Controlled Z-Image Base experiments show explicit composition prompts reshape the whole frame
Maleficent-Bowl-4841 · reddit · 2026-09-02
A Reddit user ran controlled prompt-engineering experiments on Z-Image Base in ComfyUI (INT8, Qwen3 4B text encoder, 50 steps, CFG 4), using a structured seven-part prompt (subject, composition, framing, environment, lighting, materials, style) and deliberately omitting generic quality tags.
In the composition experiment, changing only spatial instructions across seven variants (center/left/right/lower/large/small/extreme left) produced clear spatial rearrangement — and the model recomposed the environment around the subject rather than just moving the character: environments dominated in Small variants, characters dominated in Large ones. Results replicated across two seeds.
A second experiment swapped in seven radically different environments while keeping the character fixed, testing concept consistency across scenes. The author notes the small sample size and visual-only evaluation, framing it as a practical study showing structured prompts beat keyword lists for compositional control.
Related event: Controlled Experiments Test How Z-Image Base Prompts Shape Composition(2 posts)→
More from Multimodal
- Fable 5.1 nails generating a fountain-pen cursive English font via code — deedydas · 2026-09-02
- AI-Generated Short Film 'The Guardian Has Awoken' Showcases Video Model Quality — AIFilmLabsPage · 2026-09-02
- AI Group Dance MV Workflow: Syncing Multiple Dancers With Flova References and Choreography — aftahi_ai · 2026-09-02
- Seedance 2-Generated Downhill Mountain Biking Video Shows Off Motion Quality — Ok-Nerve941 · 2026-09-02
- Producing AI influencer videos with Seedance 2.5 and APOB AI — aftahi_ai · 2026-09-02
- Photoshop Beta adds masked generative editing with object reference, creators approve — jnack · 2026-09-02