REEL — sft

Project page · Code · Paper

REEL is a tool-using visual-memory planner for long-horizon, multi-shot video generation, built on Qwen3.5-9B. It retrieves and creates visual evidence, assesses generated candidates, and updates memory for later shots.

Checkpoint

• Stage: Supervised fine-tuning.
• Parent checkpoint: Qwen/Qwen3.5-9B.
• Stored precision: float32, in Hugging Face Safetensors format.

This release contains the planner, tokenizer, processor configuration, and chat template. Video generators, optimizer states, and training data are not included. The GRPO releases use BF16 inference exports of their selected training checkpoints.

Use

Use a Transformers version with Qwen3_5ForConditionalGeneration support:

import torch
from transformers import AutoProcessor, Qwen3_5ForConditionalGeneration

model_id = "redflash6/sft"
processor = AutoProcessor.from_pretrained(model_id)
model = Qwen3_5ForConditionalGeneration.from_pretrained(
    model_id, dtype=torch.bfloat16, device_map="auto"
)

This loads the planner, not an end-to-end video generator. Use the REEL tool runtime and backend adapters to execute memory actions and connect separately deployed video-generation services.

Released stages

SFT · Single-shot GRPO · Multi-shot GRPO

Citation

@misc{xu2026reel,
  title = {REEL: Learning to Retrieve, Evaluate, and Evolve Visual Memory for Any Long-Horizon Multi-Shot Video Generation},
  author = {Hang Xu and Guohui Zhang and Xiaoxiao Ma and Hu Yu and Jie Huang and Jiayin Xiao and Lin Song and Haoyang Huang and Nan Duan and Feng Zhao},
  year = {2026},
  note = {Preprint},
  url = {https://xuhang07.github.io/REEL/}
}