5 image models benched on a 12GB laptop: Nano Banana 2.1 scores 46/54
KWS_Studio · reddit · 2026-11-08
To pick keyframe models for an AI filmmaking course, the author ran 5 cinematic prompts with the same seed through Z-Image Turbo, Krea 2 Turbo, FLUX.2 klein 4B/9B (local, RTX 4080 Laptop 12GB, ComfyUI) plus hosted Nano Banana 2.1 via API, scoring skin, light, composition, and prompt adherence out of 54.
Scores: Nano Banana 2.1 (hosted) 46; Z-Image Turbo 42 (11s); Krea 2 Turbo 41 (11.5s); FLUX.2 klein 4B 40 (19s); FLUX.2 klein 9B 40 (35s).
Takeaways:
- Every model failed the low-angle + Rembrandt prompt; describing the visible effect works better than naming the technique.
- "f/8" is ignored; "sharp edge to edge" works.
- Hosted Nano Banana 2.1 clearly wins on the macro old-photo insert; local models struggle with photographic content.
Caveats: subjective scoring; Krea 2/FLUX 9B returned squares needing crops; Qwen-Image-2512 excluded for black outputs; Krea 2's community license allows free commercial use under $1M annual revenue.
More from Multimodal
- Full guide: fully local one-shot music videos with ComfyUI Video Builder on 16GB VRAM — Cheap_Credit_3957 · 2026-11-08
- MoCap-to-video pipeline: RGBMoCap + Blender + Seedance 2.5 re-renders K-pop choreography — Choidonhyeon · 2026-11-08
- MiniMax H3 fantasy fight short showcases cinematic video generation — shanytopper · 2026-11-08
- Fixing axe parries and heavy hits in MiniMax H3 with FL2VA frame guides in ComfyUI — shanytopper · 2026-11-08
- ComfyUI-from-your-Phone: free open-source mobile remote for ComfyUI via Tailscale — al-aSak · 2026-11-08
- Automatic1111 untouched for 2 years, Reddit asks what replaced it — The_LoneCoder · 2026-11-08