Image and Video Model Testing
Findings
FLUX dev won cost to quality in a six-model image bake-off
GPT Image 2 medium quality cost 9x low quality, not the projected 3.9x
Reference batching on GPT Image 2 saved nothing (50.18 vs 50.19 credits)
Seedance image to video lost the subject's identity in 5 of 7 cuts; stills held
What it is
Bake-offs across image and video models to pick the best cost to quality for each job.
Stack
- FLUX dev
- GPT Image 2
- Nano Banana 2
- Krea
- Higgsfield
- OpenArt
- Seedance
- Veo
- MiniMax
- ComfyUI on Comfy Cloud
- FFmpeg
- Remotion







