keep composition phrases in their own clause group |
| MJ V8.1 | Natural language prose | None | Same as V7; better prompt adherence overall; --raw recommended for precise framing | --no removed — exclusions must be reframed positively |
| Flux dev/schnell | Natural language prose only | (word:1.5) and (word)++ ignored | Describe spatial relationships, not just labels; Flux 2 Pro excels at "precise composition" | No negative prompt field — passing negative_prompt throws an error |
| Flux 2 (JSON) | JSON format supported | Native JSON composition key | Can specify "composition": "rule of thirds" directly in JSON | FLUX.2 only; not available on FLUX.1 |
| SDXL | Keyword labels + optional weight | (rule of thirds:1.3) works | All 5 terms recognized; multi-element scenes (depth layers, complex leading lines) may degrade | CFG 4–7 recommended; negative prompts supported |
| SD3 | Natural language | Not applicable | Poor prompt adherence — composition terms may produce near-identical output regardless of term used | Not trained with negative prompts; changing descriptors often changes nothing visible |
negative_prompt field introduces noise rather than removing elements. 9 Composition framing on SD3 hasn't been independently verified; treat it as unreliable until you've tested it for your specific prompt."composition": "rule of thirds" alongside "camera": {"angle": "slightly low angle", "distance": "medium shot"} in the same structured call. 7 If you're building pipelines on FLUX.2, this is worth knowing.(keyword:1.3) on SDXL if you want extra emphasis. Treat SD3 as unreliable. And always put composition phrases in their own clause — separate from lighting.A systematic guide covering 42 composition techniques — rule of thirds, Dutch angle, negative space, leading lines, and more — with exact prompt strings and AI-generated example images for each.
The official fal.ai guide to Flux prompting: T5-XXL encoder behavior, why keyword-weight syntax fails, JSON composition format in FLUX.2, and practical prompt structure.
このコンテンツについて、さらに観点や背景を補足しましょう。