Spaces:
Running on Zero
Running on Zero
multimodalart HF Staff
Restore to build-1 final state (bbc2e650) -- the tested working demo; cancelled the redundant rebuild that had started over it
f615c58 verified | title: InterleaveThinker | |
| emoji: 🤖 | |
| colorFrom: blue | |
| colorTo: indigo | |
| sdk: gradio | |
| sdk_version: 6.19.0 | |
| app_file: app.py | |
| short_description: Multi-agent interleaved text-image generation pipeline | |
| python_version: "3.12" | |
| startup_duration_timeout: 1h | |
| ## InterleaveThinker: Reinforcing Agentic Interleaved Generation | |
| This Space demonstrates **InterleaveThinker**, a multi-agent pipeline that endows any image generator with interleaved generation capabilities. It orchestrates a **Planner** agent and a **Critic** agent around the **FLUX.2-klein** image generator to produce step-by-step visual narratives with interleaved text and images. | |
| ### How it works | |
| 1. **Planner** (InterleaveThinker-Planner-8B, Qwen3VL) analyzes your prompt and generates a structured execution plan. | |
| 2. For each step, **FLUX.2-klein-9B** generates or edits the image. | |
| 3. **Critic** (Critic-SFT-8B, Qwen3VL) evaluates the generated image and refines the prompt if needed. | |
| 4. The result is an interleaved sequence of text and images. | |
| ### Paper | |
| [InterleaveThinker: Reinforcing Agentic Interleaved Generation](https://arxiv.org/abs/2606.13679) |