--- pretty_name: Persona Modality Mapping Qwen Image 2.1 language: - en tags: - synthetic - persona - parquet - image configs: - config_name: persona data_files: - split: train path: data/persona/train-*.parquet - config_name: description data_files: - split: train path: data/description/train-*.parquet - config_name: enhanced_description data_files: - split: train path: data/enhanced_description/train-*.parquet - config_name: qwen_pe_prompt data_files: - split: train path: data/qwen_pe_prompt/train-*.parquet - config_name: flux_pup_prompt data_files: - split: train path: data/flux_pup_prompt/train-*.parquet - config_name: pilot_40_persona data_files: - split: train path: data/pilot_40_persona/train-*.parquet - config_name: pilot_40_description data_files: - split: train path: data/pilot_40_description/train-*.parquet - config_name: pilot_40_enhanced_description data_files: - split: train path: data/pilot_40_enhanced_description/train-*.parquet - config_name: dream_enhanced_description data_files: - split: train path: data/dream_enhanced_description/train-*.parquet --- # Persona-Modality-Mapping-Qwen-Image-2.1 Generated images from [Qwen/Qwen-Image-2.1](https://huggingface.co/Qwen/Qwen-Image-2.1). Each row joins the text and generated image by `persona-id` and preserves the prompt recorded for the run. The `persona-image` column embeds the original PNG bytes. The full-run configurations use the source persona (`persona`), the source description (`description`), the PromptEnhancer rewrite (`enhanced_description`), the Qwen prompt-enhancer rewrite (`qwen_pe_prompt`), or the FLUX prompt-upsampler rewrite (`flux_pup_prompt`). The three full-run rewrites use the original persona as input. Pilot configurations use their own 40-record evaluation inputs and recorded prompts. Unknown generation parameters are null. See the [text mapping](https://huggingface.co/datasets/visual-memory/Persona-Modality-Mapping) for the text-only data. ## Configurations | Configuration | Train rows | | --- | ---: | | `persona` | 200 | | `description` | 200 | | `enhanced_description` | 200 | | `qwen_pe_prompt` | 200 | | `flux_pup_prompt` | 200 | | `pilot_40_persona` | 40 | | `pilot_40_description` | 40 | | `pilot_40_enhanced_description` | 40 | ## Provenance The full run used `claws-lab/persona-modality` at revision `1cce2b85133957584489c4367ab7d3821de98d88`. The generation code revision was `e7b9f2acac01db9d0efa8c09d2616d2ff7902a4c`. Files are sharded Parquet. Both runs retain the original persona IDs. ## Source and attribution The source text comes from [Persona Modality](https://github.com/claws-lab/persona-modality/tree/1cce2b85133957584489c4367ab7d3821de98d88) by Julius Broomfield, Kartik Sharma and Srijan Kumar. See the [paper](https://arxiv.org/abs/2502.20504) and the accompanying `LICENSE-source` for the MIT notice covering the upstream material. The [image generator terms](https://huggingface.co/Qwen/Qwen-Image-2.1/blob/main/LICENSE) describe conditions on use of its outputs. The full run used model revision `d26bb61231c349cf6b7896fa83353113880e1ba3` at 2048 × 2048 pixels and 40 steps. ## DREAM-inspired image generation The `dream_enhanced_description` configuration contains 200 images, one per persona ID, at 2048 x 2048 pixels and 40 steps. Its `prompt` is the concatenation of the five saved DREAM-inspired profile narrative sections. These profiles were expanded from the short persona using GPT-5.6 Terra. The full text mapping is available in [Persona-Modality-Mapping](https://huggingface.co/datasets/visual-memory/Persona-Modality-Mapping). The generator receives this text directly with native rewriting disabled. Seeds follow 42 + persona ID. Batch target was 8, with the actual batch, VRAM measurements, inference times and model revisions recorded in `dream-generation-manifest.json`. FLUX batches group equal padded text lengths. DREAM prompts were adapted from [Juan Mallo et al.](https://github.com/JuanMallodelaCalle/personalized-ai-visual-summaries), revision `9391ff124fa8b6ae2b7332a735fa82c80c74d322`. See `LICENSE-DREAM` for the CC BY-NC 4.0 attribution and terms for the source methodology.