QwenImage 2.1 vs Edit 2511 tested: the 7GB model doesn't lose every battle
DiagramAwesome · reddit · 2026-09-23
The author ran seven targeted prompt tests pitting QwenImage 2.1 (7GB int8) against Qwen-Image-Edit 2511 (20GB int8), generating three batches per model and picking each one's best output.
Key findings:
- It's remarkable that the 7GB model comes close to the 20GB one, but some capabilities were likely lost: 2.1 wins on some tasks and even underperforms the older model on others
- Spatial grounding — drawing directly into the image where things should go — works much better, and results generally look more realistic
- Tests covered shadow reasoning, packaging text edits, partial style transfer, precise single-object edits, outpainting, and MS Paint-to-photo conversion
- Full test prompts are included for reproducibility
More from Multimodal
- Pooling-aware training cuts retrieval vectors 4x while boosting nDCG to 97.8% of unpooled — lateinteraction · 2026-11-11
- Generating accurate 3D building models from 150 planning PDFs with new Gemini model — fofrAI · 2026-11-11
- Stanford's DyCraft Uses Agentic Decomposition to Cut Physical Errors in Long-Horizon 4D Scene Generation — StanfordUniversity · 2026-11-11
- Midjourney SREF pick: --2523103199 for loose pencil-sketch portraits — tisch_eins · 2026-11-11
- H3 Image Tips: 16MP via ComfyUI Workflow Plus Detailed Prompts Yields Best Results — gatortux · 2026-11-11
- Gemini-generated endless loop of quantum cryostats self-assembling — hapticdata · 2026-11-11